Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

66
Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1
  • date post

    15-Jan-2016
  • Category

    Documents

  • view

    223
  • download

    0

Transcript of Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Page 1: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Basic Models in Neuroscience

Oren Shriki

2010

Associative Memory1

Page 2: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Associative Memory in Neural Networks

• Original work by John Hopfield (1982).

• The model is based on a recurrent network with stable attractors.

2

Page 3: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

The Basic Idea• Memory patterns are stored as stable attractors of a recurrent

network.• Each memory pattern has a basin of attraction in the phase space of

the network.

3

Page 4: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

4

Page 5: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Information Storage

• The information is stored in the pattern of synaptic interactions.

5

Page 6: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Energy Function

The dynamics lead to one of the local minima of the energy function, which are the stored memories.

In some models the dynamics are governed by an energy function

6

Page 7: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

7

Page 8: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

8

Page 9: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

9

Page 10: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

10

Page 11: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

11

Page 12: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

12

Page 13: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

13

Page 14: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

14

Page 15: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

15

Page 16: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

16

Page 17: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

17

Page 18: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

18

Page 19: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

19

Page 20: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

20

Page 21: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

21

Page 22: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

22

Page 23: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

23

Page 24: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

24

Page 25: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

25

Page 26: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

26

Page 27: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

27

Page 28: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

28

Page 29: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

29

Page 30: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

30

Page 31: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

31

Page 32: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

32

Page 33: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

33

Page 34: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

34

Page 35: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

35

Page 36: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

36

Page 37: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

37

Page 38: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

38

Page 39: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

39

Page 40: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Important properties of the model

• Content Addressable Memory (CAM) - Access to memory is based on the content and not an address.

• Error correction – The network “corrects” the neurons which are inconsistent with the memory pattern.

40

Page 41: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

The Mathematical Model

41

Page 42: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Binary Networks

• We will use binary neurons: (-1) means ‘inactive’ and (+1) means ‘active’.

• The dynamics are given by:

N

jijiji

ii

thtsJth

thts

1

0 )()()(

)(sgn1

Input from within the network

External input

42

Page 43: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Stability Condition for a Neuron

• The condition for a neuron to remain with the same activity is that its current activity and its current input have the same sign:

0)( thts ii

43

Page 44: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Energy Function

• If the external inputs are constant the network may reach a stable state, but this is not guaranteed (the attractors may be limit cycles and the network may even be chaotic).

• When the recurrent connections are symmetric and there is no self coupling we can write an energy function, such that at each time step the energy decreases or does not change.

• Under these conditions, the attractors of the network are stable fixed points, which are the local minima of the energy function. 44

Page 45: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Energy Function

• Mathematically, the conditions are:

• The energy is given by:

• And one can prove that:

0

ii

jiij

J

JJ

N

i

N

j

N

iiijiij shssJsE

1 1 1

0

2

1)(

0)()1( tEtE45

Page 46: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Setting the Connections

• Our goal is to embed in the network stable stead-states which will form the memory patterns.

• To ensure the existence of such states, we will choose symmetric connections, that guarantee the existence of an energy function.

46

Page 47: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Setting the Connections

• We will denote the P memory patterns by:

• For instance, for a network with 4 neurons and 3 memory patterns, the patterns can be:

P1,2, ,

,

1

1

1

1

,

1

1

1

1

,

1

1

1

1

321

47

Page 48: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Setting the Connections

• Hopfield proposed the following rule:

0 ,1

1

iij

P

iij J N

J

The correlation among neurons across memory patterns

A normalization factor

48

Page 49: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Choosing the Patterns to Store

• To enhance the capacity of the network we will choose patterns that are not similar to one another.

• In the Hopfield model, (-1) and (+1) are chosen with equal probabilities. In addition, there are no correlations among the neurons within a pattern and there are no correlations among patterns.

49

Page 50: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Memory Capacity

• Storing more and more patterns adds more constraints to the pattern of connections.

