Teaching signals and information processing in the basal ganglia networks

1
Teaching signals and information processing in
the basal ganglia networks
Hagai Bergman Department of Physiology, The
Hebrew University Hadassah Medical School,
Jerusalem, ISRAEL hagaib_at_md.huji.ac.il
2
The BiG Question
Why do we need the cortex-BG-cortex loop if there
are so many cortico-cortico connections?
3
The working hypothesis The basal ganglia use
reinforcement signals and local cellular
(Hebbian, Anti-Hebbian) learning rules to reduce
the dimensionality of cortical input and pass
the compressed information to the frontal cortex.
Limbic Cognitive Motor

• Motivation
• Reduction in the needed wiring.
• Reduction in the number of synapses.
• Easier modulation of the system.

4
Dimensionality reduction ????

• Dimensional reduction in the corticostriatal
  projection implies loss of information
• To represent all possible corticostriatal input
  patterns uniquely in the output of the striatum,
  there must be at least as many striatal neurons
  as corticostriatal neurons.
• Zheng and Wilson, JNP 2002

5
Dimensionality reduction made simple
Redundancy reduction Data compression
Principal/Independent component analysis
Factor/Cluster analysis
A compression of the information encoded by a
large neuronal population to a significantly
smaller number of neurons. Efficient reduction is
achieved when all or most of the information
contained within the original space is preserved.
6
The nn input (cortical) space

• Binary neurons
• Rate coding neurons
• Rate/pattern coding neurons

7
Scaling in the woodsRuderman and Bialek, Phys.
Rev. Letters, 1994
8
Dimensionality reduction the actual input space
is smaller than the potential input space
The size of the potential input space is nn.
Dimensionality reduction will work when the
inputs actually lie within a significantly lower
(e.g., 2n) sub-space of the input space.
9
Reinforcement driven dimensionality reduction
improved detection of reinforced dimensions
10
The learning process of a 64?8 reinforcement
driven dimensionality reduction network
Vertical lines
Horizontal lines
11
Reinforcement driven dimensionality reduction
experimental predictions

• Uncorrelated activity in the output nuclei of the
  BG
• Temporal changes in BG correlation during
  learning
• Hebbian and anti-Hebbian learning rules in the
  feed-forward and lateral BG connections,
  respectively
• Multi-parameter encoding of the cortical world in
  the BG
• Coding of reinforcement probability by the BG
  teachers (critics)
• Coding of action by BG actors

12
Pallidal and SNr activities are not correlated
Nini et al, JNP 1995 Raz et al, JNS 2000 Heimer
et al, JNS 2002
Bar-Gad et al, JNS, May 15, 2003
13
The experimental method multiple electrode
recording of striatal ACh INs (TANs), DA, GP and
SNr neuron
14
Behavioral taskProbabilistic instrumental
conditioning
15
Behavioral control Kinematical and performance
parameters do not depend on reward probability
16
The monkey knows the probabilities
Relative choice
Relative reinforcement
17
DA response - example
18
TAN response - example
19
SNr response - example
20
Event related neurons - numbers
21
Reward prediction error sensitive neurons -
numbers
22
Population responses
23
Population responses linear regression
24
DA and ACh temporal correlation of responses
DA
TAN
25
DA and ACh spike correlation
1. As with the TANs DA response is very
synchronized 2. Electrical coupling between DA
neurons
TANs, n110 pairs
DA, n65 pairs
26
Summary
27
Summary and Conclusions

• ACh enhances activation and inactivation of K
  currents
• ACh freezes the up/down state of the striatal
  neurons
• Pause in ACh enables this neurons responsiveness
  to cortical input

Surmeier. 1993

• The dopamine response is indicative of the
  predictive value of various events in relation to
  reward (Teacher).
• The cholinergic pause signal may provide a
  temporal frame, defining the time period in which
  the dopamine signal will be processed (School
  bell).
• The dopaminergic and the cholinergic system
  cooperate to teach the basal ganglia actor what
  are the relevant (reinforced) dimensions of the
  cortical activity.

28
Thanks...

• Histology
• Suzanne Haber

• Electrophysiology
• Genela Morris Yoda
• David Arkadir Clara
• Alon Nevet Electra
• Guy Saban Giga

• Everything
• Moshe Abeles
• Eilon Vaadia

• NN models
• Izhar Bar-Gad
• Yaacov Ritov

29
(No Transcript)