Can human understand machine learning and develop trust? Trustworthiness of AI systems stands on three delicate pillars – Fairness, Explainability and Security. The objectives are to avoid biases due to social stereotypes, prevent misinformation and stop privacy leak. In addition, compliance framework serves as a support mechanism to the three pillars and provide better stability while evaluating Trustworthiness of AI systems.

I briefly touched upon the fairness in one of my earlier posts – Ethics in Artificial Intelligence (https://cosmicouslife.wordpress.com/2024/01/15/ethics-in-artificial-intelligence/). Fairness is judgemental as it deals with diverse discriminations. This aspect requires much better treatment, later.

Compliance with respect to regulations is nothing but interpretations of the other three attributes of trustworthiness to the local environment in which the AI operate. This will make policy makers busy and enable a lot of business opportunities to the top consulting firms, adding to the overall cost of AI systems.

Explainability requires some explanations. As long as we get value for the money or do not bother too much about the occasional small losses, we may trust the black box recommendation systems even if it keeps its mouth shut on why it recommended a particular thing. I undergo interesting explanation experience. An interesting person with whom I interact often never just answers any questions posed but comes out with a chain of reasons for the answer every time, challenging my patience. Do we need explanations for everything? No, but for some high impact situations. There is a cost attached to it.

Accuracy versus Interpretability trade-off constraints the choice of algorithms used for machine learning and so the explainability. Linear regression and Neural Networks are at both end of the spectrum, with Decision tree, K – Nearest Neighbours, Random Forest, and Support Vector Machines in-between. With neural network Large Language Models (LLM), interpretability is more of a challenge. If the system comes out with smart explanations, are we smarter enough to distinguish concocted explanations that the AI systems are capable of?

Explainability involves pointing out to the parts of an image like eyes that make an algorithm to label the image as a frog or building a decision tree of reasoning to provide traceability. Explanations are nothing but post facto confessions like ‘Why I did What I did’ derived from trained prediction models. Google came out with ‘chain-of-thought prompting’ for getting LLMs to reveal their ‘thinking.’ Thilo Hagendorff, a computer scientist at the University of Stuttgart in Germany has gone to the extent of saying that psychological investigations are required to open up the mad machine learning black boxes. Researchers are measuring the machine equivalents of bias, reasoning, moral values, creativity, emotions, and theory of mind in AI models to evaluate their trustworthiness. Like neuro imaging scan for humans, researchers are designing lie detectors that look at activation of specific neurons in neural network models to identify those set of neurons which wire together and then fire together. Anthropic, an AI safety and research company created a map of the inner workings of one of their models on May 21^st, 2024. This map can aid understanding of neuron like data points called features that affect the output.

A refreshing approach to explainability revolves around making the networks to learn from explanations rather than justifying the predictions. I was intrigued listening to Prof Vineeth Balasubramanian of IIT, Hyderabad, talking about his works on ‘Ante-hoc explainability via concepts’ in a recent IEEE event held in Chennai. Prediction models based on supervised learning are like memorizing everything and vomiting answers without adequate conceptual foundation. Researchers refer them as ‘stochastic parrots,’ meaning those models are probabilistic combination of patterns of text encountered while training, without understanding of the fundamental concepts. Models with concepts and rules seem to set the direction for explainability. This is counter intuitive to conventional prediction models. Expert systems of erstwhile era are decent implementation of explainable AI systems. Is it not like going back to the basics for human understanding of machine learning?