Seeing was Believing: Part 1

There is a tradition in the United Kingdom every year when Christmas arrives. The Queen addresses the nation on national television. But last year was a bit different. An “alternative” message was delivered by the Queen on BBC Channel 4 [1]. Let’s take a look: 

Given the astounding nature of this video, I am sure you have guessed that this message was in fact not delivered by the real Queen. Despite knowing this, we can’t stop ourselves and wonder if our eyes have deceived us, even if it was for a split second. Such videos are popularly known as DeepFakes and they are taking the internet by the storm.

Observing this increasing popularity of DeepFakes, Turning Magazine is starting this new year with a blog mini-series titled, “Seeing was Believing”. The aim of this series is to raise awareness about DeepFakes and how one can fight against them. In Part 1 of this series, I will introduce an up-and-coming area of research called Media Forensics. In Part 2, we will dive into DeepFakes, understanding their origin and its impact over us. 

Introduction to Media Forensics

Every year key innovations dominate the tech space which generate a significant amount of hype around themselves, and 2020 was no different. From recent advances in natural language understanding with OpenAI’s GPT-3 model to the groundbreaking research in protein folding with DeepMind’s AlphaFold [2], it’s needless to say that this year has seen some breakthroughs, especially in AI. But not every advancement in AI is for the good of humanity. We are living in an era of misinformation, fueled by fake news and media content. And sadly, AI has played a huge role in this. With the rise of DeepFake technology (Figure 1), all of us are left to wonder: is “seeing is believing” even relevant in today’s time?

Figure 1. The growing interest in DeepFake technology via Google Trends (keyword = “deepfake”)

Rise of Fake Media Content

Fake media content, images or videos, go back as far as the existence of digital media. In its simplest form, fake images or videos are nothing but changes made to its original (real) version which results in a depiction that is not true in reality. This fundamental concept has not changed over time, only the techniques and tools which are employed to make these changes as realistic as possible. With this point of view in mind, we should consider all the existing computer-generated movies or movies which use CGI/VFX as fake media, too. Well, not exactly. In today’s socio-political environment, the topic of fake media is more nuanced. We need to consider other aspects such as its potential to spread misinformation and potentially cause harm to individuals in society. 

Given the rise of AI-assisted generation of fake media content, there are typically two major categories for fake images or videos: CheapFakes and DeepFakes. We already know that DeepFakes are a recent advancement in AI which use deep neural networks (specifically, Generative networks) to create manipulations in original (real) media and they are getting more realistic than ever. And then there is the other category: CheapFakes. Although quite recently coined, this type of fake media has been around for a very long time. This category of fake media creates manipulations through conventional techniques such as Adobe Photoshop or even MS Paint. If you cut out a celebrity’s face from a newspaper and stick it on someone else’s photo in a manner which makes it appear realistic, you have made a CheapFake!

Both types of fake media share an equal potential to cause serious damage to our society and democracy by spreading misinformation. Nina Schick, a leading author in the realm of DeepFakes, discusses this very issue in the MIT Technology Review. She highlights that the year 2020 belonged not just to the DeepFakes, but also the CheapFakes [3]. However, we are not completely helpless in this fight against misinformation. The creation of fake media brought about the creation of a group of researchers who develop new methods and technologies to detect and monitor the usage of fake media for malicious purposes. This niche area of research came to be known as Media Forensics. Its relevance in today’s society only increases day after day alongside the usage of fake media on the internet, especially on social media platforms, to harm or defame an individual.

Impact of AI in Media Forensics

If I could boil down the crux of media forensics, I will claim that this area of research aims to assess the fidelity of any media content in focus. For added clarity, whether a certain image or video is portraying the truth and not misleading its audience. Given the digital nature of media content, pixels are the basic building blocks of images and also videos (we ignore the audio component for now). This unequivocally means that to assess the fidelity of digital media, we must investigate the pixels of an image or a video in question. Now as the tools and techniques improve to manipulate pixels in such a way that it appears true to its audience, so will the technology to detect and monitor them.

In a pre-AI era, such manipulations were possible via software tools such as Adobe Photoshop. Several operations like copy-moving or slicing were possible where elements of a source image could be transferred to a target image, thus creating a fake image portraying a lie. As real images and videos are captured through a camera, pixel-level manipulations are possible only in post-processing, i.e. after an image or video is available in its digital form. Such manipulations create, what researchers in media forensics call, artefacts. Artefacts are discrepancies in fake media which can be exploited to design detection technology. But with the use of AI, these artefacts are getting harder to detect through conventional methods. The reason behind this is that conventional methods use manual, hand-picked features which can represent a certain discrepancy. 

Now, if such discrepancies evolve to be more intricate, the process of hand-picking features inevitably fails. Hence, media forensics quickly adapted to this paradigm shift in fake media generation and adopted the use of AI to develop so-called DeepFake detection techniques to regain the ability to exploit discrepancies left behind by DeepFakes. Luisa Verdoliva recently published a survey paper which provides a brilliant overview of the ongoing research in media forensics, with more focus on DeepFake detection [4]. It is highly recommended for any reader interested in knowing more about this research area.  

Figure 2. A growing amount of research into DeepFakes (generation as well as detection) via Web of Science (keyword = “deepfake”)

Future of Media Forensics

This era of misinformation has just begun. With new emerging technologies and social media platforms, we have to accept the fact that anything we now see online should be viewed with a high level of skepticism. But like I mentioned before, we are not helpless against this threat. Figure 2 shows a heuristic representation of academia steadily increasing their contribution to media forensics. It’s only a matter of time before major tech companies and government bodies fund more research into this field. It might seem like a dark age for truth but there’s definitely a bright future for media forensics.

Note to Reader: This article was Part 1 of “Seeing was Believing” Blog Series. Part 2 will cover a more in-depth discussion about DeepFakes: how they originated, how they are used in the world and more!


[1] Deepfake queen to deliver Channel 4 Christmas message. BBC News. 23 December 2020.

[2]  Dormehl, L. (2020), A.I. hit some major milestones in 2020. Here’s a recap. Digital Trends. 

[3]  Schick, N. (2020), Don’t underestimate the cheapfake. MIT Technology Review.

[4] Verdoliva, L. (2020). Media forensics and deepfakes: an overview. arXiv preprint arXiv:2001.06564.

The Perceptron

By now, Deep Learning and Machine Learning have become synonymous with Artificial Intelligence. Neural networks seem to be what flour is to any baker. But it has not always been like this. In fact, there has been a period of time, where neural networks were considered useless for AI applications.

This post is part of “The history of AI”, which is a blog post series about scientific milestones and noteworthy people. If you are new here, check out the other posts, starting with “What is AI and where did it come from” and the blog about the big question “Can machines think?”.

With this post, I will introduce you to the perceptron. In its core, the perceptron is the smallest building block of any neural network. And also a good way to start, if one wants to understand neural networks. This blog post is also accompanied by a Google Colab Notebook. The perceptron implementation is quite straightforward and might help you understand the principle better. Play around with the code a bit to understand the individual parts. Of course, you will need to know at least some basic Python to understand the code. But the blogpost itself can be understood without any coding knowledge whatsoever.

The perceptron

The perceptron is an example of a binary classifier. It can learn to differentiate between two distinct classes. These classes need to be linearly separable. This means that if you would plot each of your data points on a graph, you need to be able to draw a straight line which separates the two classes. If that is not possible, the perceptron is not the right classifier for this type of problem (disclaimer: it still might be by using polar coordinates. But that goes beyond the scope of this article).

To fulfil all cheesy Christmas cliches with this post, let’s assume Santa wants to have a program telling him which child will deserve a present and which not. Of course, this is dependent on two different factors:

(1) naughtiness

(2) the cookies they placed on the plate for Santa.

Conveniently, this data is linearly separable. Our two classes are “gets present” (1) and “no present” (-1), so we can also check the binary classifier criterion.

The goal

Before we start looking at how we get a classifier for our Santa-problem, we should have a look at what we are trying to achieve here. Our data is linearly separable, which means that we are able to draw a straight line through the plot dividing the two classes. And this line is exactly what we want to get from the classifier. We need a function in which we can input a child’s naughtiness and cookie score and see if the child will get a present or not.

The setup

The perceptron requires a number of input values:

  • Φ(x),
  • a set of initial weights ω0;
  • and a learning rate.

The input values are the coordinates of our data points. For our example, this corresponds to naughtiness x1 and the cookie-count x2. Additionally, it is common to add a bias term x0 = 1 to the input weights, resulting in an input vector that has one more number than your amount of input values for one data point.

