SAN JOSE, Calif. — A new startup debuted software that it claims will disrupt the emerging field of machine learning. Gamalon uses complex math and algorithms instead of the high-end processors required for the deep-learning techniques becoming popular with the likes of Facebook and Google.
“This is a big breakthrough and will be a sea change in artificial intelligence,” said Ben Vigoda, who sold Lyric Semiconductor to Analog Devices in 2011.
The startup came out of stealth mode the same day a rival, GraphCore, revealed some details about the software behind its AI processor which is tailored to the current deep-learning techniques (see page 3 for more).
Today’s neural network approaches require training with the huge data sets typically only web giants possess. The job takes so much processing power that Google created its own accelerator, Facebook designed an open-source GPU server, and Intel acquired two chip startups to handle it.
Gamalon uses a statistical form of lambda calculus and proprietary algorithms to create associations in text data with virtually no training. It uses Python as an inference engine or solver algorithm, opening a door to easier interactions with programmers. The jobs run on standard servers faster than deep-learning techniques, the startup claims.
“All of the investments that people are making in hardware for neural net training are probably misplaced because this will be solved algorithmically,” said Vigoda.
The startup announced its first products using the technique it calls Bayesian Program Synthesis. It also shared initial feedback from early tests and announced more than $12 million in investments.
Pre-alpha customers accomplished in minutes with twice the accuracy what previously took large teams months or years, the company claimed in a press release. Results have been verified by Galois, Inc., a private R&D firm in Portland, Oregon, it said.
Gamalon snagged a seed round of $4.45 million from investors including Andy Bechtolsheim and Adam D’Angelo, the co-founder and CTO of Facebook. In addition, it won $7.7 million in contracts from the Defense Advanced Research Projects Agency.
Next page: Getting computers to write programs
Gamalon beats Google in accuracy inferring "Street" from "St." (Images: Gamalon)