Description:
Docket # 15-3161
Exponential increases in computing and imaging technology have both enabled and necessitated machine-based alternatives to human image analysis. To solve this challenge, computer scientists have created multi-layer artificial neural networks (ANNs), a field colloquially known as “deep learning”. The most successful of these networks are typically supervised, or trained on a large number of correctly labeled inputs. While such ANNs have achieved near-human performance, they require the laborious assembly of labeled training sets. To unlock the full potential of machine learning and allow this technology to scale with increasing computational power and data, robust unsupervised training methods must be developed.
Researchers at the Princeton University Neuroscience Institute have used fundamental research into the computations performed in the cerebral cortex to envision a novel, unsupervised, biology-inspired neural network architecture. In response to a learned visual stimulus, specific groups of neurons in the brain reinforce existing connections and predict temporal sequences of events. Violations of these predicted sequences cause broader network activation that forms the substrate for learning new temporal sequences. By studying this process of prediction, reinforcement and violation, Princeton researchers have gained insight into how biological neural networks are so incredibly sophisticated at unsupervised pattern recognition and object classification. These same insights can be applied to the creation of next-generation ANNs with an improved ability to learn through observation.
Applications:
• Image categorization
• Video analysis
Advantages:
• Unsupervised learning
• Mimics cutting edge understanding of visual cortex pattern learning
Inventors
Michael Berry began his higher education with a BS in Physics and a Minor in Philosophy from UC Berkeley. Then, he got his PhD in Physics from Harvard. At that point, he switched to Neuroscience, doing a postdoc at Harvard and finally moving to Princeton in 1999, where he started his own lab. For over 20 years, Prof. Berry has studied the image processing computations carried out in the retina. One of his most notable results was the discovery that the retina is capable of making simple predictions about the upcoming visual stimulus, all by itself. Prof. Berry also studied the code used by populations of retinal neurons to encode visual information. Starting in 2013, his lab expanded its focus to include studies of neural computation in the visual cortex. Drawing on his earlier work regarding predictive computations and population coding principles, these investigations have led to a new theory about the canonical computations in the neocortex, which forms the basis of a novel machine learning algorithm.
Intellectual Property Status
Patent protection is pending.
Industry collaborators are sought to further implement and commercialize this technology.
Contacts
Chris Wright
Princeton University Office of Technology Licensing • (609) 258-5579• cw20@princeton.edu
Laurie Tzodikov
Princeton University Office of Technology Licensing • (609) 258-7256• tzodikov@princeton.edu