The initial weights are the first guess of what the function may look like. These come in the same form as the input vector with exactly one value more than the number of input values. You can use any value here. Some starting points may be smarter than others, but you will not always know. We’ll be using (0,0,0) here.

Lastly, we need a learning rate. If you are not familiar with learning rates, don’t worry and stick around. It will be much easier to understand when going through the algorithm.

The algorithm

From a broad perspective, a perceptron tries to find the function which correctly divides the data into its two classes by iteratively adjusting the function. Considering one datapoint per iteration, the algorithm checks if this specific point is correctly classified, by predicting its class with the weights. If so, it continues to the next point, keeping the current weights. If not, it updates them with the Perceptron Update Rule.

Okay, let’s add some math to this.


To predict the class, we need two formulae.

As a first step, we calculate the dot product between the weights and our input vector. We do this by just multiplying each element of the input vector to its respective element of the weight vector (e.g. the first number in the input vector times the first value in the weights vector). And then we only have to add the resulting values.

If we were to take our first child with a naughtiness score of 2 and a Santa-cookie-rating of 1, the calculation would be the following:

Unsurprisingly, we end up at 0. If we plug this into the second function, we get f(a) = 1. But this is not right! This child was not meant to get any present (I mean, cookie rating of 1? Santa does not want to return there surely). So we need to update our function.


To update, we have to adjust the weights we’ll be using from now on. We do this with the following function:

ωk+1 = ωk + η × Φ(xn) × tn

There are a couple of symbols here, let me quickly explain.

K is the current iteration, making ωk+1 the new weights we are just calculating and ωk those which just got us a wrong prediction. We have already seen η, which is the learning rate, and Φ(xn) is still our input vector. tn is the class that would have been correct for the current x.

Here you can see what the learning rate is for. It controls the impact our update has on the new weights. High learning rates cause big updates, small learning rates cause a smaller update.

We are now left with ωk+1, which will be the weights for our next iteration. We take the next data point, plug it into the formula from above, using the new weights:

And … wrong again. This child should have gotten a present. So we update our weights again

and use them for iteration 3:

and finally, we get a correct prediction.

But this does not mean that the next will be right again. This cycle has to continue until we (1) find a set of weights that does not change after we have used it for every single data point (so if we have found weights that predicts correctly for each of the children whether or not it will get a present) or (2) if we reach a certain stopping criterion. A stopping criterion could, for example, be a maximum number of iterations we want to loop through before stopping.

I used the Santa-helping-tool (Check out the Google Colab Notebook) to calculate the final weights:

ω = (0, -0.1, 0.1)

We see that there was no change anymore after we updated ω the last time.

Now, what does this mean for Santa? Well, all he has to do now is take the data he has collected on every child and plug it into the formula from above with our weights.

The output will tell him if the child deserves a present or not.

We can transform the weights into a nice function and draw it into the graph from above to see what actually happened:

As the graph shows, our weights gave us the line which nicely separated the kids who will get a present from the poor suckers who won’t get any.

The perceptron and neural networks

Up to this point, it is maybe not really obvious how the perceptron and neural networks tie together. This becomes more apparent when looking at this graph:

What you see here is exactly what we did. The dot product of the input point and the weights are exactly the same as multiplying xn by ωn and summing all resulting values. In the final step, the sum is converted to an output value of either 1 or -1.

Going back to our Santa-example, x1 is our bias term of 1, x2 the child’s naughtiness, and x3 the cookie rating. The output, you will have guessed it, tells us whether or not the child will receive a gift.

If you are familiar with neural networks you see how this is what happens in neural networks on a fundamental level

For all those who are not, let’s have a look at this graphical representation of a simple artificial neural network (ANN). You can see that it is just a more complex perceptron, using the output of one perceptron as the input of another.

The history

Despite its rather simple mechanisms, the history of the perceptron is what makes it so interesting. The perceptron was first introduced by Frank Rosenblatt in “Principles of Neurodynamics” (1962). It combines two important works: McCulloch & Pitt Neuron and Hebb’s Rule.

The McCulloch & Pitt Neuron is a simple binary neuron that is able to do logical operations. It takes 1 or 0 as an input and performs Boolean operations with it. However, it does not have any weighting of the input nor is it able to learn as the perceptron does with the perceptron update rule.

The learning part of the perceptron comes from Hebb’s Rule. It describes a learning rule, which is commonly summarized as “Neurons that fire together, wire together”. The weights in the perceptron can be seen as the neural activity. Only if the activity reaches a certain threshold, the neuron fires.

Let’s take a quick step back to our Santa example. Here, “firing” means we get an output of 1 and a Santa-helper-elf jumps up to wrap a present. But in order for this to happen, the dot product of the child’s scores and the weights need to be bigger or equal to 0. Otherwise, the output of f(a) (remember, a is the dot product and f the second function for the prediction) will be 0. So our dot product needs to reach the threshold of 0 in order for our activation function (f(a)) to fire.

Let’s go back to 1962. Rosenblatt was extremely confident about his invention and claimed that it could solve any classification problem in finite time. Not all followed his enthusiasm. In 1969 Minsky and Papert published a book called “Perceptrons: An Introduction to Computational Geography”. It discusses the main limitation of the perceptron: it can only find a linearly separable function.

In our Santa-example the data was linearly separable (what a coincidence…). But not all, in fact, most real-life classification problems are not linearly separable. Minsky, therefore, concluded in the book that no significant advances are to be expected from studying the perceptron. This claim would put the research on perceptrons and neural networks into hibernation until the ‘80s.

Fast forward to now: neural networks are a fundamental part of today. This can be explained by an oversight of Minsky. Yes, one perceptron is not able to classify non-linearly separable data. However, if you connect several perceptrons into a neural network, it is able to solve complex classification problems and learn non-linear functions. This, together with backpropagation and hyperparameters (stories for another time), was what was needed to revive neural networks and give them a place in the history of AI.

The moral of the story (to stay in the Christmas spirit of this post)? You should better bake good cookies, it can fix a lot.

Diving into the Wave of the AI Job-Revolution

The AI revolution is here. What has been talked about for the last two decades is finally being realised. It’s here, with all its hype, and it is here to stay. Every company wants the proverbial slice of pie and wants to ingrain the use of Artificial Intelligence (AI) in its products. There has been a massive amount of funding in labs and companies across the world for developing new AI research. With this, fears have arisen about AI taking over the world as well as our jobs. There are statements like “Millions of jobs will be replaced by AI”. Or will they? To answer this question, we have to go back a few decades.

There is no reason anyone would want a computer in their home.

Key Olsen, founder of Digital Equipment Corporation, 1977.

Somehow, the microcomputer industry has assumed that everyone would love to have a keyboard grafted on as an extension of their fingers. It just is not so.

Erik Sandberg-Diment, New York Times columnist, 1985.

Both of these statements have not aged well at all. This is not to throw shade at them, but to show how predicting the impact of any new technology is extremely hard.

Every new technology, especially in the last two and a half decades, has brought about a revolution in the job market. Yes, many jobs were permanently lost. However, each new technological revolution also brought about completely new jobs which were hitherto unknown. For all the jobs lost because of the internet, like media distribution stores, encyclopedia salesman, librarians, phone-book companies, there have emerged new jobs like social media managers, web developers and bloggers. It gave so many people a platform to sell their products and services, as well the ability to reach out to a huge number of people faster. In fact, it also transformed some of the pre-internet jobs and boosted them just because of social media’s easy reach. Similarly, the smartphone revolution killed off the need to use so many devices like a separate radio, mp3 player, point-and-shoot cameras, physical maps, and even wristwatches; smartphones have fast become a behemoth contributor towards the economy, with mobile tech generating $3.3 trillion in revenues in 2014 [1]. Jobs like social media influencer and online tutors on Youtube and various educational websites are available at fingertips thanks to smartphones.

Warehouses in Guanajuato, Mexico

AI development is extremely fast; however, the background frameworks it needs has not grown at the same pace. It also has applications in the real world, and thus, paints a target on its own back for criticism. There is a huge clamour about how it will replace humans in jobs in many journals and articles. While some jobs will be lost due to AI, it will also produce new opportunities. According to the Autonomous Research Report in 2018, 70% of the front office jobs like tellers, customer service representatives, loan interviewers and clerks will be replaced by AI technologies like chatbots, voice assistants and automated authentication; however, according to Accenture, there will be a net gain in jobs among companies using AI, as well as a net gain in revenues. They claim that AI will create new roles like explaining the deployment of AI technologies, which would still be done by humans. The reasoning they give for the latter is that AI will help people with advice at investments and banking, an improvement on the human agents. In the same article from American Banker [2], the senior VP of First National Bank of Wynne in Arkansas argues that people are replaceable, AI is not. That is, if a person makes mistakes repeatedly, they can be replaced by another human more capable of doing the job; a malfunctioning AI, however, cannot be fired, it needs to be shut down, and replaced with a human, which makes companies wary of using AI. Research by the Royal Bank of Canada states that humane skills like active listening and social perception will help prospective job applicants complement AI technologies, rather than compete with them.

