A Love for Explainable Algorithms

At the start of 2022, I stopped working on a startup that I co-founded with a friend of mine. At that point, I decided to spend the next year working solely on personal projects.

These projects were designed to further my understanding of Bayesian decision theory. This may sound strange to some. Why spend a whole year on Bayesian decision theory? Is it really so interesting? Is it really so complex? And is the return on that sort of investment worth it?

The answer I have to those questions is an emphatic “YES”. The reason for that answer begins with the fact that, for much of my adult life, I have been fascinated (perhaps obsessed) with explainable algorithms. Explainable algorithms are those that are not only designed to make automated decisions, but that also provide human beings with interpretable reasons for those decisions. In the most fascinating of cases, explainable algorithms can teach human beings new things about the world around us. These days, automated explainable algorithms are often referred to as “Explainable AI”.

And so, at the start of 2022, I asked myself: “How can I contribute to the future of explainable AI? What are the possible frontiers that my humble mind can comprehend?”

Choosing Decision Theory

After much deliberation, I came to believe that those frontiers would lie in undiscovered innovations of decision theory and probabilistic reasoning. After all, aren’t the most valuable decisions regarding uncertain outcomes?

And, if we are to create truly transparent “thinking machines”, must we not find an objective way to translate uncertainty into human language (and vice versa)?

Thankfully, such a daunting task is not one that we embark on from scratch. For the past hundred years or so, Bayesian probability theory has been used as an objective means to translate subjective knowledge into numerical form. It has even been proven that these objective rules mirror the axioms of formal logic (that is, if a decision violates the rules of Bayesian inference, then it must also violate the rules of logic).

Whilst reading the works of Bayesian pioneers (such as Jeffreys and Jaynes, and even Gibbs(!) and Laplace(!!)), it suddenly became apparent to me that a large body of work from the early 50s to the late 80s contained incredibly important, but mostly forgotten insights.

Forgotten Breakthroughs

Though perhaps not completely forgotten, many important discoveries regarding Bayesian inference remain trapped in academic niches. These niches are themselves stuck in a time when the computer was still in its infancy, and when analytical (not computational) solutions limited the scope of possible progress.

When I realised this, I decided to create a website (this website), which I would use to publish lesser-known but valuable concepts: not just from Bayesian inference, but from other relevant fields also. My hopes are that the blogposts, projects and open-source code that I write will illustrate to others how valuable those concepts are. And perhaps even connect those ideas to more current and pressing issues. For example, the measurement and removal of unwanted bias from automated decision processes.

Thoughts on the Future of Explainable AI

There is no way to know when a paradigm shift will occur in Explainable AI, or what it will look like when it does occur. But I am confident that, whenever such a breakthrough will happen, it will not be due to the discovery of new and exotic methods for encoding and decoding patterns.

Rather, it will come from practical and applicable answers to questions such as: “Which patterns in a given context constitute knowledge?”, “What do we mean when we say we learn knowledge?” and “How can all forms of subjective knowledge be described by an objective and universally interpretable mathematical language?”.