In December 2015, three researchers from MIT and NYU found a new way to make AI relevant.

Computers typically need a long time and hundreds or thousands of examples to understand concepts (Think millions of images and multiple processors for a computer to learn that whiskers are different from a tail).

But these researchers, using a model called the Bayesian Program Learning framework, built an algorithm that gets machines to learn the way humans do: using only a few examples. They used probabilistic programming – code that deals in probabilities rather than specific variables— not only to recognize but also reproduce over 1500 characters written in many languages such as Tibetan, Gujarati, Sanskrit, among others.

At the end of this experiment, the researchers victoriously declared that “fewer than 25 percent judges” could distinguish between the handwritten characters with those drawn by the computer. Euphoria!

But once the chatter around the research died down, it became obvious that if computers can recognize handwriting today, there’s a lot more they will be able to do tomorrow.

While this study was purely academic, other researchers, stirred by the revelation, have been finding ways to make a buck out of what the 18th century statistician Bayes formulated and what probabilistic programming can do.

In February 2017, a startup called Gamalon emerged from stealth mode (it was started in 2013) to reveal some of its own plans based on the Bayesian Program. And last month, this Boston-based startup had raised $20 million in a series A funding led by Intel Capital, with participation from .406 Ventures, Omidyar Technology Ventures, Boston Seed Capital, Felicis Ventures, and Rivas Capital, driving its funding to a total of $32.1 million.

The New Wave of AI

Most data is unstructured. It’s just blobs of text full of what’s called data dirt. Unstructured data cannot be visualised, nor can it be used for machine learning or to run business processes. It can take many computers months to access millions of data points, and even then, the information that churns out may or may not be useful.

But Ben Vigoda, an MIT-trained computer scientist and Gamalon’s CEO says he has a quick-fix.  

CEO Ben Vigoda

In a talk, Vigoda gives the example of a grocery app that connects drivers to grocery stores to pick up supplies for consumers. Say, you need the 12 ounce diet cola but the app, having sent a signal to every cash register’s data system, received results populated with other types of colas. Vigoda says this is a difficult problem to solve with AI because the machine is not educated about the different brands of beverages, the variety among brands, or that ounce can be abbreviated as OZ. This is where Gamalon’s products with its probabilistic programming come into the picture.

Vigoda told Bloomberg last year, “You can run our software on a laptop, and it takes 100 times less horsepower to find an answer.” And companies do. They use Gamalon’s products – Structure, which picks up concepts from raw text efficiently and can adequately describe the product, and Match – which categorizes data and learns concepts quickly. For example, Match will learn the numerous abbreviations, brands, and types and correctly seek out the 12 ounce diet cola.

Vigoda had earlier founded a company called Lyric Semiconductor that did projects for the Department for Defense and used probability to guess user behaviour, purchase patterns, improve results etc. commercially. It was acquired by Analog Devices.

Pictionary with Probabilities

Another idea Gamalon is using probabilistic programming is for its easy-to-create drawing app.

Recognizing drawings through deep learning

In 2016, Google launched its “Quick, Draw” app which used deep-learning to recognize drawings. However, the sketch of what you drew had to be similar to what it had already seen. Gamalon, on the other hand, used Bayesian Program Synthesis to teach the machine concepts such as a line, triangle, rectangle, etc., along with the shapes it could make like a lamp or a chair. The app then used probabilistic programming to recognize key features. As long as the shapes remained the same, the app, Gamalon claims, could guess correctly. It also corrected your drawing of say, a triangle if your hand wavered midway.  

In 2017, MIT Technology Review’s listed Gamalon over companies like Facebook and Tesla as the 21st Smartest Company, crediting its technology’s efficiency advantage over other machine-learning methods. Its efficiency comes in from its probabilistic programming algorithms that learn from less-data, perhaps just a few examples.

But along with groceries or the drawing app, the startup uses its technology to help e-commerce, manufacturing and other companies structure and match text data from disparate sources, such as inventory databases. It can perform numerous roles from optimizing supply chains to doing competitive price analytics, but it’s its role in recognizing speech command that’s most intriguing.

Conversing With Machines

Today’s virtual assistants and chatbots search through large amounts of text and then follow assigned rules in order to respond to questions. But what if they could understand language, make guesses, be less confused, and have conversational memory? Vidoga told MIT that Gamalon has built an interface that defines a tree of conversation and the “various different ways the dialogue might unfold”. This software has far-reaching commercial and practical applications and takes Vigoda closer to his dream of using AI “to get all the participants quickly and seamlessly connected (on a conference call), stay connected, (when the assistant) takes notes, captures hand-drawn figures and re-draws them beautifully, and sends everything to everyone in a nice summary afterwards.”
But AI, like anything with an eye on world domination, is divided into camps: those who are Bayesian like Vigoda, follow probabilistic programming, or those that use neural networks, or reinforcement learning, among others. The future can be none, either or a combination of them all. But for now, Gamalon’s less-data approach, that can also convince privacy campaigners to disincentivize companies from accumulating tons of personal data, to its focus on teaching machines the way humans learn can have significant impact on how we perceive AI and the new ideas that leap in from the periphery.