• Diversity
• Emerging Tech

Written by Jillur Quddus, Data Scientist, Methods

Technology provides us with the toolkit to change lives for the better – but if unchecked, it also has the power to discriminate and reinforce stereotypes and bias.

This is true of any technology past, present and future. But it is the advancement of distributed systems capable of storing and processing massive amounts of previously disparate data coupled with the emergence of artificial intelligence (AI) into our everyday lives that requires us to urgently refocus on how technology is being architected, engineered, tested, deployed and governed, to ensure that its impact remains positive – and only positive.

First Principles

Whether we explicitly asked for it or not, the fact is that, today and right at this very moment, artificial intelligence is impacting your life. The majority of the time, AI is invisible to us – whenever we enter a search term into Google, visit websites, use social media, browse Netflix, use online banking, use public transport or walk the streets of any major city, AI algorithms are busy working away in the background to return relevant search results, analyse your browsing habits, social network, viewing preferences, detect signs of fraud and to capture an image of, and process, your face. With such a deep and far-reaching, yet paradoxically invisible, global footprint, the dangers of AI if unchecked and not properly governed are significant and tangible, not least to the groups of people it will adversely affect (usually those already at a social disadvantage).

But before we being to explore these dangers, let us go back to first principles and definitions. AI is a broad term given to the theory and application of machines that exhibit intelligent behaviour. Machine learning (ML) is an applied field of study within the broader subject of artificial intelligence that focuses on learning from data by detecting patterns, trends, boundaries and relationships in order to make predictions and ultimately deliver actionable insights to help decision making. A fundamental tool used in machine learning is that of probabilistic reasoning – creating mathematical models that help us to draw population inferences to understand or test a hypothesis about how a system or environment behaves. As such, machine learning is data-dependent and, supervised and unsupervised machine learning both make an implicit assumption that trends and patterns identified in historic data in order to learn a mathematical function, can be used to make inferences and predictions about the future using that same mathematical function.

Deep learning

(DL) is a sub-field within machine learning where the goal remains to learn a mathematical function that maps inputs to outputs. However, deep learning learns this function by employing an architecture that mimics the neural architecture found in the human brain in order to learn from experience using a hierarchy of concepts or representations, and gradient-based optimisation algorithms. Deep learning is still data-dependent however, and the more data used to train a neural network the higher the predictive accuracy.

Bias Amplification

Now that we have an understanding of what artificial intelligence is, it is clear that data – and lots of it – is required for machine learning to be effective. The more data, the better the predictive powers of the resultant model.

But what if the data itself contains bias? In this case, the model itself will not just perpetuate that bias but in many cases will amplify it. So why do machine learning algorithms amplify bias? To answer this question, we must revisit the mathematical nature of their training structure.

Recall that the goal of machine learning algorithms is to learn a mathematical function to map input data points to outputs. Normally these outputs take the form of a value or classification. In order to learn this mathematical function and to maximise its accuracy, machine learning algorithms seek to minimise the number of errors it makes by reducing the variance in the predictions made by the final model. To reduce variance, algorithms introduce assumptions, referred to as bias in the model, to make the function easier to approximate. By introducing more bias, there will be less variance in your final model, but at the expense of greater divergence from the real-world reality in which your model operates. But not enough bias, and the predictive accuracy of your final model is reduced as there is greater variance. As such, many machine learning algorithms expose a bias-variance trade-off through the manipulation of hyper-parameters so that the developer or data scientist may control this balance.

However therein lies a problem – if a goal of machine learning is to train models to maximise their predictive accuracy, then any bias in the data that feeds these algorithms will be amplified due to the algorithm generalising the data in order to reduce variance. Therefore if the data contains gender or racial bias for example, this bias will be preserved and amplified to help make the model more accurate, perpetuating a vicious circle of bias and discrimination.

Gender bias in Natural Language Processing

Let us take a look at an example algorithm to highlight how the problem of bias amplification is a physical reality in today’s artificial intelligence-driven world. Natural language processing (NLP) refers to a family of computer science disciplines, including information engineering, linguistics, data management and machine learning, with the goal of helping computers to understand natural language used in speech and text. Natural language processing is employed in a variety of scenarios, from interaction with chatbots and virtual agents such as Alexa, to foreign language translation and understanding and classifying written text such as legal contracts and health reports.

One common family of feature-learning techniques used in natural language processing is called word embeddings. A word embedding represents each word or phrase found in a given collection of texts as a vector. Words with similar semantic meaning will have vectors that are close together, and the relationship between words can be quantified through the vector differences. As such, word embeddings are an extremely useful tool for a wide variety of prediction tasks related to natural language. For example, when we wish to determine sentence similarity, such as when asking a chatbot or Google a question, it will employ word embeddings to see if that question has been asked before and take action accordingly based on a similarity index. Another example is analogical reasoning where we wish to predict semantic relationships. A widely used implementation that learns vector representations of words is word2vec which provides semantic relationships between words as follows:

Comments are closed.