Learning
This presentation is the property of its rightful owner.
Sponsored Links
1 / 25

Learning PowerPoint PPT Presentation


  • 75 Views
  • Uploaded on
  • Presentation posted in: General

Learning. Computational Cognitive Neuroscience Randall O’Reilly. Overview of Learning. Biology: synaptic plasticity Computation: Self organizing – soaking up statistics Error-driven – getting the right answers. Synapses Change Strength (in response to patterns of activity). What Changes??.

Download Presentation

Learning

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.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.


- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -

Presentation Transcript


Learning

Learning

Computational Cognitive Neuroscience

Randall O’Reilly


Overview of learning

Overview of Learning

  • Biology: synaptic plasticity

  • Computation:

    • Self organizing – soaking up statistics

    • Error-driven – getting the right answers


Synapses change strength in response to patterns of activity

Synapses Change Strength(in response to patterns of activity)


What changes

What Changes??


Gettin ampa d

Gettin’ AMPA’d


Opening the nmda receptor calcium channel

Opening the NMDA receptor calcium channel

Glutamate

Na+

CA2+

AMPAr

NMDAr

Mg2+


Which way

Which Way?


Causal way

Causal Way?


Let s get real

Let’s Get Real..


Urakubo et al 2008 model

Urakubo et al, 2008 Model

  • Highly detailed combination of 3 existing strongly-validated models:


Allosteric nmda captures stdp including higher order and time integration effects

“Allosteric” NMDA Captures STDP(including higher-order and time integration effects)


What about real spike trains

What About Real Spike Trains?


Extended spike trains emergent simplicity

Extended Spike Trains =Emergent Simplicity

S = 100Hz

S = 50Hz

S = 20Hz

r=.894

dW = f(send * recv) =

(spike rate * duration)


Xcal linearized bcm

XCAL = Linearized BCM

  • Bienenstock, Cooper & Munro (1982) – BCM:

  • adaptive thresholdΘ

    • Lower when less active

    • Higher when more..

      (homeostatic)


Threshold does adapt

Threshold Does Adapt


Computational self organizing and error driven

Computational: Self-Organizing and Error-Driven

  • Self-organizing = learn general statistics of the world.

  • Error-driven = learn from difference between expectation and outcome.

  • Both can be achieved through XCAL.


Floating threshold long term average activity self org

Floating Threshold = Long Term Average Activity (Self Org)


Self organizing learning

Self Organizing Learning

  • Inhibitory Competition: only some get to learn

  • Rich get richer: winners detect even better

    • But also get more selective (hopefully)

  • Homeostasis: keeping things more evenly distributed (higher taxes for the rich!)


Limitations of self organizing

Limitations of Self-Organizing

  • Can’t learn to solve challenging problems – driven by statistics, not error..


Floating threshold medium term synaptic activity error driven

Floating Threshold = Medium Term Synaptic Activity (Error-Driven)


Fast threshold adaptation late trains early

Fast Threshold Adaptation:Late Trains Early

Essence of Err-Driven: dW = outcome - expectation


Backpropagation

Backpropagation


Where does error come from

Where Does Error Come From?


Leabra

Leabra


Hebbian learns correlations

Hebbian Learns Correlations


  • Login