Two MIT researchers sue Facebook over neural-networking code they say a former employee stole from them

A Somerville start-up founded by two MIT researchers who say their software could revolutionize complex computing charge that their first employee stole their proprietary algorithms when he left for a job at Facebook.

In a lawsuit filed this week in US District Court in Boston, Neural Magic - started by computer-science professor Nir Shavit and research scientist Alexander Matveev - Neural Magic charges that neural networking software Facebook posted to the GitHub open-source software repository uses the code they developed.

Neural Magic says what makes the code particularly valuable is that it can run on off-the-shelf computers without requiring expensive specialized processors, through "proprietary and heretofore confidential techniques that enable commodity multicore machines, that is, servers with multiple computing units, to execute complex mathematical operations in a computationally and storage-efficient manner."

Shavit and Matveev say that after they started the company in 2017, they hired Aleksandar Zlateski, then an MIT post-doc working on computer vision, as their first employee - paying him $165,000 a year and giving him stock options to work as the company's technology director.

Neural Magic says Zlateski also signed an agreement not to disclose any of the code should he leave the company - which he did in 2019, after, Neural Magic claims, he was reminded of his NDA and after he said his new job had nothing to do with neural networks

This past November, Neural Magic charges, Facebook posted what it said was open-source code on GitHub that was, in fact, largely based on its proprietary algorithms. A Facebook software engineer speaking at a conference in December, praised Zlateski's code for speeding up Facebook software, the suit alleges.

The code and compiler Facebook posted to GitHub implement the same Neural Magic Algorithms used in Neural Magic’s compiler code, to achieve the same computational and storage efficiencies running on commodity hardware (CPUs). Indeed, Neural Magic has tested the Facebook compiler side-by-side against its compiler, and the results from this direct comparison establish that the algorithms implemented in the Facebook compiler are the Neural Magic Algorithms.

The company adds:

Neural Magic can only see what Facebook has already published to GitHub, which includes further Neural Magic technology in addition to the said compiler. Zlateski could disclose (and on information and belief has disclosed) even more to Facebook, for use internally or in future open source releases. Those continued disclosures would further damage Neural Magic's business and market opportunities. Trade secrets at risk of further disclosure (perhaps already disclosed) include trade secret techniques used by Neural Magic to run fully connected neural networks and convolutional neural networks at speed and to speed up the training of such networks.

Neural Magic said Facebook rebuffed its request to take down the code, which the start-up said left it with no choice but to sue.

The company is seeking treble damages as a start, along with punitive damages, attorneys' fees and "seizure of property incorporating Neural Magic’s trade secrets."