Author | Liu Yan's lawsuit-ridden Facebook has another trouble. Facebook accused of stealing algorithm
Recently, the startup company Neural Magic filed a lawsuit against Facebook. Neural Magic accused Facebook of stealing and disclosing its own core algorithm secrets and intellectual property rights. The algorithm is mainly used to accelerate the use of hardware. It has now become a part of PyTorch. And key technologies.
Some people speculate that this key technology may be Facebook's deep learning compiler Glow.
The indictment information shows that Neural Magic is a small AI algorithm startup founded in 2017. The founders are two professors from the Massachusetts Institute of Technology, Nir Shavit, a professor in the Department of Electrical Engineering and Computer Science at MIT, and MIT Computer Science and Artificial Intelligence. Alex Matveev, a research scientist in the laboratory.
The core technology of this company is a technology called "neural magic", and the "neural magic algorithm" is the core of the neural magic technology. It is understood that the neural magic algorithm can make certain types of neural networks run efficiently, and engineers can handle the calculation problems of deep learning without using special chips such as GPUs. The algorithm can reduce hardware costs in the fields of medicine, network economy and artificial intelligence.
This is not the first time Facebook has faced accusations of stealing algorithms.
In June 2019, a Lithuanian data company sued Facebook for using a stolen 3D object database to train AI projects. This is a SUNCG dataset composed of 3D data of furniture such as doors, windows, tables, chairs, etc. Facebook uses it to train robots to recognize three-dimensional spaces.
The SUNCG data set used by Facebook was not directly stolen by it. The initiator was Princeton University. People at Princeton University illegally downloaded more than 1 million digital objects from the Planner5D website by writing code. Facebook used these data sets without permission. Eventually Facebook was sued for infringing on trade secrets and copyright laws, and the use of the SUNCG data set was also banned.
Is the core algorithm of the PyTorch compiler "stolen" by employees?
This is a dispute case involving infringement of trade secrets caused by a former employee's job-hopping.
Also pushed to the dock with Facebook was Aleksandar Zlateski, a Facebook technician.
Zlateski was the first employee to join Neural Magic. When he served as the technical director of the company, Zlateski was trusted by the team, with an annual salary of $165,000 and enjoys company options.
During the work, with the status of technical director, Zlateski has access to all trade secrets, proprietary information and future business plans of neuromagic technology. A very important point is that Zlatesk has also been exposed to the core software for developing neural magic, and its source code encapsulates the compiler of neural magic algorithms.
In July 2019, after working for a year and a half, Zlateski notified Neural Magic that he had obtained a new position at Facebook and promised that the next technical work has nothing to do with the work at Neural Magic. Taking into account the differences in the nature of the positions, Neural Magic felt that Zlateski did not violate the competition agreement and only reminded him to keep the professional information of the neural magic algorithm confidential when he left.
Six months later, Facebook announced the open source of PyTorch on GitHub. But what Neural Magic did not expect is that PyTorch contains the technology and intellectual property rights that constitute the core neural magic proprietary algorithm, which is essentially the core algorithm of Neural Magic.
In the lawsuit file, Neural Magic also mentioned a "real hammer" detail. Facebook publicly thanked Zlateski at a meeting in December 2019 and highly praised him for his significant contributions to solving sparse kernel and unified code caching.
On January 18, 2020, a LinkedIn post from a Facebook employee made Dr. Matveev of Neural Magic realize that Facebook and Zlateski had stolen the neural magic algorithm. Subsequently, Neural Magic reviewed the open source code released by Facebook and confirmed that Zlateski had misappropriated neural magic algorithms and gave them to Facebook.
After discovering the infringement, Neural Magic asked Facebook in writing to delete the infringement part on Github, but it was rejected by Facebook.
Neural Magic decided to protect its legitimate rights and interests in the form of law. It believed that Facebook’s actions violated the Massachusetts Trade Secrets Act. Zlateski violated the confidentiality and non-competition agreement signed during his tenure at the company. Neural Magic requested three times the actual damages, and asked Facebook to delete the code and stop further use of Neural Magic’s proprietary confidential information.
Currently, Facebook has not yet responded to the allegations. PyTorch is one of the most mainstream machine learning frameworks at the moment. It is widely used in academia and industry. If the infringement facts are established in the subsequent case, the development history of PyTorch will undoubtedly be cast a shadow of plagiarism.