global
As most of us know, bias continues to be a challenge in machine learning and artificial intelligence (AI). Remember the 2015 Google scandal? When Google’s facial recognition system tagged black people as gorillas, and a flaw in camera AI software labeled Asian faces as “blinking” when they were smiling.
Understanding this kind of bias requires both demystifying its underlying processes and the political will to force structural and institutional change. This is argued by former Google employee Timnit Gebru, who was fired for failing to be fair in AI. model training.
These prejudices are often reduced to race and gender, but they actually extend to aspects such as the homogenization of linguistic norms. This is due to the decreasing online footprint of languages by small or poor countries.
In other words, bias, in computer terms, can take many forms, including overfitting or underfitting, recall bias, observer bias, selection bias, exclusion bias, association bias, measurement bias, and outliers. So how do these occur?
A modern approach to AI
The modern approach to AI is very different from the methods used before 2010. Prior to 2010, the methods used aimed at representing human knowledge symbolically using explicit rules that relied on semantic nets, logic programming, symbolic mathematics, etc.
Since 2010, a family of machine learning (sometimes called AI) techniques, deep learning or deep neural networks, has become the dominant paradigm. AI and deep learning are often confused, but technically they are different.
For clarity, artificial intelligence is an interdisciplinary method of modeling and replicating cognitive processes in the creation of “intelligent” machines using computational, biological, mathematics, and other theories and applications. .
Deep learning, or deep neural networks (DNNs), on the other hand, is one of the approaches taken to achieve this, which together with computer architecture combine multiple or “deep” layers of units, or “neurons” within layers. It works by using (or system design) and highly optimized algorithms.
Innovative Backpropagation Advances
Backpropagation is a method of training neural networks in so-called “supervised learning” that neural networks use to update their parameters. In other words, by propagating backward responsibility for observed errors in the output units, we can weight the initial parameters to make the network’s predictions more accurate.
Neural networks are, in a sense, “learning” using forms of reasoning that people also use, such as feedback and error correction. So, just as students use feedback from their instructors to update their knowledge parameters, so do neural networks. This is why I mention “algorithmic reasons” or “deep learning”.
But what a neural network “learns” through this backpropagation process is not necessarily what a human learns. This is because the statistical methods that neural networks rely on to learn are dependent on the dataset they are fed. We willingly learn those prejudices and, more worryingly, reflect them on us.
This dynamic is evident, for example, in our personal social media feeds, generating filter bubbles that distort our view of reality.
Deep learning and bias propagation
This can also lead to extreme political bias, often with the prevalence of conspiracy theories. Because the surveillance capabilities of these new technologies can be used in sophisticated ways to exploit and weaponize existing biases in society and datasets.
One reason this happens is when the network strictly clusters the patterns it finds in the training dataset.
For example, imagine that a learning model needs to identify and analyze pictures and pictures of cats. However, models are primarily trained on long-haired cats, and hairless cats are classified as separate animals such as dogs.
What this means is that the machine learning model failed to extrapolate patterns from the dataset and, as a result, was unable to effectively generalize what it learned. Of course, this is exactly what happened with the camera software and Google’s facial recognition system I mentioned at the beginning of this article.
The obvious answer to fix this might seem simple: have a better dataset. Including more information from countries in Asia, Africa and the Global South, or using Africa-specific or Global South-specific datasets can be very beneficial for creating more accurate and less biased AI systems. It seems obvious that
But the more complex answer is that even if this kind of inclusivity positively impacts machine learning, it doesn’t really address the underlying systemic predispositions of our society, and that few people Nor does it change the concentration of power, information and wealth given to “Tech giants” – companies such as Apple, Facebook, Google, Microsoft, and Amazon.
It also fails to adequately address exploits associated with correcting data set skew. For example, Amazon Mechanical Turk is known for exploiting workers in a variety of ways, including paying incredibly low wages for the repetitive and tedious task of labeling images used in datasets. I’m here.
These tasks include exposure to graphic and violent images, often leading to work-related post-traumatic stress disorder, as discussed below.
It may not need to be added, but much of the AI and machine learning development is concentrated in the countries of the Global North, while much of the raw workforce for this expansion is outsourced to the countries of the Global South. I’m here.
Content moderators who review graphic media exposed to horrific content such as suicide, murder, and rape are often in locations such as Nairobi, India, and the Philippines, and earn as little as US$1.50 an hour.
AI ethical imperative
Much of the work being done to address these bias issues, as well as issues such as opacity and concentration of wealth, is covered by what is collectively known as “AI ethics.”
While this work is highly valuable, it does not adequately address the deeper individual and collective impacts of digitization on our societies, psychological health, and even thought processes.
The late technology philosopher Bernard Stiegler found these more subtle effects very disturbing. spent on He explained it in terms of a kind of “generalized developmental arrest” that materializes as a symptom of widespread dissatisfaction.
These include, but are not limited to, decreased ability to experience pleasure. Depression and hopelessness; lack of concentration due to cognitive saturation; new dysmorphia such as Snapchat dysmorphia – to a perfect appearance by erasing perceived imperfections using social media platform filters and other enhancements A body image disorder characterized by an obsession with A new social phobia like hikikomori, or acute and prolonged social withdrawal, was first discussed by Tamaki Saito in his 2013 book. Hikikomori: Endless Youth.
My point is that while we still have to address issues such as fairness in AI and machine learning, we also have to address the underlying issues that are driving these issues.
It asks not only how we can eliminate biases in datasets, but also why those biases are reflected in datasets in the first place, and the powers and political and economic arrangements that keep these biases from changing in society. We should ask what this tells us about . .
From this perspective, the question is not so much what is a good algorithm as what we think a good society is and what place technology should occupy in that society.
As Dan McQuillan argues, the ethical challenges that arise from AI arise not from the results of machines and computational processes, but from the way machine learning processes augment prejudices and other tendencies that already exist in our society.
Chantelle Gray is Professor of Philosophy, Faculty of Humanities and Chair of the Institute for Contemporary Ethics at Northwest University, South Africa.