These discussions come from management employees, but that doesn’t mean that we have to agree to them necessarily. They state that there will be more jobs available. However, we also need to look at what kind of jobs AI will create. If we look at job markets as a whole, the trend would be that indeed, new jobs will be created. If we decide to look closely, however, we can see that the kind of jobs AI will take away are the main source of livelihoods for many working-class people. A taxi driver would not be too amused to see an autonomous truck taking away their passengers and wages, especially if they’re struggling to pay their rent each month with a job. I would also respectfully disagree with the aforementioned VP of the Bank; the sentence ‘humans are replaceable’ is not a good look with the current job market prospects. We should be protecting our workforce with the help of technology, not replace them with more technology. We also need to consider that people need to adapt to the new technologies; some might not have the resources to do so. Do they then try to find different jobs, or risk being left behind with the new revolution? 

There are discussions in companies like Accenture, that AI will maximise profits for companies, which will lead to maximization of growth and in turn more employees as companies join more global markets. With even more widespread use of AI, it will be necessary to regulate the ethics behind the usage. This will necessitate employing humans to monitor the decisions taken by AI tools. An example of AI being limited to a single purpose job is given in this New York Times article [3], which quotes an important MIT report [4], “A finely tuned gripping robot can pluck a glazed doughnut and place it in a box with its shiny glaze undisturbed, but that gripper only works on doughnuts, it can’t pick up a clump of asparagus or a car tyre.” A general-purpose AI is still a few years away. AI is normally designed as a tool to help humans do better at their jobs. In fact, ASAPP, a New York-based developer of AI-powered customer service software trains human call centre representatives with the help of AI to be more productive, rather than replacing them with AI [5]. Combining human jobs with AI seems to be the best way to go forward, delicately achieving a balance between productivity and human ingenuity.

The effects of AI will be seen in almost all industries. Since AI is a tool which can be applied in so many industries, there is a huge push to apply AI in various domains like medical imaging, neuroscience, business analysis to even sport science. The very essence of AI is such that it pervades all kinds of job markets without discrimination. It is going to change the landscapes of many job markets. Whether you are a taxi-driver, or a doctor, or an assembly-line worker, it is going to affect your job. The hardest hit will be people who are in no position to learn the new skills needed to complement AI entering our lives. This is because as of now, AI is great at doing single purpose repetitive tasks efficiently. AI is not a messiah which will take us to a proverbial promised land, nor is it a weapon of mass destruction wiping out the planet, it is somewhere in the middle, and like with every new technology, we need to adapt. In fact, the balance is more towards the negative. Have we thought about the sheer displacement of people with more and more jobs utilizing AI? Is AI, and especially complex deep learning models, even necessary in tasks? We need to ensure that we’re not just falling victims to following trendy buzzwords and trying to incorporate the latest technology in our services. We have lots of experience in dealing with revolutions, we need to find a way to deal with this latest revolution and ensure it becomes close to the messiah. Thankfully, we still are some years away from AI pervading all job markets, so we can make concrete plans to handle smoother takeovers. 

To tackle the wave of AI revolution and ensure that we’re not left behind on the kerb while AI takes our jobs away, we need to keep reinventing ourselves. Yuval Harrari, in his article the “Reboot for the AI revolution”, sums it up beautifully. He says, “…humankind had to develop completely new models [to adapt to the Industrial Revolution] — liberal democracies, communist dictatorships and fascist regimes. It took more than a century of terrible wars and revolutions to experiment with these, separate the wheat from the chaff and implement the best solutions.” [6] He then adds that we need a working model to overcome the challenges introduced by the new technology. These words may prove to be seminal in the coming years as AI starts disseminating in all job markets. We may need to turn the wheel- or reinvent it completely, as we adapt to the new revolution.

This article delved into how AI is affecting the job industries. The next follow-up article, “Riding the Wave of AI Job Revolution”, of the same series will cover how we can improvise and adapt to this challenging situation. 


[1] Dean Takahashi, “Mobile technology has created 11 million jobs and $3.3 trillion in revenues”, (2015),

[2] Penny  Crosman,  “How artificial intelligence is reshaping jobs in banking”, (2018),

[3] Steve Lohr, “Don’t Fear the Robots, and Other Lessons From a Study of the Dig-ital Economy”, (2020),

[4] Elisabeth Reynolds, David Autor, David Mindell,  “The Work of the Future: Building Better Jobs in an Age of Intelligent Machines”, (2020),

[5] Kenrick Cai Alan Ohnsman, “Meet The AI Designed To Help Humans, Not Replace Them”, (2020),

[6] Yuval Noah Harari, “Reboot for the AI revolution”, Nature News550.7676(2017), p. 324.

A Post-Pandemic Outlook on the Future of AI Research

The year 2020 is most likely to play an instrumental role in how Artificial Intelligence (AI) will evolve in the years to come. It won’t just be how it is perceived in the world’s view (be it good or bad), but also how the AI research community will change. The world right now is in a state of crisis, the COVID-19 pandemic, and we are yet to observe the end of this crisis and any lasting repercussions which might follow.

Where Are We?

A myriad of reports and articles are published on a daily basis which discuss the socio-economic implications of the COVID-19 pandemic, for example [1]. Additionally, this pandemic has brought forth the rise of AI in ways which are astonishing as well as disconcerting. Since the outbreak of COVID-19, AI has significantly contributed to the healthcare industry [2] through applications such as predicting new cases, drug discovery and more. But the pandemic has also inadvertently fuelled the world of surveillance [3] we live in today. Such a rapid development and controversial usage of AI raises concern regarding our privacy and security in the future. Yuval Noah Harari, an Israeli historian and the author of Sapiens (2011), discusses [4] the future of humanity as well as poses some daunting questions regarding our response towards the expected impact after COVID-19. The article discusses how governments are using the pandemic as an excuse to abuse the state of emergency by forcing upon “under-the-skin” surveillance and essentially, setting up a totalitarian regime.

It’s quite natural to find yourself in a state of confusion as to how this grim description of a post-pandemic world connects to the future of AI research. But it should be considered vital and even instrumental in shaping the future of AI research. 

We find ourselves in a thicket of strategic complexity, surrounded by a dense mist of uncertainty

– Nick Bostrom, Superintelligence: Paths, Dangers, Strategies

What Was and What Is?

The evolution of any industry is ever-changing, but certain trends and historical patterns have enabled us to forecast their future, at least to an extent. The AI industry is not stand-alone but rather closely associated with a multitude of other industries, such as healthcare, business, agriculture and more. There has been a significant rise in the adoption of AI by such industries in just the last decade and this has increased the need to forecast AI’s impact in the future and how it will inevitably shape our society. Books like Superintelligence (Nick Bostrom, 2014) and Human Compatible (Stuart J. Russell, 2019) have discussed this and presented their predictions. Stakeholders who fund the advancement of AI make decisions based on such predictions. This in turn drives the AI research community. 

But with the crisis we face today, the status quo is going to change. Seeing how AI research has been conducted since the start of COVID-19, the focus should not only be on the outcomes of a certain study but also the way such studies are conducted. For the sake of simplicity, this article focuses on AI in healthcare but also emphasizes that the presented arguments apply to AI research in general. 

In Focus: AI in Healthcare

The astounding increase in the number of cases during this pandemic pushed the AI community, both academia and industry, to divert their resources in providing any form of support which could be essential in this fight against the virus. AI research in medical imaging [5], specific to COVID-19, has enabled researchers as well as doctors to train and deploy predictive models which can contribute towards patient diagnosis. There is also an increase in drug discovery research which will help researchers to identify vaccines which can be tested and distributed to everyone. AI in drug discovery has been very favourable when compared to conventional clinical trials as AI speeds up the process of developing new drugs as well as drug testing in real-time [2].

So What’s the Issue?

Once the crisis is over, companies are expected to invest even more into AI research [6] and government bodies are expected to increase their involvement [7] to use AI to plan strategies against future pandemics as well as empower other industries which could benefit from AI. Dana Gardner [8] discusses in his podcast that the data collected over the pandemic will be a key factor of how AI will shape the post-pandemic world.

