Neural Computation & Adaptive Perception Accomplishments
The most widely acclaimed conceptual breakthrough of this program has been the introduction of “deep belief networks.” This model, which builds on the characteristics of biological vision systems, has generated considerable excitement in the field of computational neurobiology. It has triggered intense collaborations among CIFAR theorists and experimentalists who are exploiting its potential as an algorithm for “unsupervised learning.”
- Collaborations between several program members have led to significant advances in teaching computers to identify motion in video – like tracking the movement of a human being against a cluttered background. Researchers have developed one type of model that allows computers to represent complex motions with just a few parameters, making it easier for a machine to know what it is looking at. Another model uses comprehensive data describing how a person’s joint angles change as they walk in various styles. Computers can learn these styles so well that they can produce convincing new motions that retain the original style. This work is of great interest to animation and video surveillance companies.
- A fundamental visual ability that is effortless for human beings but very challenging to computers is the ability to identify objects in images – cars, dogs, trees, etc. Search engines such as Google do offer image searchers, but such engines do not actually search image content. Instead, they use caption information and textual context to generate search results. New methods pioneered by Neural Computation and Adaptive Perception members empower computers to categorize the actual content of images themselves, which allows faster, more accurate image retrieval.
- Over the years, the program has established relationships with other internationally based research institutes working in related areas, such as the Redwood Neuroscience Institute in California and the Gatsby Computational Neuroscience Unit in the UK. It has also developed a successful summer school program, which provides specialized training to approximately thirty graduate students and postdoctoral fellows per year, and facilitates collaboration across disciplines between disparate research groups.