Transcript Pattern Recognition using Hebbian Learning and Floating-Gates

Pattern Recognition using Hebbian Learning
and Floating-Gates
Certain pattern recognition problems have been shown to be easily solved by
Artificial neural networks and many neural network chips have been made and sold.
Most of these are not terribly biologically realistic.
Output layer neurons
weights
Hidden layer neurons
weights
Input layer neurons
A 2-dimensional example…
x (-10:10) y(-10:10)
w1 = 0.3, w2 = =0.7
x
y
Output Decision Space
50
50
100
100
Y
Y
Summation of weighted inputs
150
+1, -1
150
200
200
50
100
X
150
200
50
100
X
150
200
x (-10:10) y(-10:10)
w1 = 0.5, w2 = 0.11
x
y
Outputs over the Input Space
50
50
100
100
Y
Y
Weighted Summation over the Input Space
150
+1, -1
150
200
200
50
100
X
150
200
50
100
X
150
200
Putting the two together…
x
y
50
-.53
y
0.53 -0.25
We respond to a smaller region of this
2-D input space.
+1, -1
100
150
200
50
100
x
150
200
So in general, we can apply this type of operation on an N-dimensional input
With the hidden units defining hyperplanes in this input space. The individual
output units combine these hyperplanes to create specific subregions of this N-dim
space. This is what pattern recognition is about.
100 x 77 pixels = 7700 dimensional
input space
As you might expect, these two images
live very far apart from each other in
this very high dimensional space.
10
20
30
40
50
60
70
unit1
80
90
Easy Task
100
10
20
30
40
50
60
70
unit2
10
20
30
But if we had a set of 100 faces that we
wanted to recognize, this might be harder.
40
50
60
70
80
90
100
10
20
30
40
50
60
70
What happens if the faces are
rotated, shifted, or scaled?
How do I pick the weight matrices to solve these tasks??
One way is to present inputs and adjust the weights if the output is not what we want.
wiknew  wikold  wik
where
input k, example 
wik  2 i k
= 0
Learning rate
if
 i  Oi
otherwise
Output unit i, example 
Target output for unit i
example   /- 1
This is known as the perceptron learning rule
A training set of examples with target output values is defined and presented
one by one, adjusting the weight matrix after each evaluation. The learning rule
Assigns large value weights to components of the inputs that allow discrimination
between the classes of inputs.
e.g., many faces and many helicopters
Face vs. Helicopter Example
5000
Red = unit 1 response to face
4000
Blue = unit 1 response to helicopter
3000
Green = unit 2 response to face
2000
Black = unit 2 response to helicopter
1000
0
-1000
-2000
-3000
-4000
-5000
0
10
20
30
40
50
60
70
80
90
100
Associative Memory and Energy Functions
Si
inputs
10
10
20
20
30
30
40
40
50
50
60
60
70
70
80
80
90
90
100
100
10
wij
matrix
20
30
40
50
60
70
10
20
30
40
50
60
70
The Hopfield Recurrent Network
The concept of an energy function of a recurrent neural network was introduced
by Hopfield (1982) to describe the state of a network. By studying the dynamics,
it is possible to show that the activity in the network will always decrease in energy,
evolving towards a "local minima".
H -
1
wijSiSj

2 ij
The network defines an
'energy landscape' in which
the state of the network settles.
By starting close to minima (stored patterns) compared to other points in the landscape
The network will settle towards the minima and 'recall' the stored patterns.
This view of neural processing has its merits, provides insight into this type of
computational structure and has spawned new fields on its own, but does not describe
the current neurobiological state of knowledge very well.
In particular, neurons communicate with spikes and the backpropagation learning rule
Is not a good match to what has been found.
So what do we know about neurobiological learning?
Hebbian learning
If both cells are active,
strengthen the synapse
If only the post-synaptic cell
is active, weaken the synapse
In fact, learning at some synapses seems to be even more specific.
Temporal ordering seems to play a role in determining the change in the synapse.
strengthen
weaken
Dw
Time between pre-syn and post-syn spikes
Abbott and Blum, 1996
Chip Idea:
1. Design a spiking neural network that can learn using the spike-timing rule
to solve a particular temporal pattern recognition problem
2. Design a floating-gate modification circuit that can implement the learning rule