Despite the extensive amount of AI research in such a short time, they pose an inherent flaw of being products of AI black-box systems. Such systems spit out numbers and leave us humans to derive meaning out of it. A majority of research in AI is purely focused on the final results (such as how they perform on a benchmark dataset) rather than how they arrived at it, especially the research conducted by AI start-ups. Even though this attitude was due to the urgency of this crisis, it has been around for quite some time. There are multiple applications such as image recognition which require only numbers to make sense but when we enter human-centric applications such as healthcare, mere numbers are simply not enough. A high accuracy model is not guaranteed to attain high efficacy in such areas and we need to ask ourselves: Will this style of research, a pursuit to train a model with highest accuracy in the shortest amount of time possible, continue in a post-pandemic world? 

Explainable AI, commonly referred to as XAI, deals with research into developing interpretable systems which can provide explanations for its decision making in any given scenario. This is a step ahead of the current “black-box” AI systems.

This argument brings forth Explainable AI, an area of research which is still in its infant stage. What has been already done cannot be changed, but we can and should learn from these past few months. The amount of data in the future, especially in industries such as healthcare, is going to explode and how we handle this data and design new AI systems with it should be the crux of future AI research. Until this pandemic, the questions posed by most AI critics regarding the black box nature of high accuracy models were mostly hypothetical, for example, the trolley problem. But now, the decisions made by these AI systems during the pandemic affected very real humans. We are clearly outside the hypothetical debate and the need to tackle this issue is of utmost importance.

How Do We Move Ahead?

Besides the importance of AI systems being more explainable, at least to the involved stakeholders, new policies and legislations are required which can dictate the plan of action of AI research during a crisis mode, such as COVID-19. Such policies around AI research will introduce important standards and guidelines, similar to other areas like healthcare and environmental sustainability, to tackle problems which are not only ad hoc but also consider the societal as well as ethical implications of the research in question.

[1] Nicola, M., Alsafi, Z., Sohrabi, C., Kerwan, A., Al-Jabir, A., Iosifidis, C., … & Agha, R. (2020). The socio-economic implications of the coronavirus pandemic (COVID-19): A review. International journal of surgery (London, England), 78, 185.

[2] Vaishya, R., Javaid, M., Khan, I. H., & Haleem, A. (2020). Artificial Intelligence (AI) applications for COVID-19 pandemic. Diabetes & Metabolic Syndrome: Clinical Research & Reviews.

[3] Yuan, S. (2020). How China is using AI and big data to fight the coronavirus. Al Jazeera.

[4] Harari, Y. N. (2020). The world after coronavirus. Financial Times, 20.

[5] Bullock, J., Pham, K. H., Lam, C. S. N., & Luengo-Oroz, M. (2020). Mapping the landscape of artificial intelligence applications against COVID-19. arXiv preprint arXiv:2003.11336.

[6] Global companies will invest more in AI post-pandemic. Smart Energy International, Sept 18, 2020.

[7] Jad Hajj, Wissam Abdel Samad, Christian Stechel, and Gustave Cordahi. How AI can empower a post COVID-19 world. strategy&, 2020.

[8] Dana Gardner. How data and AI will shape the post-pandemic future. July 7, 2020.

Can Machines Think?

The rise of digital computers motivated scientists to reason about the possibilities and limitations of such machines. In 1936, Church and Turing independently proved that there are problems which are not possible to solve with a digital computer. Other questions remained unsolved, one of which still concerns society today: Can machines think?

In 1950, Turing published “Computing machinery and intelligence”. In it, he proposed a practical test which is to this day synonym with testing machine intelligence: the Turing test. But the paper offers more. It gives an overview of common objections against Artificial Intelligence. Many of them are still brought up today. Others get weaker and weaker with every new challenge computers tackle. And it is the paper which makes Turing one of the founding fathers of Artificial Intelligence, as he proposes to teach machines just as we would teach a child.

“Computing Machinery and Intelligence” is a seminal paper, incredibly ahead of its time and most of all: still worth a read. As it is one of the earliest works in computer science and Artificial Intelligence, it is approachable without much background knowledge. But it also touches so many aspects that even the more AI-advanced readers will be blown away by. This blog post will only give an overview of the paper; I can only encourage everybody to pick it up and read it for yourself.

The context

Before we dive into the paper itself, it is worth considering the time the paper was published.

In 1950, the term ‘computer’ most commonly described humans, women mostly. It was a job description for somebody who did computations, using computation tables. Human computers reached their peak during the world wars. But with the rise of digital computers, their downfall began. Digital computers soon outperformed human computers and in 1954 most humans were already replaced by digital computers.

Turing’s paper was published just in this transition time. Two years prior to the first electronic computer, the Manchester Mark I ran for the first time. But it was still humans who did the most of the computations. His paper, therefore, contains several sections explaining the notion of a digital computer to the reader and describing it as a machine following a human-computer rule book.

The Manchester Machine, commonly referred to as Manchester Baby
Can machines think?

To answer this question, the terms ‘thinking’ and ‘machines’ would have to be explained. But instead of doing so, Turing proposes to replace the question altogether. Instead of asking whether machines can think, we should be asking if the machine can succeed to win in the imitation game.

The imitation game

The imitation game requires three players, a woman (A), a man (B) and an interrogator (C). A and B are separated from C in a different room. C’s objective in the game is to figure out which of the two players (X and Y for him) is female. C may ask any question to A and B, which they have to answer. A tries to fool the interrogator by either pretending to be male or by suggesting that B is actually female. B on the other side will try to help C to correctly identify the woman. All communication is done in written form so that the voice will not give the sex of X and Y away.

Turing then proposes to replace A by a computer. Is there an imaginable digital machine which is able to fool a human interrogator as often as a woman is able to fool a human judge that she is a man? In Turing’s eyes, this is the question we ought to ask.

What does it take to win the imitation game?

From today’s perspective, one can see that a number of AI sub-disciplines would be needed for a computer to participate in the imitation game.

First of all, it would need Natural Language Processing to communicate with the interrogator. Some form of knowledge representation and automated reasoning would be needed to reason about the questions asked and provide satisfying answers. Lastly, machine learning will be necessary so that the machine can adapt to new circumstances.

Taking a step back here, Turing wrote his paper when even the notion of a digital computer was foreign to the majority of people. I assume today’s reader is familiar with those machines, but there is more to Turing’s excursion to computing.

On Computing

You might have heard about Turing machines. In a prior paper, Turing proved that every definite-state machine can mimic every definite-state machine’s behaviour given enough resources. All computers are basically the same. This means the computation unit in your microwave could, with the right code and enough resources, do the same computations as your fancy MacBook pro.

Following this logic, if we can imagine one Turing complete machine (as digital computers are) to succeed in the imitation game, we will have answered the question for all digital computers. And Turing believes that such machines are possible.

Contrary views

Turing anticipates that not everyone will agree with either the replacement of the original question or with the possibility that machines can think. He addresses theological, mathematical and even objections from extrasensory objections (he actually takes ESP more serious than religion).

A few quite interesting ones are the argument from consciousness, the objection of various disabilities and the Lady Lovelace objection.

The argument of consciousness

For this argument, Turing quotes Professor Jefferson who wrote in 1949: “Not until a machine can write a sonnet or compose a concerto because of thoughts and emotions felt, and not by the chance fall of symbols, could we agree that machine equals brain—that is, not only write it but know that it had written it. No mechanism could feel (and not merely artificially signal, an easy contrivance) pleasure at its successes, grief when its valves fuse, be warmed by flattery, be made miserable by its mistakes, be charmed by sex, be angry or depressed when it cannot get what it wants.”

Turing takes this quote as a justification for his imitation game. In his eye, the question-answer approach of the imitation game does allow the interrogator to ask the machine questions about art, feelings and perceptions. As stated before, it does not matter if the machine is actually able to feel the way we humans feel. It only matters if it can make humans believe it feels and perceives as we do.

The argument from various disabilities

This argument is a particularly interesting one, as it gets weakened with every advancement in the field of AI. The argument claims that a machine will never be able to do X, where Turing states X to be:

Be kind, resourceful, beautiful, friendly (p. 448), have initiative, have a sense of humour, tell right from wrong, make mistakes (p. 448), fall in love, enjoy strawberries and cream (p. 448), make someone fall in love with it, learn from experience (pp. 456 f.), use words properly, be the subject of its own thought (p. 449), have as much diversity of behaviour as a man, do something really new (p. 450). (Some of these disabilities are given special consideration as indicated by the page numbers.)

Page numbers refer to the original publication in Mind, Volume LIX, Issue 236, October 1950e

From today’s perspective, we see, however, that machines proved they can do some of the X’s. It seems less likely that they won’t be able to achieve the others. They will just never be able to be human.

