Data Science Investment Counter —Funding raised by UK data science companies in 2018.
£ 5.640 Million

Can Bayesian AI Outsmart DeepMind?

A US Artificial Intelligence startup has exited stealth mode and it is now on a quest to challenge the dominant AI paradigm.

Boston-based Gamalon has created AI tech that does away with machine learning, embracing a technique known as Bayesian program synthesis. This makes it possible to build AI systems able to learn new concepts by being shown relatively few examples.

Machine learning algorithms— like those used by Alphabet-owned AI lab DeepMind—are trained to spot patterns and recognise images or words by being fed massive amounts of data.To use a textbook instance: if you want to teach an AI to learn what a cat looks like, you will have to flood it with thousands of pictures of different cats in different environments, taken from different angles.

That actually works pretty well, but it takes a lot of (often manually labelled) images and a lot of (expensive) computing power. Would not it be easier if computers were able to learn stuff the way people do— that is, without going through huge dataset?

That is what Gamalon has set out to accomplish. Its Bayesian (after mathematician Thomas Bayes) approach means that its systems behave in a probabilistic fashion. After being shown only a few images of, say, an armchair, Gamalon’s AI will begin hypothesising what a random armchair would look like— for instance, it will rate as highly probable that it has a seat and two armrests. The system then refines its hypotheses as it meets more examples.

The company’s founder and CEO Ben Vigoda  claims that this model is 100 times more efficient than Google’s deep learning algorithms, and the MIT Technology Review wrote that Bayesian AI could make it simpler  for developers without access to powerful data-processing computers to create AI system.

But the journal underlined  that there are potential pitfalls, too: formulating several different hypotheses every time the AI is presented with a new concept will require a great deal of computational power nonetheless.

Among the upsides, WIRED highlighted how Gamalon’s architecture makes it easier for researchers to understand what is the logic behind the AI’s decision—in a usual machine learning scenario, AIs’ actions are often somewhat unintelligible even for their creators.

In a world that is steadily marching towards entrusting smart machines with powers as serious as bestowing life or death on passersby (remember all those “trolley dilemmas” about self-driving cars?), being able to grasp why they do what they do is a rather useful feature.

That is why, while Gamalon’s only released products so far are data-cleaning software, it is safe to assume its technology is bound to play a significant role in the development of the AIs to come. And the fact the company has received a $7.7 million investment by DARPA —besides $4.5 million in a seed round from VC firms— seems to confirm that.

Image via Gamalon (screenshot)

SHARE THIS ARTICLE:
BY SHACK15

Co-working space and blog dedicated to all things data science.

Subscribe to our newsletter