1 / 13

A Theory of Cerebral Neocortex

A Theory of Cerebral Neocortex. Marr 1970 Proc Roy Soc. David Marr 1945-1980. Three Key Theories. cerebellum (1969) [2341 citations] feedforward associative memory anatomy: granule cells=codons climbing fibres as teaching signals

laurir
Download Presentation

A Theory of Cerebral Neocortex

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. A Theory of Cerebral Neocortex Marr 1970 Proc Roy Soc

  2. David Marr 1945-1980

  3. Three Key Theories • cerebellum (1969) [2341 citations] • feedforward associative memory • anatomy: • granule cells=codons • climbing fibres as teaching signals • substantial attention to subtractive/divisive inhibition; threshold setting • neocortex (1970) [384 citations] • archicortex (1971) [1594 citations] • recurrent autoassociative memory • offline consolidation process for neocortex • MIT: famously abandoned modelling style -> Vision

  4. Neocortex • general theory • canonical microcircuit • unsupervised learning • what’s the statistical structure? • how can we be sure that this matters? • how do we initialize learning? • umbrella of information theory • realization in terms of clustering • mountains • numerical taxonomy (Jardine & Sibson) • what counts as evidence (codons) • mapping onto the anatomy

  5. Recycling • output cells (P-cells) no fixed meaning • ‘some aspect of the structure of information’ • codons (granule cells) could support probabilistic classification

  6. Redundancy (1) • redundancy reduction is fine near periphery • ‘diversity of specialized coding tricks’ • centrally: (iv)

  7. Redundancy (2) • information storage via probabilistic recoding • no general probabilistic model • but feature-based similarity is common • ‘some objects share features’ • implies mountains

  8. Five Questions • determining the appropriate features (codons) • finding the mountains (spatial recognition) • diagnosing new input in terms of its associated mountain • interpreting partial input in terms of its associated mountain • justifying the mountain in the first place

  9. Justification • internal biological utility • redundancy • uniformity of processing • external biological utility • extrinsic vs intrinsic properties

  10. Mixture of Mountains • spatial recognition, diagnosis, interpretation via EM; ML density estimation • but: • what is if has never been seen before? • similarity is not the same as frequency:

  11. Marr’s Solution • a bit unconvincing

  12. Discussion • gets to the heart of the problem of unsupervised learning • oddly not self-consistent • ‘special’ recoding mechanisms from spatial/temporal contiguity are different? • oddly not ML DE • before EM had been invented • much more to be said: notably recoding; initializing; mountain climbing…

More Related