Of course, there is also the mathematical objection. There are certain things which are mathematically impossible for a machine to solve. Turing claims that humans are also not flawless and frequently make mistakes. So even if a computer cannot solve the Entscheidungsproblem, they could be programmed to give a random answer and pretend to have done a mistake in case they were mistaken.

Lady Lovelace’s objection

Lady Lovelace is commonly considered to have been the first programmer. She wrote algorithms for the “Analytical Machine”, a mechanical computer. In her memoir written in 1842, she states “The Analytical Engine has no pretensions to originate anything. It can do whatever we know how to order it to perform”.

But what if we tell the machine to learn and originate? In the refutation of this argument, Turing introduces learning machines, which he dedicated a whole chapter in his paper.

Learning Machines

Turing admits that he cannot make a sufficient refutation of Lady Lovelace’s objection. But he proposes a machine which would a) give a solution to the imitation game and b) address Lady Lovelace’s objection.

In his eyes, creating a machine that does well in the imitation game is merely a question of the program. He believes that the computational requirement will improve in the future, that they will not restrict the aim.

The real challenge will be to write a program that can fool a human interrogator. He proposes to go back to the roots of what shapes the intelligence and ability of the human mind:

  • the initial state of the mind at birth
  • the education the human obtains
  • and the experiences made

Instead of trying to write a program, mimicking an adult brain, we should mimic a child’s brain: a blank notebook with only a couple of inference mechanism, which allows the machine to learn.

The child-machine would obtain an education by a teacher as a normal child would do. Turing believes that this education process would probably take as long as for a human child. It would get rewarded if it produced good results and answered correctly. It would get punished for disobeying and wrong answers. He basically proposes to implement a machine that can learn. Or slightly reformulated, he proposes Machine Learning. Or for the real nerds: Reinforcement Learning.

Turing expects that the teacher will not always know what exactly happens in the child’s mind, much as a teacher cannot always tell what the human pupil thinks. And if we cannot tell what happens inside the machine when it learns, how could we possibly have told it to do exactly this action? So if we can make a machine learn as Turing proposes, it would do something we did not specifically tell it to do, which would refute Lady Lovelace’s argument.

From the Imitation Game to the Turing Test

What Turing proposed as the imitation game has in some form become a practical approach for measuring intelligence. Turing believed that by the end of the 20th century, nobody would be contradicted when stating that machines think.

We are not quite there yet. No machine has officially passed the Turing test. Eugene Goldman, an AI pretending to be a Russian boy has fooled judges by claiming that his bad communication skills are due to his inferior English level. But expert’s opinion is that it was cheating the game.

As Turing said, winning the imitation game is a matter of the right program. And the AI world did not focus on producing the right program for passing the Turing test. The focus lays on solving real-world problems. We have computers that can detect cancer, personal digital assistants, robots that help improve the social skills of autistic children and even sex robots.

Turing asked, whether there is any imaginable machine which would do well in the imitation game. We are not there yet. But given the advances in the field of AI, most people would be able to imagine such a machine.

“Computing Machine and Intelligence” helped to lay the foundation of what is today known as Artificial Intelligence. It is yet another example of Turing’s genius.

You can find the original paper here: Give it a try!


Turing, A. M. (1950). I.—Computing Machinery and Intelligenec. Mind, LIX(236), 433–460.

Russell, S., & Norvig, P. (2009). Artificial Intelligence: A Modern Approach (3rd ed.). Prentice Hall Press.

Oppy, Graham; Dowe, David, “The Turing Test”, __The Stanford Encyclopedia of Philosophy (Fall 2020 Edition)__, Edward N. Zalta (ed.), URL=

Grier, D. A. (2001). The human computer and the birth of the information age. Joseph Henry Lecture of the Philosophical Soc. of Washington, 42-43.

Warwick, K., & Shah, H. (2015). Human misidentification in Turing tests. Journal of Experimental & Theoretical Artificial Intelligence, 27(2), 123–135.

Ewald, W. and Sieg, W., 2013. David Hilbert’s Lectures on the Foundations of Arithmetic and Logic 1917-1933. Springer Berlin Heidelberg.

The face recognition complication

Artificial Intelligence is a subject which is often, inaccurately, talked about in terms of ending the world.  Several films have a theme of a rogue AI destroying the whole world and threatening the existence of humanity. They show a very physical manifestation of AI; however, the problem is much more personal and all the more mordacious. Our faces are collected and are being used to build increasingly powerful, yet error-ridden and human bias induced AI models. A recent survey found that on average, a Londoner is captured on 300 security cameras a day [1]. With this ominous beginning, its time to talk about a specific form of dangers posed by AI- facial recognition based discrimination especially among minority groups.

There has been centuries of oppression of minority groups by humans, it now seems like AI is also adding to it. Amazon created a bot to help with resume sorting, keyword searches based on skills and job profiles, and then recommend candidates. The caveat? The AI showed a bias against women, because it had been fed male-dominated resumes [2]. This kind of bias is an unwanted result, a danger posed by relying too much on AI to help with tasks.

Speaking of blatant discrimination, Michael Kosinski, an AI researcher, published a paper in 2017, where an AI model guessed whether a person was straight or otherwise, just by looking at their facial features [3]. The political, ethical and moral implications of this are dangerous [4], [5]. Especially in the countries where homosexuality is still a criminal offence, and AI which outs you just by looking at your facial expressions- this could set a terrible precedent to any hopes of getting homosexuality decriminalised, and could strengthen the already existing institutional bias against LGBT folks. What was even more surprising about the model, is that it was 81% accurate in correctly identifying men, and 74% for women, using just a single facial feature. With five facial images of a person, this rate increased to 91%. All of these rates are far ahead of identification done by humans.

Another major misfire by AI in face recognition is the discrimination against people of colour [6], which has countless examples of not recognising people of colour as accurately as white males. Detroit Police wrongfully arrested a man of colour using facial recognition software [7] with an inherent bias. There are countless others who are being subjugated to police searches after using AI as a facial recognition software. Another facial recognition software misidentified a person of colour behind bombings in Sri Lanka. The college student received severe abuse when his name was leaked to the public, all over a faulty output.

Joy Buolamwini, the founder of Algorithmic Justice League, an organization trying to raise awareness about the social implications of AI, used AI facial recognition models from a few leading companies like Amazon, Microsoft, etc [8]. All the models performed poorly for minority groups and women. In fact, some of the results were just shocking. Figure 1 shows how two famous black women, Oprah, and Michelle Obama, were labelled by the leading facial recognition software as “appears to be male”, and “young man wearing a black shirt”. The reasons for the underwhelming performance are explained later in the article.

a)Amazon Face Recognition                                                                                        b) Microsoft Face Recognition

Figure 1: Discrimination by Facial Recognition Software

These episodes lead to a necessary discussion about the use of facial recognition software, especially with AI development still in the nascent stage, and above all, such software being unregulated. There haven’t been laws against the use of such software yet, and there has been no proper consolidation of rules regarding the ethics of using image recognition. Yes, there are many benefits to using face recognition software, but currently, the problems outweigh them. The accuracy of any data-based model depends on the data it is fed; facial recognition software depends on the faces provided to it. We as AI researchers dive into every new technology and application we feel AI can be applied to, however, there is a need to stop and have a look around at the legal, moral and philosophical aspects of the technology we are building, before we actually start with our models.

We arrive at the same question again, Do we really need facial recognition? Is AI for detecting faces is at the right threshold to be used effectively in mainstream applications? Who is responsible if a person is discriminated against by a facial recognition algorithm? Who is to blame, when AI accidentally outs the sexual identity of a person, when they are not ready to come out yet? Humans can be taught to be less racist, less sexist and less homophobic, by education. Who is going to teach the AI? On the other hand, is the AI really at fault, being fed predominantly white faces as input to recognise who enters a building, in a cosmopolitan city? Maybe, instead of trying to get that validation accuracy of 99%, instead of trying to  focus on pushing publications, we should have a look at ourselves and ask, are we doing it right? Are we working with the right data? Maybe the fact that AI discriminates based on race, gender, colour, is that because we are letting our human biases into the dataset [9]. AI is fast becoming a proverbial looking glass, reflecting the society’s values. However, with the social media revolution, the extremities of the society are amplified and reach more people than ever. This opens a box of questions. Have we thought about the implications of what we are creating? Have we thought if the law frameworks exist to regulate AI? Have we thought of the ethical implications of the dataset we are trying to collect? So many tech companies are shutting down their facial recognition software. Amazon has decided to shutdown Recognition, IBM has shut down its facial recognition research as well, both because of human rights concerns. Instead of shutting down the research altogether, we could build more robust data to fix historical societal problems. We should focus more on building the legal and ethical foundations and understand more on how we use our data.


