Swiss Computational Neuroscience Seminar: Haim Sompolinsky
Event details
| Date | 22.11.2018 |
| Hour | 16:05 › 16:55 |
| Speaker | Haim Sompolinsky |
| Location | |
| Category | Conferences - Seminars |
Title: Untangling of Perceptual Manifolds in Deep Networks
Abstract: An object perceived under different physical conditions (i.e. location, pose, size, orientation,
background) creates different responses in sensory neurons, resulting in a “perceptual
manifold” in the response-space of a neuronal population. A prominent theory asserts that
hierarchical sensory systems untangle these manifolds, allowing downstream systems to
perform perceptually invariant tasks such as object recognition and classification. However,
it is unclear who to quantify the process of untangling and measure it in biological and
artificial deep networks. In my talk, I will describe recent theoretical advances that relate the
ability to perform invariant object classification to the geometric properties of the perceptual
manifolds. I will show numerical results that use the theory to evaluate the population based
changes of object representations in the successive layers of deep convolutional neuronal
networks (DCNNs) as well as in neural data in visual cortex. This work contributes to the
construction of a systematic theory of sensory processing in deep networks in both AI and
the brain.
Abstract: An object perceived under different physical conditions (i.e. location, pose, size, orientation,
background) creates different responses in sensory neurons, resulting in a “perceptual
manifold” in the response-space of a neuronal population. A prominent theory asserts that
hierarchical sensory systems untangle these manifolds, allowing downstream systems to
perform perceptually invariant tasks such as object recognition and classification. However,
it is unclear who to quantify the process of untangling and measure it in biological and
artificial deep networks. In my talk, I will describe recent theoretical advances that relate the
ability to perform invariant object classification to the geometric properties of the perceptual
manifolds. I will show numerical results that use the theory to evaluate the population based
changes of object representations in the successive layers of deep convolutional neuronal
networks (DCNNs) as well as in neural data in visual cortex. This work contributes to the
construction of a systematic theory of sensory processing in deep networks in both AI and
the brain.
Links
Practical information
- Informed public
- Free
Organizer
- Laboratory of computational neuroscience