CODED BIAS -Academic Reflection

Coded Bias is about exploring the fallout of MIT Media Lab Researcher Joy Buolamwini’s aim to prove inherent gender and racial bias within face detection systems used by law enforcement. After a great research into the root causes of the above problems, According to me, “Coded Bias” serves as both a “wake-up call” for those who don’t yet realize how those methods are implemented and a “call to action” for implementing ubiquitous system with balanced and more generalized dataset.

The film was sparked by the work of Joy Buolamwini, who conducted facial recognition experiments using A.I. and had difficulty getting the technology to accurately process her face. Investigating further, she experienced that the faces of lighter skinned persons are recognized more accurately than the darker skinned persons. Moreover, the male were more accurately recognized than females, which results in gender and racial bias. Thus, I agree with Joy Buolamwini that algorithms like Face Recognition are highly susceptible to bias and the major cause behind this is the technology which is created and designed by one demographic tends to be mostly effective on that demographic only.

Personally speaking, the way AI is currently implemented in every aspect such as advertisement, online recruitment, financial services, securities and many more can lead to inequalities. The film covers both practical and larger political critiques, arguing that it’s not just that the technology is faulty, even if it works ubiquitously and accurately, it would infringe dangerously on people’s liberties and personal lives.

To put it in a nutshell, There is a need for an increase in contextual awareness in AI developers and limiting the implementation of AI in such fields where it can hinder people’s personal space, freedom of expression and rights to equality. Moreover, no one should be forced to submit their base data to access widely used platforms or economic opportunities.

Bias in data or algorithms can generate from the imbalanced or incomplete training data set or the dependence on flawed information that reflects historical inequalities. Such biased algorithms can lead to unfair decisions as well as unfair representation of certain groups which can have a collective, disparate impact on certain groups of people even without the developer’s intention to discriminate. We could detect the bias by checking whether the dataset is balanced and contains ubiquitous training data. To overcome this problem, the training data should be as diverse as possible and it should include as many relevant groups of data as possible. Moreover, we can cross-examine the set of outputs produced by the algorithm with different groups to check for anomalous results.

AI is based on data, and data is a reflection of our history.”- Joy Buolamwini. From my perspective, such problems occur more with those people belonging to certain groups or communities which have historical racism and inequalities in the criminal justice system as these realities are also reflected in the training data. For instance, African-Americans are more likely to be suspected of having criminal history by AI surveillance due to their historical racism.

Nowadays, the major civil rights concern is to make sure that the Algorithmic Justice is oversight in the age of automation. Algorithmic Solution to such bias problems could be fair if we make sure that the algorithm we are building has a proper way to deal with bias. Moreover, to reduce the bias there should be an increase in human involvement in the design and monitoring of algorithms.

As a Data Science students, it is our responsibility as well, to always consider and anticipate bias while developing machine learning models and maintain AI ethics. Moreover, to mitigate the bias, we should try to find what drives biases, and which irrelevant information distracts the model from making accurate decisions.

Further, we should rely upon a cross functional work team to identify bias before and during the model’s fitting. Apart from this, as said earlier, it is preferable to increase human involvement in the design and monitoring of algorithms.

If I talk about bias in Machine Learning for Finance, there are many biases that can affect the performance of a backtested strategy. For instance, there exists optimization bias in algorithmic trading using machine learning, we can mitigate it by keeping the number of parameters to a minimum and increasing the quantity of data points in the training set. Moreover, taking more than 5 years of historical stock market data could be irrelevant and biased. Instead, we can try to use the most recent data up to 3 years to get accurate results.