Transcript ppt
Neural basis of Perceptual
Learning
Vikranth B. Rao
University of Rochester
Rochester, NY
Research Group
Alexandre Pouget
Jeff Beck
Wei-ji Ma
Perceptual Learning in Orientation
Discrimination
► Orientation
learning.
► Perceptual
learning.
discrimination is subject to
Learning (PL) is one such form of
Repeated exposure leads to decrease in
discrimination thresholds (Gilbert 1994).
Central Question
► Perceptual
learning is a robust phenomenon in a
wide variety of perceptual tasks.
► When
applied to orientation discrimination, how
do we relate the learned improvement in
behavioral performance, to changes in population
activity due to learning at the network level?
► This
is the question we aim to answer.
Approach
►
We assume behavioral improvements are due to
information increases in sensory representations.
(Paradiso 1998, Geisler 1989, Pouget and Thorpe 1991,
Seung and Sompolisky 1993, Lee et al. 1999, Schoups et al.
2001 Adini et al. 2002, Teich and Qian 2003).
►
By information, we mean Fisher Information
It clearly relates to discrimination thresholds
It can be directly computed from first and second-order
statistics (mean and variance).
It can be computed for a population of neurons.
Fisher Information
►
By information, we mean the information about the
stimulus feature (orientation θ), in a pop. of neurons.
►
Response of one neuron in the pop. can be written as:
ri fi ni
►
(Seung and Sompolinsky, 1993)
The Fisher Information for this neuron is:
I
2
f
Activity
ri fi ni
► For a population of neurons with independent noise:
2
N
N
I50 100 150I i
1
Orientation i(deg)
i 1
2
fi
i2
Problems
► We
know that neurons are not independent.
2
fi1
Q 2 f' tr Q 1 Q Q 1 Q
I f
N
T
i 1
► Mechanisms
i
which…
Change tuning curves may also change the correlation
structure
Change correlation structure may also change tuning
curves
Change cross-correlations but not single-neuron statistics
can increase information drastically (Series et. al. 2004)
Investigative Approach
►
We want to use networks of biologically plausible
spiking neurons with realistic correlated noise to study
the neural basis of PL.
►
Therefore, we consider:
Two spiking neuron network models:
► Linear
Non-Linear Poisson (LNP) neurons – analytically tractable
but less biologically realistic
► Conductance-based integrate and fire (CBIF) neurons –
biologically very realistic but analytically intractable
Biologically plausible connectivity
Biologically plausible single-neuron statistics (near unit Fano
factor)
Enough simulations to produce a reasonable lower bound on
Fisher information
Exploring candidate mechanism(s)
for PL
►
We want to investigate changes in Fisher Information
as a result of the following manipulations to network
dynamics:
Sharpening
► Via
feed-forward connectivity
► Via recurrent connectivity
Amplification
► Via
feed-forward connections
► Via recurrent connections
Increasing the number of neurons
►
We use the analytically tractable LNP network to
generate predictions and the CBIF network to confirm
these predictions
Sharpening – LNP Simulations
rmax
Activity spikes/s
0.4
0.35
20
0
-45
0
45
0.25
Orientation (deg)
0.2
I
0.15
0.1
Activity spikes/s
Information (deg-2)
0.3
40
0.05
0
19
20
21
22
23
24
Tuning curve width (Deg)
25
26
40
20
0
-45
0
45
Orientation (deg)
Results - Sharpening
Sharpening by adjusting feed-forward thalamocortical
connections
20
10
0
Activity spikes/s
-45
0
45
Activity spikes/s
Orientation (deg)
Log (variance)
Information
(deg-2)
Orientation (deg)
Activity spikes/s
Activity spikes/s
►
Orientation (deg)
20
10
0
-45
0
45
Orientation (deg)
Log (mean)
Tuning
curve width (Deg)
Results - Sharpening
20
10
0
-45
0
45
Activity spikes/s
Orientation (deg)
Log (variance)
Information
(deg-2)
Orientation (deg)
Activity spikes/s
Activity spikes/s
Sharpening by adjusting recurrent lateral connections
Activity spikes/s
►
Orientation (deg)
20
10
0
-45
0
45
Orientation (deg)
(mean)
TuningLog
curve
width (Deg)
Activity spikes/s
Comparing sharpening schemes
3
2.8
10
0
-45
2.4
0
45
Orientation (deg)
2.2
2
1.8
1.6
1.4
1.2
1
20
22
24
26
28
Tuning curve width (Deg)
30
32
Activity spikes/s
Information (deg-2)
2.6
20
20
10
0
-45
0
45
Orientation (deg)
Future Work
► Exploring
result of:
changes in Fisher information as a
Amplification
Increasing the number of neurons
► Exploring
other ways of increasing
information
► Exploring
Early versus Late theories of
Visual Learning
Conclusion
►
We are interested in investigating the changes at the population level,
that sub-serve the improvement in behavioral performance seen in PL.
►
We follow the prevalent view that improvement in behavioral
performance is due to information increase in the population code.
►
Relaxing the independence assumption no longer allows us to relate
changes at the single-cell level to changes at the population level, in
terms of information throughput.
►
An exploration of the mechanism of sharpening at the population level,
using networks of spiking neurons with realistic correlated noise, yields
the following results:
Sharpening through an increase in feed-forward connections leads to an
increase in information throughput
Sharpening by changing the recurrent lateral connections leads to a
decrease in information throughput