[1]Luke Dormehl. “Surveillance on steroids: How A.I. is making Big Brother biggerand brainier”. In: (2019).url:

[2]Julien Lauret. “Amazon’s sexist AI recruiting tool: how did it go so wrong?” In:(2019).url:

[3]Yilun Wang and Michal Kosinski. “Deep neural networks are more accurate thanhumans at detecting sexual orientation from facial images.” In:Journal of person-ality and social psychology114.2 (2018), p. 246.

[4]Sam Levin. “New AI can guess whether you’re gay or straight from a photograph”.In: (2017).url:

[5]Paul Lewis. “I was shocked it was so easy”. In: (2018).url:https : / / www

[6]Fabio Bacchini and Ludovica Lorusso. “Race, again: how face recognition technology reinforces racial discrimination”. In:Journal of Information, Communication and Ethics in Society(2019).

[7]V Steeves J Bailey J Burkell. “AI technologies — like police facial recognition —discriminate against people of colour”. In: (2020).url:

[8]Joy Buolamwini. “Arti cial Intelligence Has a Problem With Gender and RacialBias. Here’s How to Solve It”. In: (2019).url:

[9]Brendan F Klare et al. “Face recognition performance: Role of demographic information”. In:IEEE Transactions on Information Forensics and Security7.6 (2012),pp. 1789–1801

Neuromorphic Computing: Future of AI?

Autonomous Vehicles, or commonly known as Self-Driving Cars, are one of the most popular applications of Artificial Intelligence (AI) today. One can find numerous articles where a Self-Driving Car is used as an example to describe a certain aspect of AI which makes it very monotonous for readers (including me) over time. So let’s switch some gears, pun intended, and look at another application which seems quite interesting as well!

The Mars Rover Mission, which is slated to launch in July 2020, is a NASA project to explore the surface of the planet Mars, collect samples from its surface and find proof of life in the past. Well, it’s not much different from the mission of its predecessors but what is evidently different, is the technology behind this mission. The rover has made advancements in every aspect such as its sensors, communication modules and even its mechanics. But this is an AI blog, so I’ll focus on the AI aspect of this rover.  Although there’s not much information about the “brain” of the rover on their information website, there is a clear mention about the hardware specifications of the rover’s “brain”: Radiation-hardened central processor with PowerPC 750 Architecture (BAE RAD 750). 

Despite its remarkable features, it’s an upgraded version of processors originally built for traditional computing. Using such hardware for AI, although fast, does not bring out the true potential of what AI can achieve. This need demands a new type of computing hardware which can create a new generation of AI.

A Brief History of Neuromorphic Computing

Neuromorphic Computing, or otherwise known as Neuromorphic Engineering, has piqued interest in the AI community for some time now but the field itself has been kicking around since the early 1990s. The original idea was to create VLSI systems which could mimic the neuro-biological architectures that are present in the nervous system. Hats off to Carver Mead who coined the term “neuromorphic” in his 1990 paper [1]. Mead’s motivation for looking into a new approach at understanding the biological information processing systems lies in the fundamental fact that they behave quite differently from how engineers have designed electronic information processing systems, aka traditional computing. 

Mead further drives the importance of a paradigm shift in computing by demonstrating that biological information processing systems are capable of solving ill-conditioned problems in a more efficient manner than the digital systems. Thus, showing that neuromorphic systems are capable of handling complex problems and execute such computations in an energy-efficient manner. But the constant rise and success of Moore’s Law in the industry overshadowed the importance of neuromorphic computing. This caused the neuromorphic computing to go into a long period of hibernation. So, let’s just skip a few years to when things get really interesting. 

The neuromorphic computing research got an overhaul when the SyNAPSE project was established in 2008. Funded by DARPA, this project’s main goal was to explore ways of “organizing principles” which can be used in practical applications. Their various collaborations with other organizations have been covered in the ACM Publication by Don Monroe [2]. The Human Brain Project, established in 2013 and funded by the European Union, took a different route and focused on understanding the cognitive processes and conducting Neuroscience research via modelling and simulations. 

Despite the differing goals, both projects work on the physical realization of biological neurons and spikes [2] as means of transmitting information between neurons. This is drastically different from traditional computing systems which can simulate neurons and spikes, making it computationally expensive when compared to neuromorphic computing systems. Unlike traditional computing, which widely uses the Von Neumann architecture where the processing and memory units are located separately, neuromorphic computing employs a “Non-Von Neumann” architecture where processing and memory units are co-located in a neuron core on the neuromorphic chip.

Fantastic Neuromorphic Computing Systems and How to Use Them

Since the conception of Neuromorphic Computing, there have been numerous hardware implementations of neuromorphic chips. Some are purely analog in nature whereas some are purely digital along with some as a combination of analog and digital components. A survey paper by Schuman et al. (2017) [3] extensively covers the hardware implementations. See Figure 1 for a visual representation of the overview from [3]. As one can guess, there’s a lot to discuss but let’s focus on the popular hardware implementations for neuromorphic computing for now.

Figure 1: Overview of Hardware Implementations of Neuromorphic Computing

IBM TrueNorth [4] is one of the most popular digital neuromorphic implementations. Developed in 2014, it is a fully custom chip with nearly a million neuron cores and 256 million synaptic connections. This is a significant accomplishment given that an average human brain has around 86 billion neurons and around 100-1000 trillion synaptic connections [5]

SpiNNaker [6], part of the Human Brain Project, is yet another digital neuromorphic chip which is massively parallel in its computations with around a million ARM9 cores (neurons). Both TrueNorth and SpiNNaker are good digital hardware implementations but they come at a cost: energy. As the complexity and size of a neural network architecture increases, these systems use a huge amount of energy for the computations, such as learning. BrainScaleS [7], also part of the Human Brain Project, is a mixed analog-digital hardware implementation using wafer-scale implementation (like a microprocessor) along with analog components with nearly 200 thousand neurons per wafer.

Although these above systems are great at simulating neurons at a hardware level, it still requires massive space and energy to perform their computations. This is where Intel’s Loihi [8] Neuromorphic chip enters. Produced in 2017, it claims to be 1000 times more energy efficient when compared to its competition. Currently in the research phase, Intel aims to release such neuromorphic chips for commercial use in the near future. Although only having scratched the surface of this topic, it is enough to understand the extent of research ongoing in this field. But what’s the point of neuromorphic computing when current traditional computing systems such as GPUs are pushing AI to new heights?

Neuromorphic computing provides a brain-inspired computation which is biologically-plausible when compared with the Artificial Neural Network (ANN) models which are run on traditional computing systems. This is possible due to the third generation of Neural Networks, Spiking Neural Networks (SNN). The basic units of a neuromorphic computing system are like  neurons which can be connected to each other via synaptic connections. Such a network simulates a spiking neural network model that exists in biological brains. As these networks which are trained on such systems are biologically-plausible, it drives the research in AI from a probabilistic viewpoint (See Bayesian Brain Hypothesis) which attempts to solve the issues related to fundamental uncertainty and noise that is present in the real world [9].

Return of The Rover

Well, this article is almost at its end. So it’s time for some wrap-up and a  “In a nutshell” session where the Mars Rover is back in focus again. Like I mentioned before, the current Rover project uses an exceptional computing hardware but still follows the traditional “Von Neumann” like architecture to run AI algorithms for the execution of certain tasks. As this does not unlock the full potential of AI, a new form of computing is required. This is where Neuromorphic Computing would be a perfect fit. Not only does it allow simulation of neurons on a hardware level but also focuses more on biologically-plausible neural network models. Using such models is more probabilistic in nature and thus highly beneficial in real world applications where uncertainty and noise is very prominent and can not be avoided.

Intel also explains in their article [8] that tackling problems via a probabilistic approach deals with representing outputs as probabilities instead of deterministic values (which is very common in areas such as Deep Learning). Working with probabilities also creates a path for explainable and general AI which are the next big milestones in this world of AI. So, solving tasks such as maneuvering around the surface of Mars and explaining the decisions taken by the Rover to human operators back on Earth can be a reality with neuromorphic computing systems. Besides this, a neuromorphic chip is more energy efficient than a traditional microprocessor chip. 

In conclusion, the field of Neuromorphic Computing is highly relevant in today’s age where AI demands a new generation which can solve tasks with high performance and low energy consumption.

Disclaimer: Some sections of this article have been adopted from the “Neuromorphic Computing”, a masters AI course at Radboud University.


[1] Mead, C. (1990). Neuromorphic electronic systems. Proceedings of the IEEE, 78(10), 1629-1636.

