Learning Transformational Invariants from Time-Varying Natural Images
Bruno A. Olshausen · Neural Information Processing Systems · 2008
How does the brain represent and learn the structure contained in dynamic visual scenes? Previous work using unsupervised learning in the time domain has shown that sparse coding can uncover direction-selective components that are tuned to specific spatial and temporal frequency bands. However, these models do not capture more complex motion selectivity such as speed tuning or pattern motion, which are response characteristics found in higher visual area MT.