By: Michael Feldman
Gamalon Inc, emerged from stealth mode this week, announced two machine learning products, based on an in-house technology known as Bayesian Program Synthesis (BPS). The company claims BPS can perform machine learning tasks 100 times faster than conventional deep learning techniques, while providing more accurate results.
“We call our way of doing this Bayesian program learning,” said Gamalon founder and CEO, Ben Vigoda at a recent TED talk. “And it’s already becoming an important new idea and technology in machine learning. It’s also incredibly beautiful mathematically.”
He believes using Bayesian probabilistic modeling is a much more efficient way, that is, a much less computationally intensive way, to infuse intelligence into machines. Unlike deep learning, which often needs millions of data examples to train a neural network, a Bayesian model can be built with much fewer examples. The examples have to be of better quality, but the process cuts down on computation time and human effort dramatically. So instead of collecting a million tagged data items and commandeering a GPU-accelerated cluster for a few days, a model can be built on a laptop in just minutes using a handful of examples.
An MIT-trained computer scientist, Vigoda is no stranger to probabilistic thinking. In 2006, he founded Lyric Semiconductor, which developed logic circuitry for hard-wired probability computation. The technology’s initial application was performing error correction in flash memory chips. In 2011, Lyric was acquired by Analog Devices. Two years later, Vigoda turned to the software side of probability computation and founded Gamalon. To date, Gamalon has collected $7.7 million worth of research contracts from the US Department of Defense’s DARPA program, as well as $4.5 million in investor seed money from Felicis Ventures, Boston Seed Capital, and Rivas Capital, as well as individual investors like Adam D’Angelo, Andy Bechtolsheim, Steve Blank, Ivan Chong and Georges Harik.
BPS, the technology that Gamalon invented, is derived from Bayesian probabilistic techniques. The new products, dubbed Gamalon Structure and Gamalon Match, represent the company’s first commercial offerings. According to the press release, “Gamalon Structure converts streams of text paragraphs found in databases or documents, into clean, structured data rows ready for use in the enterprise. Gamalon Match can then deduplicate/link these data rows.”
While that may seem a little mundane, Gamalon says the software can structure almost any type of business data and enable new avenues for scientific research. One example Gamalon offers is for company inventory data, which is often stored in different formats and uses different nomenclature for identifying the same thing (i.e., CA=Calif=California, 1teaspoon=5ml=5cc, and so on). The Structure product can normalize all the data into a master inventory and do it very quickly. According to Gamalon, their pre-alpha customers have used the software to “accomplish in minutes with twice the accuracy what previously took large teams of people months or even years.”
In a broader sense, Vigoda thinks his company’s approach will eventually replace the brute-force neural networks that are so popular today for AI applications. In fact, he seems to believe that probabilistic modeling is the approach that will lead the industry to artificial general intelligence (AGI), that is, the ability to mimic human brainpower. “Probabilistic [models] are, I believe, are the path to making machines that will someday, themselves, have ideas worth spreading,” said Vigoda, and then joking, “I’m at least 80 percent sure of it.”