Sign in or Join FriendFeed
FriendFeed is the easiest way to share online. Learn more »
Keynote: Tomaso Poggio - Computational Neuroscience: Models of the Visual System
Not a talk about bioinformatics but computational neuroscience - Roland Krause
Learning in supervised and non-supervised settings to understand the processes in the brain. - Roland Krause
Just now computational neuroscience can feed back into more classical disciplines. - Roland Krause
15 years ago, machine learning techniques were developed for face recognition. Now present in many consumer digital cameras. - Roland Krause
Now, detection of pedestrians with Daimler and the future will be detecting people in cars. - Roland Krause
No brain research involved so far, only engineering. - Roland Krause
Basics on learning theory, the mathematics of learning: Poggio & Smale, 2003 - Diego M. Riaño-Pachón
[...] How the ventral stream works. - Roland Krause
Build quantitative models, only modeling feed forward connections. - Roland Krause
A good place to start because we can quickly recognize e.g. animals in pictures, no time for real feedback[?]. - Roland Krause
A model of visual representation based on the neuroscience of the cortex, described by Riesenhuber and Poggio in 1999, 2000 and Knoblich, Kreimann and Poggio in 2005 - Roland Krause
Iterations between tuning and maximization steps in the step wise neuronal graph. - Roland Krause
agreement of the hierarchical model was surprising given that it used only data from anatomy and physiology, not psychophysics - Ruchira S. Datta
adding layers decreases the number of examples needed for a given accuracy - Ruchira S. Datta
Empirical motivation for hierarchical learning: sensitivity increases by adding more layers. - Gabriele Sales
consider images as functions mapping the square to {0,1} - Ruchira S. Datta
similarly, function spaces on subpatches - Ruchira S. Datta
there are multiple embeddings of the subpatches into the larger patch, e.g., from different translations - Ruchira S. Datta
have finite sets of templates, which can instantiate on different subpatches - Ruchira S. Datta
now define a kernel based on composing the embedding with the template and taking the dot product - Ruchira S. Datta
just made this clean mathematical formulation, now there are many open problems - Ruchira S. Datta
One interesting conjecture: small effective dimensionality at each layer. - Gabriele Sales
extend model to videos: sequences of images - Ruchira S. Datta
Extensions of the model to videos and sequences of images - Allyson Lister
computational mechanisms of invariant recognition - Ruchira S. Datta
system: computer vision of mice in cages - Ruchira S. Datta
The system based on this model is looking for mice in cages. - Roland Krause
classify simple behaviors by watching a couple of seconds of video: grooming, walking, etc. - Ruchira S. Datta
Patterns like grooming or walking should be quantified. - Roland Krause
Collected ~100 hours of videos and then try for an automated analysis - Allyson Lister
Hand labelling of the videos to train the system. - Gabriele Sales
The automated analysis of 100 hours agrees at 71% with a human labeler. Two human laberlers agree at 72% only. - Roland Krause
Two human labelers agree 72% of the times. The system achieves 71%. - Gabriele Sales
Can use markov chains to analyze movement / mouse behaviour - Oliver Hofmann
24/7 survey of well-being of mice or other animals possible. - Roland Krause
have 4 different strains of mice; can infer the strain from behavior given long enough video - Ruchira S. Datta
You can infer the mouse strain from the behaviour with about 50% accuracy with 10 mins of video - Allyson Lister
Behavioral comparisons of mice being model systems for psychiatric diseases. - Roland Krause
Vision still remains more than simple categorization: it is image understanding / inference / parsing. - Gabriele Sales
Limits of present FF models: vision is more than categorization or identification: it is image understanding/inference/parsing. Our visual system can "answer" almost any kind of question about an image or video (a Turing test for vision) - Allyson Lister
want to *understand* the image, not just categorize or identify - Ruchira S. Datta
Vision is more than categorization, the system would fail a visual Turing test. - Roland Krause
vision has to do with general intelligence - Ruchira S. Datta