Commercialization - The Soft Underbelly of Artificial Intelligence

Artificial intelligence has grabbed business by the lapels and is demanding its time in the spotlight. Breathless headlines in the mainstream media feed the flames and it’s not hard to see why. A.I. is now able to dominate masterful human players in an increasing long list of games of skill. We’ve also watched as A.I. proves itself a match to the skills of physicians in identifying radiological abnormalities, detecting cancer and recommending treatment.

Despite these advancements, business uptake of AI in operational settings is still relatively low. The issue is that AI has still not broken free of its academic origins. The AI of academia is an almost artistic creation, assembled carefully and delicately to achieve a specific, narrowly-defined predictive capability. The construction of an A.I. starts with the data, but not just any data. All data must first be filtered, conditioned, normalized, vectorized and deposited into storage such as Hadoop before it even meets the AI model. The model itself can easily be compared to a hand-crafted clockwork of equal parts science and intuition. If the model is based on an artificial neural network, it will consist of layer after layer of weighted activations whose function is so complex, their makers are strained to explain exactly how they do what they do. Despite the internal complexity, A.I.s are tuned with scores of dial-like hyperparameters to coax the A.I. towards its maximum performance.

While commercially interesting A.I. models routinely spring from labs, ask the academic creator of the A.I. to then scale it and replicate it to many instances and you will see them rapidly lose interest. Their interest is simply to come up with better and more creative models.

The current challenge with A.I. is to take an interesting model and make it robust, appealing and usable to as wide an audience as possible, that is to say, be able to productize it. It's a lot of painstaking time-intensive work. You will likely have to integrate it into a wider framework and to establish a standard API through which customers can interact with it. But we’re just getting started. It must be made scale to industrial levels. If the A.I. is to handle credit card transactions or spam, there’s no sense commercializing a product whose design limits it to only a few users. Then there are the edge cases and “black swans”. How are these handled? How is it queried?

For A.I. to truly enter business mainstream, practical A.I. technology must be implementable without the current obligatory complement of PhDs to go with it. It is exciting that Canada is making groundbreaking advances in A.I. research through University of Toronto, University of Alberta and recently the newly formed Vector Institute. Indeed, we expect great things but equally important is to develop the parallel track of technical skills and acumen to successfully productize this research and to generate broad mainstream and affordable demand.

Originally published on LinkedIn, April 1, 2017.