Ex1 2017 1

Neural networks
Statistics, Optimisation, and Learning

Examples sheet 1
1. Deterministic Hopfield model. Write a computer program imple-

menting the Hopfield model with synchronous updating according to

N
X
Si sgn wij Sj
j=1
where Si is the state of neuron i, N is the number of neurons, and wij are
the synaptic weights given by Hebbs rule for all i and j. Generate and store
p random patterns.
1a. Following the calculation on p. 18-22 in Lecture notes, derive an ap-
proximation for the one-step error probability PError valid for large p and N.
Take into account that wii 6= 0. (1p)
1b. Check the approximation derived in 1a as follows. Set in your com-
puter program the number of neurons to N = 200, and use the following
values of p: 1, 20, 40, 60, 80, 100, 120, 140, 160, 180, 200, 220, 240, 260,
280, 300, 320, 340, 360, 380, 400. For each value of p, estimate the one-step
error probability PError based on 10000 bits (note that this requires that you
generate multiple samples of p random patterns, because the number of bits
in a single sample is equal to pN). Plot your numerically determined PError
as a function of p/N. Plot also the corresponding theoretical curve that you
derived in 1a. Discuss your result. (1p)
2. Stochastic Hopfield model: phase diagram. Write computer pro-

gram implementing the Hopfield model (take wii = 0) with asynchronous
stochastic updating.
2a. Set the number of neurons to N = 200 and store p = 5 random patterns,
denoted by (i) (i = 1, . . . , p). For noise parameter = 2 feed the stored
pattern (1) to the network and monitor the corresponding order parameter
m1 as a function of time (one unit of time corresponds to one update). Re-
peat this experiment 20 times. For each experiment, plot the resulting m1
as a function of time (make one plot with 20 curves). Describe the results.
Discuss: How long time does it take for the network to reach the steady
state? Based on m1 that you obtain in the different experiments, explain
whether or not the network performs well. Is your finding consistent with
the phase diagram illustrated on page 60 in Lecture notes? Note that the
phase diagram on page 60 in Lecture notes is valid in the limit of N ,
whereas in your simulation N is finite. Discuss the effect of finite N on the
order parameter m1 . (1p)
2b. Repeat the task 2a but now for p = 40 (keeping all other parameters the
same as in 2a). Plot m1 as a function of time for the different experiments
1
made with p = 40. Describe the results and compare to the results in 2a.
Is the order parameter closer to unity in this case or in the case analysed in
2a? Why? Explain by referring to the phase diagram illustrated on page 60
in Lecture notes. (1p)
3. Back propagation. The task is to write a computer program solving
a classification problem using back-propagation. The goal is to achieve a
classification error that is as small as possible for a given validation set by
training the network on the corresponding training set. The classification
error Cv for the validation set is defined as
p
1 X
Cv = | () sgn(O () )|
2 p =1
where |x| stands for the absolute value of x, p denotes the total number of
patterns in the validation set, () is the target output for pattern , and
O () is the network output for pattern . The classification error for the
training set is defined similarly. Implement your program on the data set
available on the course web page: train data 2017.txt containing training
data, and valid data 2017.txt containing validation data. Patterns are two-
dimensional (first two columns in the data files). Each row corresponds to
a different pattern. The third column contains target outputs. To improve
the training performance, normalise the input data so that both the first and
the second column have zero mean and unit variance. Tasks:
3a. Train the network without hidden layers. Use asynchronous back-
propagation with the following parameters: learning rate = 0.02, = 1/2,
weights initialised randomly from the uniform distribution in the interval
[0.2, 0.2], biases initialised randomly from the uniform distribution in the
interval [1, 1]. Take activation functions to be g(b) = tanh(b) (Lecture
notes, p. 103). Iterate for 106 iterations (one iteration corresponds to feed-
ing one randomly chosen pattern to the network). Perform ten independent
training runs. During each training, plot the energy of the training set as
well as of the validation set as a function of the training time. Discuss the
trends observed. Explain: how is the energy expected to change over time
according to the gradient-descent rule? (0.5p)
3b. Plot the classification errors obtained during these ten training experi-
ments both for the training set and for the validation set. Discuss how well
the network performs. (0.5p)
3c. Now train a network with one hidden layer that has four neurons. As
in task 3a, perform ten independent runs. For each run, plot the energy in
the training set as a function of the training time, as well as the energy of
the validation set as a function of the training time. Discuss the results in
comparison to those obtained in 3a. Is the performance better? Why? Use
a graphical illustration of the classification problem to support your answer.
(0.5p)
3d. Make a table showing the classification error at the end of each train-
ing for both the training set and the validation set. Discuss: would adding
another hidden layer improve the classification? Why? (0.5p)

Ex1 2017 1

Uploaded by

Copyright:

Available Formats

Ex1 2017 1

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ex1 2017 1

Uploaded by

Copyright:

Available Formats

Neural networks

Statistics, Optimisation, and Learning

1. Deterministic Hopfield model. Write a computer program imple-

2. Stochastic Hopfield model: phase diagram. Write computer pro-

You might also like