[2] Monroe, D. (2014). Neuromorphic computing gets ready for the (really) big time.

[3] Schuman, C. D., Potok, T. E., Patton, R. M., Birdwell, J. D., Dean, M. E., Rose, G. S., & Plank, J. S. (2017). A survey of neuromorphic computing and neural networks in hardware. arXiv preprint arXiv:1705.06963.

[4] “The brain’s architecture, efficiency on a chip”, IBM Research Blog, 8 Feb. 2019,

[5] Scale of the Human Brain

[6] “SpiNNaker Project”, APT,

[7] “The BrainScaleS Project”,

[8] “Intel unveils Loihi neuromorphic chip, chases IBM in artificial brains”, 17 Oct. 2017,

[9] “Neuromorphic Computing”,

Sex robots are coming: why we should all play an active role in the creation of sexbots

Technology has been used for human pleasure for centuries (figure 1). At the end of the 19th century, English physician Joseph Mortimer Granville invented the first vibrator as a medical instrument for pain relief [1]. In the 1960s vibrators became a tool for women to masturbate. Ten years later, the oldest internet-connected sex toys, called teledildonics, were first mentioned by pioneer Ted Nelson.

Figure 1. Timeline of technological developments used for human pleasure.

The arrival of improved internet connection was an undeniable breakthrough for sex and technology. People could now watch pornography and take part in roleplaying in online chatrooms [2]. Next was virtual reality (VR): technology that allows the consumer to watch pornography as if they were in the same room or even as a participator. Given the huge development in artificial intelligence, designers now focus on a new project: sex robots, sexbots for short.

Current developments

Sexbots are robots designed to have sex and form a connection with humans [3]. This connection can be made because of their human-like behaviour: sex robots are designed to talk with humans, respond to touch, and move their head, eyelids and lips in a natural way [4]. This may sound futuristic, but in reality the American company TrueCompanion already put sex robot Roxxxy on the market. Roxxxy has five different personalities, customized hair and eye colour, and the ability to converse in English with humans. The first Dutch speaking sex robot was introduced in 2018 by manufacturer Mytenga. She is called Robin, has brown hair, brown eyes and speaks with a Flemish accent (figure 2). Mytenga and TrueCompanion are not the only companies to create sexbots. Multiple entrepreneurs all over the world are joining them in the race to design the best sex robot, similar to the race of designing the first fully autonomous driving vehicle.

Figure 2: Robin, the first Dutch sex robot [5].

Positive impact

The introduction of sexbots into society has various advantages. For example, people with social anxiety or sexual inexperience could first practice on a sexbot before being intimate with a real person. Furthermore, lonely people, due to living in a rural area or being socially isolated, could use a sex robot to fulfil their social and sexual needs. An example in pop culture is the film Lars and the Real Girl (2007), where main character Lars has a sex doll as his girlfriend (figure 3). Sex dolls are sex robots that cannot move, speak a language or react to touch. However, lonely people finding comfort in having a sex doll as their partner suggests that sex robots could increase quality of life for those people as well. Sex robots also show promise for people with a certain mental or physical disability that have difficulty finding a partner.

Figure 3: Shot from the film Lars and the Real Girl (2007).

In addition, sex robots could be used in education and healthcare. To illustrate, sex therapists could use sexbots to treat men with erectile disorders or premature ejaculation problems. Sexbots could also be used to teach young adults consent. Imagine a female robot that will tell you to stop if you touch her when she does not want it and that only climaxes when you stimulate her appropriately.
Naturally, sex robots do not only have to be used by one person. People that share an intimate relationship might also find a sex robot useful and even fun. For example, couples that share the same fantasy to have a threesome might first practice with a sex robot. In long-distance relationships, people could have sex with a robot if they cannot see their partner(s), but still want to have sex. Sex robots could also open up new forms of sexual experience. Together, partners could discover new ways of experiencing intimacy with others and in general, introduce sexbots to their bedroom to spice up their sex life.


Of course, new technologies do not come without serious concerns that urge to think about ethical questions. One of the main concerns is that people will use robots to act out their darkest fantasies, such as rape and paedophilia. Consequently, these people might act out these urges in the real world more quickly, increasing the number of (children) victims of sexual abuse. However, there is no answer to the question whether sex with a child sex robot will fuel exploitation of real children, and it is not possible to determine its answer either, since that would result in a highly unethical study [6].
Another great concern is that sex robots will make men objectify and mistreat women. Given that sex robots can be customized to the buyer’s preferences, these robots could be programmed to be extremely submissive and have disbalanced body proportions.
These two concerns ask for regulations that do not yet exist. The absence of such rules complicates law suits regarding sex robots and allows for child sex robots to be distributed. For example, over the course of two years, 42 child sex dolls were seized by Canadian border agents [7].
These and other concerns about sex robots horrify some so much that Kathleen Richardson, a British robot ethicist, started The Campaign Against Sex Robots [8].

Get in formation

The critique of sex robots and concerns about the effect of sex with robots are understandable. If indeed all doom scenarios are true, sex robots will have a deeply negative impact on society. However, it is important to keep in mind that technology is not inherently evil. Ethicists are right to be critical about sex robots stereotyping the female body and the serious complications around child robot sex, just like entrepreneurs are right to see the positive impact sex robots have on lonely people or couples.
A campaign against and ban on sex robots, however, will not solve concerns about sexbots. If anything, sex robots will very likely stay female and child sex robots will continue to be shipped across the border. To effectively confront these issues, we have to face them head-on and answer the questions surrounding sexbots ourselves. It is up to us to discover the possibilities of this new technology and create laws to remove all ambiguity concerning sexbots. To strive for a sex positive future, where women, men, and everything in between and beyond with various sexual preferences equally enjoy sexbots, a group of designers, ethicists, psychologists and entrepreneurs representing this diversity should play an active role in the creation of sex robots.

For Dutch speaking readers:

Mytenga uploaded an interview with sexbot Robin on YouTube


[1] Maines, R. P. (1999). The technology of orgasm. Baltimore: Johns Hopkins University Press.
[2] Griffiths, M. D. (2001). Sex on the internet: Observations and implications for sex addiction. Journal of Sex Research, 38: 333–342.
[3] Levy, D. (2009). Love and sex with robots: The evolution of human-robot relationships. New York.
[4] Sex robot manufacturer Realbotix FAQ. Retrieved from
[5] Picture of Dutch sex robot Robin. Retrieved from
[6] Sharkey, N., van Wynsberghe, A., Robbins, S., & Hancock, E. (2017) Our sexual future with robots: a foundation for responsible robotics consultation report. Retrieved from
[7] Celli, R., & Harris, K. (2018, December 12). Dozens of child sex dolls seized by Canadian border agents. CBC News. Retrieved from
[8] Richardson, K. (2015). The asymmetrical ‘relationship’: Parallels between prostitution and the development of sex robots. Special issue of the ACM SIGCAS newsletter, SIGCAS Computers & Society, 45(3): 290–293.

What is AI and where did it come from?

Even though the term Artificial Intelligence is in nearly everybody’s everyday vocabulary, most people will not be able to explain it. And even amongst those who can explain what AI is, there are many who go silent if you asked for the history of AI.

If you are one of those people, do not worry and keep reading.

This is the first blog post in a blog post series about the history of AI. During the next month, we will publish a number of articles dealing with the biggest milestones in the history of AI. This is your chance to learn something about history which is not taught in school but has implications on our everyday life and the future. We will try our best to make this series understandable for readers with only limited knowledge of the concepts of Artificial Intelligence, introducing concepts as they appear.

The Definition of Artificial Intelligence

Looking into the most well-known encyclopaedias and dictionaries, Artificial Intelligence is usually defined as the ability of a machine (a computer, a robot etc.) to perform intelligent tasks which simulate human thinking or would require some form of intelligence to be performed.

This itself is not too surprising, as it is somewhat entailed in the term itself. But what is behind this definition? How does Machine Learning, Neural Networks and all the other increasingly popular terms fit into this image?

To understand where we are now, it might be a good idea to start at the beginning. And in the case of AI, the beginning is rather close but its foundations go back to the time of Aristotle.

Dartmouth Workshop Proposal Quote
The Dartmouth Workshop is now commonly referred to as the founding event of AI

The foundations of AI

There are a number of disciplines which are considered to form the foundation of AI. Without a doubt the oldest is Philosophy. In the time of the ancient Greek, Aristotle’s Syllogism formed the basis of a field which is now known as Logics. Next to Logics, also philosophical questions of what knowledge is and where it comes from are ways Philosophy contributed to Artificial Intelligence.

