340 likes | 562 Views
Computational Modeling of Place Cells in the Rat Hippocampus . Nov. 15, 2001 Charles C. Kemp. Talk Overview. Give an introduction to place fields and the hippocampus Review two models both with navigation using place fields one with a model for generating place fields
E N D
Computational Modeling of Place Cells in the Rat Hippocampus Nov. 15, 2001 Charles C. Kemp
Talk Overview • Give an introduction to place fields and the hippocampus • Review two models • both with navigation using place fields • one with a model for generating place fields • Critique these two models • Look toward the future of the field
Hippocampus & Place Cells • Big Field • Journals, books • Research groups • Long History • 1800’s for hippocampus • 1971 for place fields • Few Answers
Importance of the Hippocampus • Episodic memory • HM • Spatial tasks • allocentric system • Navigation • taxi drivers • VR
What about Rats? • Hippocampal anatomy is very similar in rats and humans. • Rats show similar deficits from hippocampal lesions. • Monkeys have view cells.
Zooming into the Rat Brain • 2g • 150 million neurons • 500,000 pyramidal neurons in hippocampus
Review of Two Explicit Computational Models Arleo & Gerstner (2000) Foster, Morris & Dayan (2000)
Navigation with Place Fields • Both models of navigation • Create functions that associate actions with locations in the environment. • Train these functions while the simulated rat navigates in an environment looking for a reward. • Foster, Morris, and Dayan’s also learns coordinates for each location
Functions For Behavior • Location -> Action • A(p)={a1(p), a2(p), ... an(p)} • P[A, p](i) = probability of action i • Location -> Value • c(p) ~ • v(p) = Max[A(p)] • Location -> Metric Coordinate • {x(p},y(p)}
Learning the functions • Recursion Trick • Gradient Descent and Hebbian Learning
reset Vision place cells Path Integration place cells Linear combination CA1 Hippocampal place cells Arleo’s model of Place Fields
Feature Vector Maker Ii fi Filter Bank Magnitude Subsample Feature Vectors for Snapshots • Collects four images at each position it visits • Converts all images to feature vectors prior to use.
Snapshot cells are combined to make sEC cells. • A radial basis function is put around each of the four feature vectors from a new location • the outputs from these 4 radial basis function are combined as a weighted average • the weight vector is adapted by a hebbian update rule
Weighted average of PI cells and sEC cells makes Place cells • PI cells: • use integration of wheel turns • represent as a set of radial basis functions • strongly responding PI cells and sEC cells are combined using sEC cell method
Static Navigation can’t learn from a single example • Rat’s can • Water maze • 2 meter diameter • opaque water • hidden platform, 1.1 cm diameter • After 3 days of 4 trials a day • minimal latency after a single example
Single example learning without metric navigation? • Topological • Cue, Action sequence • note distal cue from example • swim to center • look for cue • swim towards it until proper distance from the wall Real paths (steele&Morris 1999) Simulated paths (Foster et al 2000)
Self-motion information • Save et al, Hippocampus 2000 • olfactory information is more important • self-odor has been neglected • place cells go unstable • 39% (dark/cleaning) • 80% (light/cleaning) • few remain stable • 10% (dark/cleaning) • 0% (light/cleaning) • Both models assume • accurate self-motion information • stable place fields • Arleo & Gerstner assign too much importance to PI cells
More problems with RBF place fields • Wood et al, Nature 1999 • smell cup • if matches last cup smell ignore • if it doesn’t match last cup smell dig for food
A better model for place cells? • Hartley et al, Hippocampus 2000
Getting there faster. • quantify the input • robot • rat VR • model the environment • record the output • at least head position, body position, eye position • camera array to record the rat • observe the computations • improved multi-electrode arrays • chronic implantation • multi-region • larger number (1/2 million cells) • facilitate collaboration
Robots • Are they a good model • better methods of quantifying the input exist • poor models of rat senses and actions • convenient, cool looking • Can they help this research? • indirectly, yes • elucidate issues • explore complex tasks • for example, Sebastian Thrun and Hans Moravec
Navigating the Microstructure • compartmental models • statistical characterizations • 3D reconstruction and data sets Ascoli et al. (1999) Fiala & Harris (2001)
Conclusion • Introduced place fields and the hippocampus • Reviewed two models • both with navigation using place fields • one with a model for generating place fields • Critiqued these two models • Tried to look toward the future of the field
Other Points of Interest • Abstract Neighborhoods • Generalized Snapshots • Searching through states • Beyond simple navigation