Agents of the Mind

1
Agents of the Mind
Jim Houk, Northwestern University. Dept of
Physiology at NU Med Schl

• Agent-based modeling has been a productive
  approach for illuminating emergent properties of
  complex systems
• The mind clearly is a complex system, and
  concepts about the brains agents have been
  elusive to identify synapses, neurons, clusters
  of neurons?
• I am suggesting that a more opportune choice is
  the set of anatomically defined brain networks
  that I refer to as distributed processing modules
  (DPMs)

2
Distributed processing modules (DPMs) based
on viral anatomy
mind
thinking
moving
3
Where I am heading in this talk Abstract signal
processing operationsthat I will posit for each
DPM agent
Houk (2005) Agents of the Mind, Biological
Cybernetics 92 427
4
The M1 distributed processing module (M1-DPM)
5
Transverse view of cerebellar circuitry
molecular layer
Purkinje cell layer
granular layer
white matter
6
Parasaggital view of CB circuitry
black Purkinje cells
purple mossy fibers
red Climbing fiber
red cerebellar nuclear neuron
7
One Microscopic Module
Cerebellar Cortex
Cerebellar Nucleus
Motor Cortex
Elemental Motor command
Sensory cue
8
Regulation of movement direction
Purkinje cell discharge Premotor neuron discharge
9
CB signal processing scheme one
macroscopic module
M1
10
Coupling learning in one spine to movement
11
Unique computational features in the cerebellum

• Amplification and refinement
• A huge state vector is presented to PCs
• 3-factor learning rule with great credit
  assignment
• --gt basal ganglia

12
M1-DPM
13
A medium spiny neuron in caudate/putamen (GABA)
The dendrites are black. They have 20 thousand
spines.
The axon collaterals are red. They mediate
competitive pattern classification.
14
Schematic of a cortical-basal
ganglionic module
15
3-factor learning rule in the striatum
Glu
one spine
NMDA
LTP
depolarize
neuron
consolidates
Dopamine neurons
reward propensity
16
Dopamine also does this!
Nonlinear amplification
Gruber, Solla, Surmeier Houk, 2003
17
Functional imaging of serial order recall
Replicate Task Timeline
The two contrasts are Execution and Decoding
18
ROI Differential BOLD blood flow
Dave Fraser Paul Reber --gt Houk et al 2005
19
Model of serial order processing
Derrick Fansler-Wald Alon Fishbach
Tested under 3 conditions 1. No collateral
inhibition 2. Postsynaptic inhibition 3.
Presynaptic inhibition
The latter yielded the best noise tolerance, and
it also decreased energy requirements,
which could explain the significant decrease in
caudate ..
20
Schizophrenia patients
Sensorimotor
Serial order recall
Fraser, Park, Clark, Yohanna, Houk, 2003
21
Unique computational features in the basal ganglia

• A large cortical vector projects to spiny neurons
  -- competitive pattern classification of this
  input is like a parallel search through a very
  large data base
• Spiny neurons have a 3-factor learning rule with
  fantastic temporal credit assignment
• Bistability nonlinear amplification is also
  induced by dopamine in spiny neurons
• The loop implements a recursion-like operation --
  well-suited to detect serial order of
  events/thoughts

22
M1-DPM
23
Rehearsal of a task
Response time RT MT -- Practice ?
?
cortical remapping
24
Intracortical connectivity
25
Learningrules
reinforcement learning
supervised learning
Hebbian learning
26
Summary of Agents

• Cortical-basal ganglionic loops perform
  competitive pattern classification to enable
  initiation of cortical patterns, and include a
  recursion-like operation that is sensitive to
  serial order
• Cortical-cerebellar loops amplify and refine the
  enabled patterns to shape their spatiotemporal
  patterns of discharge into appropriate output
  vectors
• Intracortical circuitry learns from practice --
  faster and more accurate
• All the loops have the same neuronal architecture
  -- all the DPMs should
  perform the same signal processing operations on
  their particular input vectors
• A network of DPM agents may be an appropriate
  architecture for exploring the dynamics of the
  mind

27
For more on this topic, see .

• Sule Yildirim -- next talk -- A Computational
  Model of a Microzone of the Cerebellum
• Greg Dam -- poster -- A Computational Model of a
  Cortical-Cerebellar Microscopic Module
• Jun Wang --poster -- Model of a Cortical-Basal
  Ganglionic Module that Encodes the Serial Order
  of Events

28
The signal processing operations posited for each
distributed processing module (DPM)
29
Extras follow
30
Parasaggital view of CB circuitry
black Purkinje cells
purple mossy fibers
31
J. Neurophysiology (2000) Bob Holdefer LT
Chen Lee Miller
32
Regulation of movement direction
Purkinje cell discharge Premotor neuron discharge
33
Schematic of a cortical-cerebellar module
34
A dendritic spine on a PC
35
Distributed processing modules (DPMs) based
on viral anatomy
36
More detailed anatomy
37
Submovements
delayed
overlapping
Novak et al, J Neurophysiol, 2000, 2002,
2003 Fishbach et al, Exp Br Res, 2005
38
M1 unit
CUSUM
39
Paired-site recordings

• Identify sites in GPi that project, via thalamus,
  to M1
• Record from GPi neurons at those sites
• Eventually record pairs of GPi --gt M1 neurons
  from the coupled sites

40
GPi unit -- single trial
41
Time course of loop operations
Working Memory Task
Houk (2005) Agents of the Mind, Biological
Cybernetics 92 427
42
A network of DPM agents
may be an appropriate architecture for exploring
the dynamics of the mind
43
Primary movements submovements
44
Output of the BG selectively facilitates
corrective submovements
Because practice in a task exports repeatedly
practiced knowledge from the basal ganglia to the
cerebral cortex for more automatic execution
Alon Fishbach Stephane Roy Christina Bastianen
45
M1-DPM
Tentative ballpark estimate of an intended action
(primary movement followed by 0-N corrective
submovements)
Amplification and refinement of the movement
primitives