• There is a limit on the number of stable patterns that can be stored.

• In practice, a some point a new pattern will not be stable even if we set the network to this pattern.

50

Page 51: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Memory Capacity

• If we demand that at every pattern all neurons will be stable, we obtain:

N

NP

ln2max

51

Page 52: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Memory Capacity• What happens to the system if we store more patterns?

• Initially, the network will still function as associative memory, although the local minima will differ from the memory states by a few bits.

• At some point, the network will abruptly stop functioning as associative memory.

52

Page 53: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Adding “Temperature”

• It is also interesting to consider the case of stochastic dynamics. We add noise to the neuronal dynamics in analogy with the temperature in physical systems.

• Physiologically, the noise can arise from random fluctuations in the synaptic release, delays in nerve conduction, fluctuations in ionic channels and more.

53

Page 54: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Adding “Temperature”

-5 -4 -3 -2 -1 0 1 2 3 4 50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1S=1

h

P(s)

54

Page 55: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Adding “Temperature”

-5 -4 -3 -2 -1 0 1 2 3 4 50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1S=-1

h

P(s)

55

Page 56: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Adding “Temperature”• Adding temperature has computational advantages:

It drives the system out of spurious local minima, such that only the deep volleys in the energy landscape affect the dynamics.

• One approach is to start the system at high temperature and then gradually cool it down and allow it to stabilize (Simulated annealing).

• In general, increasing the temperature reduces the storage capacity but can prevent undesirable attractors. 56

Page 57: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Associative Memory - Summary

• The Hopfield model is an example of connecting between dynamical concepts (attractors and basins of attraction) and functional concepts (associative memory).

• The work pointed out the relation between neural networks and statistical physics and attracted many physicists to the field.

57

Page 58: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Associative Memory - Summary

• Over the years, models that are based on the same principles but are more biologically plausible were developed.

• Attractor networks are still useful in modelling a wide variety of phenomena.

58

Page 59: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

References• Hopfield, 1982

– Hopfield, J. (1982). Neural networks and physical systems with emergent collective computational properties. Proceedings of the National Academy of Sciences of the USA, 79:2554 - 2588.

• Hopfield, 1984 – Hopfield, J. (1984). Neurons with graded response have

collective computational properties like those of two-state neurons. Proceedings of the National Academy of Sciences of the USA, 81:3088 - 3092.

59

Page 60: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

מקורות

• Amit, 1989 – Amit, D. Modeling Brain Function. Cambridge

University Press, 1989

• Hertz et al., 1991 – John Hertz, Anders Krogh, Richard G. Palmer.

Introduction to the Theory of Neural Computation. Addison-Wesley, 1991.

60

Page 61: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Associative Memory of Sensory Objects – Theory and

ExperimentsMisha Tsodyks ,

Dept of Neurobiology, Weizmann Institute, Rehovot, Israel

Joint work with Son Preminger and Dov Sagi

61

Page 62: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Non Friends

… …

Experiments - Terminology

Friends

62

Page 63: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Experiment – Terminology (cont.)

•Basic Friend or Non-Friend task (FNF task)–Face images of faces are flashed for 200 ms –for each image the subject is asked whether the

image is a friend image (learned in advance) or not .–50% of images are friends, 50% non-friends, in

random order; each friend is shown the same number of times. No feedback is given

F/NF

?

200ms

F/NF

?F/NF

?

200ms 200ms 200ms 200ms 200ms 200ms 200ms 200ms63

Page 64: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Morph Sequence

1 … 20

Source(friend)

Target(unfamiliar)

… 40 … 60 … 80 100…

Experiment – Terminology (cont.)

64

Page 65: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Two Pairs: Source and Target

Pair 1Pair 2

65

Page 66: Basic Models in Neuroscience Oren Shriki 2010 Associative Memory 1.

Subject HL -------- (blue-green spectrum) days 1-18

FNF-Grad on Pair 1

Bin number

Num

ber

of ‘F

riend

’ res

pons

es

66