Neurons and neural networks II. Hopfield network

Transcription

1 Neurons and neural networks II. Hopfield network 1

2 Perceptron recap key ingredient: adaptivity of the system unsupervised vs supervised learning architecture for discrimination: single neuron perceptron error function & learning rule gradient descent learning & divergence regularisation learning as inference 2

3 Interpreting learning as inference So far: optimization wrt. an objective function where 3

4 Interpreting learning as inference So far: optimization wrt. an objective function where What s this quirky regularizer, anyway? 3

5 Interpreting learning as inference Let s interpret y(x,w) as a probability: 4

6 Interpreting learning as inference Let s interpret y(x,w) as a probability: in a compact form: 4

7 Interpreting learning as inference Let s interpret y(x,w) as a probability: in a compact form: the likelihood of the input data can be expressed with the original error function function 4

8 Interpreting learning as inference Let s interpret y(x,w) as a probability: in a compact form: the likelihood of the input data can be expressed with the original error function function the regularizer has the form of a prior! 4

9 Interpreting learning as inference Let s interpret y(x,w) as a probability: in a compact form: the likelihood of the input data can be expressed with the original error function function the regularizer has the form of a prior! what we get in the objective function M(w): the posterior distribution of w: 4

10 Interpreting learning as inference Relationship between M(w) and the posterior interpretation: minimizing M(w) leads to finding the maximum a posteriori estimate w MP The log probability interpretation of the objective function retains: additivity of errors, while keeping the multiplicativity of probabilities 5

11 Interpreting learning as inference Properties of the Bayesian estimate 6

12 Interpreting learning as inference Properties of the Bayesian estimate The probabilistic interpretation makes our assumptions explicit: by the regularizer we imposed a soft constraint on the learned parameters, which expresses our prior expecations. An additional plus: beyond getting w MP we get a measure for learned parameter uncertainty 6

13 Interpreting learning as inference Demo 7

14 Interpreting learning as inference Demo 7

15 Interpreting learning as inference Demo 7

16 Interpreting learning as inference Demo 7

17 Interpreting learning as inference Making predictions Up to this point the goal was optimization: 8

18 Interpreting learning as inference Making predictions Up to this point the goal was optimization: 8

19 Interpreting learning as inference Making predictions Up to this point the goal was optimization: Are we equally confident in the two predictions? 8

20 Interpreting learning as inference Making predictions Up to this point the goal was optimization: Are we equally confident in the two predictions? The Bayesian answer exploits the probabilistic interpretation: 8

21 Predictive probability: Interpreting learning as inference Calculating Bayesian predictions 9

22 Predictive probability: Interpreting learning as inference Calculating Bayesian predictions Likelihood: Weight posterior Partition function: 9

23 Predictive probability: Interpreting learning as inference Calculating Bayesian predictions Likelihood: Weight posterior Partition function: Finally: 9

24 Interpreting learning as inference Calculating Bayesian predictions How to solve the integral? 10

25 Interpreting learning as inference Calculating Bayesian predictions How to solve the integral? Bad news: Monte Carlo integration is needed 10

26 11

27 12

28 Interpreting learning as inference Calculating Bayesian predictions 13

29 Interpreting learning as inference Calculating Bayesian predictions Original estimate 13

30 Interpreting learning as inference Calculating Bayesian predictions Original estimate Bayesian estimate 13

31 Interpreting learning as inference Gaussian approximation Taylor expansion around the MAP estimate 14

32 Interpreting learning as inference Gaussian approximation Taylor expansion around the MAP estimate The Gaussian approximation: 14

33 Interpreting learning as inference Gaussian approximation Taylor expansion around the MAP estimate The Gaussian approximation: 14

34 Neural networks Unsupervised learning Capacity of a single neuron is limited: certain data can only be learned So far, we used a supervised learning paradigm: a teacher was necessary to teach an input-output relation Hopfield networks try to cure both Unsupervised learning: what is it about? Hebb rule: an enlightening example assuming 2 neurons and a weight modification process: This simple rule realizes an associative memory! 15

35 Neural networks The Hopfield network Architecture: a set of I neurons connected by symmetric synapses of weight w ij no self connections: w ii =0 output of neuron i: x i Activity rule: Learning rule: Synchronous/ asynchronous update ; 16

36 Neural networks The Hopfield network Architecture: a set of I neurons connected by symmetric synapses of weight w ij no self connections: w ii =0 output of neuron i: x i Activity rule: Learning rule: Synchronous/ asynchronous update alternatively, a continuous network can be defined as: ; 16

37 Are the memories stable? Neural networks Stability of Hopfield network Necessary conditions: symmetric weights; asynchronous update 17

38 Are the memories stable? Neural networks Stability of Hopfield network Necessary conditions: symmetric weights; asynchronous update 17

39 Are the memories stable? Neural networks Stability of Hopfield network Necessary conditions: symmetric weights; asynchronous update Robust against perturbation of a subset of weights 17

40 Neural networks Capacity of Hopfield network How many traces can be memorized by a network of I neurons? 18

41 Neural networks Capacity of Hopfield network 19