|AFFILIATIONS||Professor of Vision Science, Optometry and Neuroscience|
Computational models of sensory coding and visual perception
Each waking moment, our brain is bombarded by sensory information, estimated to be nearly one gigabit/sec. Somehow, we make sense of this data stream by extracting the forms of spatiotemporal structure embedded in it, and from this we build a model of the world containing objects, surfaces, and other relevant information for guiding action. The overarching goal of research in my laboratory is to understand how this process occurs in the brain, focusing especially on the thalamo-cortical system.
One major line of work is to develop probabilistic models of natural images, and to construct neural circuits capable of representing images in terms of these models. For example, we have developed a model of natural images based on the principle of sparse coding — in which the retinal image is explained in terms of a small number of events at any given point in time — and we have shown that the receptive field properties that emerge in such a system match those found in the primary visual cortex (V1) of mammals. The suggestion then is that V1 may be operating, at least in part, according to a similar principle. We are currently working on extending this model to learn invariances from natural image sequences, in addition to building models composed of multiple layers to capture the hierarchical structure of visual cortex.
Another goal of our work is to build physical computing and memory systems that work more like the brain. Current computing hardware revolves around exact, discrete representations consisting of 1’s and 0’s, with a central clock synchronizing all operations in the system. By contrast, the brain does most of its computation by manipulating analog signals (the membrane voltage) asynchronously in continuous-time, and it works all of its magic utilizing very low amounts of power – the human brain consumes just 20 watts! We are currently working together with electrical engineers to develop methods for encoding image data and other natural signals in an efficient manner that can exploit the intrinsic analog storage properties of low-power, nanoscale memory devices. We are also developing new models of neural computation based on high-dimensional representation for building a holistic, internal representation of a scene from multiple fixations of the eye.
Engel JH, Eryilmaz SB, SangBum Kim, BrightSky, M, Chung Lam, Hsiang-Lan Lung, Olshausen BA, Wong H-SP (2014) Capacity optimization of emerging memory systems: A Shannon-inspired approach to device characterization. Electron Devices Meeting (IEDM), 2014 IEEE International, pp.29.4.1,29.4.4, 15-17 Dec. 2014 doi: 10.1109/IEDM.2014.7047134 PDF
Cheung B, Livezey JA, Bansal AK, Olshausen BA (2015) Discovering hidden factors of variation in deep networks. International Conference on Learning Representations 2015 Workshop. arXiv:1412.6583v4 PDF
Köster U, Sohl-Dickstein J, Gray CM, Olshausen BA (2014) Modeling higher-order correlations within cortical microcolumns. PLOS Computational Biology, 10(7): e1003684. doi:10.1371/journal.pcbi.
Lewicki MS, Olshausen BA, Surlykke A, Moss CF (2014) Scene analysis in the natural environment. Frontiers in Psychology, 5, article 199. PDF
Olshausen BA (2014) Perception as an inference problem. In: The Cognitive Neurosciences V, M. Gazzaniga, R. Mangun, Eds. MIT Press. PDF
Olshausen BA, Lewicki MS (2013) What natural scene statistics can tell us about cortical representation. In: The New Visual Neurosciences. J. Werner, L.M. Chalupa, Eds. MIT Press. PDF
Olshausen BA (2013) Highly overcomplete sparse coding. In: SPIE Proceedings vol. 8651: Human Vision and Electronic Imaging XVIII, (B.E. Rogowitz, T.N. Pappas, H. de Ridder, Eds.), Feb. 4-7, 2013, San Francisco, California. PDF
Olshausen BA (2012) 20 years of learning about vision: Questions answered, Questions unanswered, and Questions not yet asked. In: 20 Years of Computational Neuroscience. J. Bower, Ed. (in press) pdf
Cadieu CF, Olshausen BA (2012) Learning intermediate-level representations of form and motion from natural movies. Neural Computation, 24(4):827-66 pdf
Charles AS, Olshausen BA, Rozell CJ (2011) Learning sparse codes for hyperspectral imagery. IEEE Journal of Selected Topics in Signal Processing, 5, 963-978. pdf
Tosic I, Olshausen BA, Culpepper BJ (2011) Learning sparse representations of depth. IEEE Journal of Selected Topics in Signal Processing, 5, 941-952. pdf
Wang CM, Sohl-Dickstein J, Tosic I, Olshausen BA (2011) Lie Group Transformation Models for Predictive Video Coding. In: Data Compression Conference 2011 proceedings. pdf
Culpepper BJ, Olshausen BA (2010) Learning transport operators for image manifolds. In: Advances in Neural Information Processing Systems, 22, Y. Bengio, D. Schuurmans, J. Lafferty, C. K. I. Williams, A. Culotta, Eds. PDF
Olshausen BA, Cadieu CF, Warland DK (2009) Learning real and complex overcomplete representations from the statistics of natural images. In: SPIE Proceedings, Vol. 7446: Wavelets XIII, (V.K. Goyal, M. Papadakis, D. van de Ville, Eds.), August 2-4, 2009, San Diego, California. pdf
Rozell CJ, Johnson DH, Baraniuk RG, Olshausen BA (2008). Sparse Coding via Thresholding and Local Competition in Neural Circuits. Neural Computation, 20, 2526-2563. pdf
Olshausen BA, Field DJ (2005) How Close Are We to Understanding V1? Neural Computation, 17, 1665-1699. pdf