A Reddit post has triggered discussions regarding a patent Google filed on Dropout supposedly becoming active last week. Many were shocked and concerned regarding potential patent infringement if they kept using the popular neural network training technique — L’Oréal Lead Data Scientist Louis Hénault incredulously tweeting “Seriously @GoogleAI ?”
So what is Dropout? For years, machine learning researchers grappled with overfitting when training their neural networks. In 2014, a Google research team led by renowned deep learning pioneer Geoffrey Hinton introduced Dropout, which addresses overfitting by randomly removing units from the training process. Dropout has since became an efficient and widely used regularization technique for training neural networks.
Synced discovered that Google first filed its Dropout patent application in 2013, and the patent actually became active not last week but rather back on August 2, 2016. The patent expiration date is listed as September 3, 2034.
An attorney with a focus on corporate and securities law who asked not to be identified told Synced that “even if the patent is granted, you can still use it as long as Google does not sue you. It is possible that Google can wait until your company grows up and come back to sue you, which could potentially make you or your institution liable for consequences and damages, not Google.”
Tech giants patent their neural network algorithms just like other techniques in order to legally commercialize their inventions and to defend themselves in the event their algorithms are misused. The number of patent application filings related to AI and machine learning has been growing rapidly over the past several years. Google field 99 applications related to machine learning or neural networks in 2016, up from just one in 2010, Wired has reported. Facebook meanwhile filed 55 patent applications the same year, up from zero in 2010.
Google and DeepMind, the UK based research team owned by Google’s parent company Alphabet Inc., have already patented the following important ML techniques:
- Processing images using deep neural networks
- Computing numeric representations of words in a high-dimensional space (word2vec)
- Generating audio using neural networks
- Processing sequences using convolutional neural networks
- Generating video frames using neural networks
- Neural networks for selecting actions to be performed by a robotic agent
- Processing text sequences using neural networks
- Training action selection neural networks
- Reinforcement learning systems
- Scene understanding and generation using neural networks
- Recurrent neural networks
- Sequence transduction neural networks
- Reinforcement learning with auxiliary tasks
- Environment navigation using reinforcement learning
Although the question of usage rights for these and other AI and machine learning properties have not yet come to litigation, tech companies’ moves to patent fundamental algorithms in neural networks have become an area of concern for many in the machine learning community, who fear that companies claiming control over the tech might one day lead to monopolization.
Journalist: Fangyu Cai & Tony Peng| Editor: Michael Sarazen
0 comments on “Concerns on Social Media Over Google ML Patents”