Transcript My final poster

Memory Network Maintenance Using
Spike-Timing Dependent Plasticity
David Jangraw, ELE ’07
Advisor: John Hopfield, Department of Molecular Biology
category
i(mA)
10
• Each neuron is modeled as a parallel RC circuit with a
firing threshold
5
6
0
0
0.02
15
Br
Bl
2
Gr
4
6
property
0.08
0.1
time(s)
output voltage
0.12
8
10
• The firing frequency ‘f’ is our measure of activity.
5
Each black square in this grid represents an
active neuron encoding something about a
person. For example, if the bottom row of
neurons encodes eye color, and the ‘green’
neuron is active, this person has green eyes.
-5
0
0.02
0.04
0.06
0.08
time(s)
0.1
0.12
70
frequency (Hz)
Synaptic Drift (δT):
Small, random changes in synaptic strength due to “noisy”
cellular processes
• This will make the neuron fire more quickly than the
neurons connected to it (-δt)
50
40
• STDP would transform this spike timing delay into
negative changes in synaptic strength (-δT)
30
• This would produce a proportional negative change in
input current (+δI) that could stabilize the memory!
20
10
Normal memories have equal activity in each participating
neuron, but synaptic drift causes unequal activity (a
corrupted memory) or even memory loss.
0
0
2
4
6
8
10
12
Io = DC input current (mA)
14
16
18
20
Non-Memory
+δT  +δI  –δt  –δT  –δI
-2
• STDP applied based on average delay
after 50 spikes of each neuron
-3
-4
• STDP successfully used to
synchronize firing of many neurons
-5
-6
-7
0
0.5
1
1.5
2
2.5
3
applications of STDP rule (x50 spikes)
3.5
• Slope of STDP rule affects speed and
stability of synchronization
4
Synchronization of two neurons
0.4
0.2
1
-0.6
-0.8
m=5
m=30
m=50
m=80
m=100
-1
-1.2
1
2
3
4
5
6
7
8
applications of STDP rule (x50 spikes)
0.6
18
0.2
delta T
- Neuron Activity
0.4
12
-0.6
-0.8
11.5
12
12.5
Io (mA)
13
13.5
14
Changes in input current (while still on the f-I plateau)
lead to changes in spike timing.
13
12.8
12.4
12.2
12
11.8
11.6
11.4
11.2
1
1.5
2
2.5
3
3.5
applications of STDP
4
Future Directions
-0.4
10
current through active neurons
0
-0.2
14
10
Drift Correction
• This indicates that spiking neurons could
correct for synaptic drift using STDP
20
16
9
Two neurons with slightly different input currents are synchronized using STDP rules with
different slopes (m). Low m produced slow convergence; high m produced damped
oscillation; extremely high m (not shown) produced instability.
• Each drifted memory did converge
to an ideal memory
0.8
22
time of spike relative to input max (ms)
-0.4
• Applied random synaptic drift (≤5%) to
each active connection in a memory.
A Simple STDP Rule
24
8
11
-0.2
12.6
final timing of spikes: A=5, w=30
Figure (except titles) from J. J. Hopfield (2006). “Searching for memories, Sudoku, implicit check-bits,
and the iterative use of not-always-correct rapid neural computation.” arXiv.org, 19 Sep 2006.
-1
• Simplified memory network using
continuous-variable neurons with spiking
neuron’s f-I curve and spike timing patterns
AC input produces a plateau on the neuron’s f-I curve.
…Can the system correct itself?
Drifted Memory
IAC
60
The Problem:
Ideal Memory
0.16
Stability Analysis:
• A small positive change in T (synaptic drift, +δT) will
produce a proportional positive change in the input
current (+δI) to the receiving neuron
frequency response of neuron: A=5mA
w=30Hz
I = Io + 0.05
sin(2π wt)
Synaptic Strength (matrix T):
T21 is proportional to the size of the electrical response in
neuron 2 evoked by an electrical spike in neuron 1
T12
0.14
Spike-Timing Dependent Plasticity
Synapse:
A connection between two neurons
2
• If we feed an AC input of the form:
I=Io + A sin(2π wt)
into a neuron, a large range of DC offsets (Io) will drive
the cell to fire at a frequency w, creating a plateau on
the f-I curve (below left).
• Modeled disconnected neurons given
slightly different input currents, causing
varying delays in spike timing
0
80
1
• When the membrane voltage exceeds threshold, the
cell fires.
0.16
0
Synaptic Drift
T21
0.14
10
…
v (mV)
2
0.06
t
4
Eyes:
0.04
Synchronization of 51 neurons, m=0.1
0
input current (mA)
• We will model the activity of these neurons
using MATLAB
15
8
• Each property of the person or thing being
remembered is represented by the activity of
one neuron
• All neurons active in a memory are
connected, so thinking of one thing about a
memory (i.e. a name) will recall other things
(i.e. height, eye color)
20
10
• Associative memory is not fully understood;
here we refine a simplified model of it
Spike-Timing Dependent Plasticity (STDP):
Experimentally observed phenomenon by which relative
spike timing changes synaptic strength
input current
Sample Memory
Synchronization
average delay (t2-t1), in ms
Associative Memory:
Learned connections between ideas one
remembers as being associated
Stabilization of Firing
average delay (t2-t1), in ms
Associative Memory
-1
-4
• Load multiple, overlapping memories into network and test performance
-3
-2
-1
0
delay (ms)
1
2
3
Changes in spike timing are converted to
proportional changes in synaptic strength.
Spikes too far separated are ignored.
4
• Create converging memory network of spiking neurons
• Use more realistic STDP rule
4.5
5