Machine Learning as Art

I watched a very interesting Ted Talk today by Paul Bloom. In it, he touches on the fact that pleasure is directly tied to our knowledge of the history of events leading to the moment in question. For instance, we view an original work of art much higher than a forgery even though the forgery is nearly identical. The reason for the discrepancy is that the original artwork is connected to a history that involves a strongly creative act. This had me thinking about why I love machine learning so much.



History in an Instant

Most machine learning algorithms I deal with these days are reinforcement learning algorithms. Basically, you have an agent and an environment and the agent must learn how to perform well in that environment. Learning takes place over time via a series of trials and rewards. Sometimes it’s a single agent iteratively improving its policy, while other times it’s a population of agents that are selected for fitness. Either way, both are generating something that I see as beautiful: a history.

An algorithm’s journey through a search space is filled with all the makings of an epic. There’s a desperate struggle, countless failed attempts, and occasionally an ultimate triumph. Unfortunately, it’s difficult for storytellers to due the tale justice. Often it’s passed on via a simplistic line graph with an understated Y-axis label of “Performance”. The nail-biting competition between our protagonist and his nemesis, “Baseline”, is boiled down to an anti-climactic red-colored line surpassing a black one.


Different Perspectives

I think this is often why machine learning algorithms are mostly unusable. To their creators, the battle has been won, “The good guys were victorious! The evil Baseline has been statistically significantly defeated! Now go forth and spread the word via this C++ library with little to no instructions!”

I suppose we all need to step back and realize that we experienced the journey like no one else can. So unlike some abstract paintings or scuptures, our results need to be masterpieces independent of their history. That seems to be what defines the line between changing your paper count and changing the world.

Also read...


  1. Wesley,

    I enjoyed reading this blog-post. A unique perspective.

    I just completed my first course in machine learning (support vector machines, multi-layer perceptrons, regularization, semi-supervised learning, etc.) and I enjoyed it immensely.

    Reflecting now, each result I obtain in this area of work will have a provenance embodying all contributions made to the field from the past to the present.


Comments are closed.