But the list of other disciplines which contribute noticeably to AI is long and can be found in Formal Science (Mathematics, Computer Science) as well as Social Science (Psychology, Economics, and Neuroscience, which is an interdisciplinary discipline itself). This diversity shows that Artificial Intelligence is a somewhat special discipline. It does not entirely fit into the natural sciences area but it also does not fit any other discipline.

The Beginnings

The term AI was first mentioned in 1956 in a two-month workshop proposal at Dartmouth College:

“We propose that a 2-month, 10-man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.”


This proposal was preceded by several works which are now considered to be highly influential for the founding of AI as its own disciplines. Examples for this are the Turing Test (1950), by Alan Turing as well as his theory of computation (1948); Shannon’s information theory (1948 ) and Norbert Wiener Cybernetics (1948). Surprisingly enough, from today’s point of view, the mentioning of the first neural networks also fall in this time. Pitts and McCulloch showed how networks (which researchers later would call neural networks) can perform logical functions in their paper “A logical calculus of the ideas immanent in nervous activity” (1943).

The Dartmouth Workshop

“We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.” This last sentence from the above-quoted proposal will just be the beginning of a number of overestimations in the field of AI. Despite the high ambitions, the Dartmouth workshop did not lead to any breakthroughs. However, this workshop is now considered to be the founding event of Artificial Intelligence and for this, it earned its place in the history of AI.

After the Summer

Since the 1956, AI has seen ups and downs. The first initial hype ended with the AI winter when researchers came to realize that their high expectations could not be met. It went so far that the United Kingdom stopped AI programs at all but two universities. That this time is over now is obvious. If the time of overestimation of AI possibilities is over, however, is something yet to be discovered.

Next in the Series…

Are you interested in a deeper dive into the history of AI? The next blog posts in this series will take a deal the events preceding the Dartmouth college. What did Shannon’s information theory entail? What does the Turing test actually test? And how are Pitt’s and McCulloch’s networks related to today’s neural networks?
Afterwards, we will continue our journey through the history of AI, continuing our story with the Dartmouth conference and eventually reaching the state of the art.

Follow us on LinkedIn or Facebook to stay up to date with new Blog publications!


Russell, S. J., & Norvig, P. (2016). Artificial intelligence: a modern approach. Malaysia; Pearson Education Limited,.


Understanding the Importance of Attachment Theory in Social Robotics

A post-apocalyptic world where humans are nearly extinct and a humanoid robot is tasked with the mission of repopulating humans on the planet Earth. This is not a figment of my imagination but the plot of a Netflix movie, “I am Mother”. Discussing this intense thriller movie would be really engaging but unfortunately, it would be a bit tangential to the topic at hand. Instead, it is quite interesting to focus on this one aspect of the movie, which is the relationship between the humanoid robot and the human child.  

In the movie, the robot portrays the role of a sort of a surrogate mother and a caregiver of the newly born infant. This intriguing bond which is shared between them is the crux which will be explored in this article.

Attachment Theory and HRI

One of the defining characteristics of human beings which separates us from other animals on this planet is the social interaction amongst humans. A major aspect of survival depends on the social interaction one human has with another. Such interactions were pretty simple back in the prehistoric ages but in the modern world, they have evolved and taken up a complex form. And understanding social human interaction has been one of the major fields of neuroscience and psychology. 

Attachment Theory is one such study of social interaction which explores the attachment behaviour portrayed by humans. John Bowlby, the psychiatrist responsible for the conception of this theory, shifted the classical theory of associating human attachment shown in infants from a stimulus (for example, food provided by a human caregiver) to a more emotional connection with a human.  This theory was confirmed to a great extent by Harry Harlow in his work involving newly-born monkeys (McLeod, 2017). 

The need to understand human cognitive behaviour gave rise to the field of Social Robotics and Human-Robot Interaction (HRI). These fields are, in some sense, quite similar to each other as HRI can be considered as a subfield of social robotics with the main motivation of understanding human cognition via interaction of humans with robotics. Emerged around the 1990s, HRI has gained a lot of recognition in contribution of understanding human cognition via understanding and testing robotic systems which dynamically interact with humans.

An arousal-based model controlling the behaviour of a Sony AIBO robot during the exploration of a children’s play mat was designed based on the research in developmental robotics and attachment theory in infants. When the robot experienced new perceptions, the increase of arousal triggered calls for attention from its human caregiver. The caregiver could choose to either calm the robot down by providing it with comfort, or to leave the robot coping with the situation on its own. When the arousal of the robot has decreased, the robot moved on to further explore the play mat. Hence, the study presented the results of two experiments using this arousal-driven control architecture. In the first setting, it is shown that such a robotic architecture allows the human caregiver to influence greatly the learning outcomes of the exploration episode, with some similarities to a primary caregiver during early childhood. In a second experiment, it was tested how human adults behaved in a similar setup with two different robots: one needy, often demanding attention, and one more independent, requesting far less care or assistance.

Long Term Dyadic Robot Relations with Humans

In Attachment Theory, the caregiver-infant relationship (Bowlby, 1958) is widely popular due to the paradigm shift of knowing how infant attachment to their mothers or caregivers works and the factors which play a role in it. This relationship was explored with the use of a Sony AIBO robot where an arousal-based model is created for a robot to stimulate responses from human caregivers, (Hoile et al., 2012). The study was successful in showcasing that the robot running on the arousal-based model was able to elicit positive caregiving behaviour from the humans instead of being left to cope with the situation the robot at any particular time. The arousal-based model essentially turned the robot either needy or independent and the human caregiver responses were recorded for either of the behaviours portrayed by the robot. 

While the above study dealt mainly with this dyadic relation of human and robot, effects of long-term HRI and it’s association with the Attachment Theory was studied by exploring various factors such as attachment styles, formation and dynamics (McDorman et al., 2016). This study has thus proposed Attachment Theory as a somewhat generalised framework for understanding long-term HRI.

Influence of Human Attachment Patterns on Social Robotics

As mentioned before, the Sony AIBO robot experiment (Hoile et al., 2012) was successful in stimulating human caregiver responses but this showcased the human to be the response system in the human-robot relation whereas it is also important to understand how a robot might behave as a response system based on a human’s actions. This aspect was explored as well where EMYS type robots were set up to spend 10 days with humans with different attachment patterns and the robots’ operations were assessed based on their response to the various styles of attachment displayed by the humans (Dziergwa et al., 2018). The above two studies in a way represent the two sides of a coin as understanding the behaviours of a social robot playing the “infant” as well as the “caregiver” role might provide a more articulate knowledge of the Attachment Theory and its association with HRI. 

Importance of Attachment Theory in Social Robotics

Another study involving human interactions with the PARO robot (Collins, 2019) explored the Attachment Theory and HRI by drawing parallels with other forms of human interactions and bonds, such as with other humans, animals and objects. Although the results weren’t conclusive, it demonstrated how important Attachment Theory can be in understanding and developing HRI methodologies. 

The Wrap-Up

In conclusion, multiple studies have shown the importance of taking inspiration from Attachment Theory to better understand HRI and developing cognitive models which follow the norms such as increased attachment towards emotional stimuli and not simple, materialistic stimuli (for example, food). Advancements in HRI by considering the Attachment Theory shows great potential in more successful assistive robots which can display a personalised attachment behaviour towards humans. 

Although studies similar to Harlow have not been attempted on humans where they are isolated from other humans and placed in the care of only robots, it poses an interesting question whether prolonged interaction and attachment to a social robot might reduce a human’s ability to create as well as retain other attachments with humans.


Bowlby, J. (1958). The nature of the childs tie to his mother. International Journal of Psychoanalysis, 39, 350-371.

Hiolle, A., Cañamero, L., Davila-Ross, M., & Bard, K. A. (2012). Eliciting caregiving behavior in dyadic human-robot attachment-like interactions. ACM Transactions on Interactive Intelligent Systems (TiiS), 2(1), 3.

McLeod, S. A. (2017, Feb 05). Attachment theory. Simply Psychology.

McDorman, B., Clabaugh, C., & Mataric, M. J. (2016). Attachment Theory in Long-Term Human-Robot Interaction.

Dziergwa, M., Kaczmarek, M., Kaczmarek, P., Kędzierski, J., & Wadas-Szydłowska, K. (2018). Long-term cohabitation with a social robot: A case study of the influence of human attachment patterns. International Journal of Social Robotics, 10(1), 163-176.

Collins, E. C. (2019). Drawing parallels in human–other interactions: a trans-disciplinary approach to developing human–robot interaction methodologies. Philosophical Transactions of the Royal Society B, 374(1771), 20180433.