Psychologische Modelbouw

Connectionism Lecture by Jaap Murre

Connectionism Demos

Lab Assignment for Connectionism

Teach the two patterns to the Hopfield network, then try to retrieve a distorted version. Do you observe the pattern retrieval (= content-addressable memory)? Why does the Hopfield network sometimes give the complement of the pattern learned (i.e., the 'negative')? How much distortation can the network take? Now, repeat this after having taught it other random patterns. How many random patterns can you store additionally before errors appear in the retrieval of the face?

Observe how the Perceptron can learn some of the data sets well, but not the others. Which one can be learned? Which one gives trouble? Why?

Now repeat this with the MLP (multi-layer perceptron or backprop network). Does the system manage this well now? How long does it take to learn?

Second Connectionism Lecture: Models of Memory by Jaap Murre

The file was too large to upload from home; I will do it later.

Lab Assignment for Second Connectionism Lecture

Image Compression

Image compression is a bit akin to what the hippocampus may be doing: a rich and complex firing pattern is 'indexed' by a relatively small hippocampus. This demo shows how an auto-encoder neural network can compress an image, form a compact representation. Things to try after studying the principles. (1) First try the demo as provided. (2) How many training steps do you need for a good generalization? (3) How does the training generalize to other faces? What if you train on a woman and test on a man? Which training-test exemplars generalize badly and why do you think this is the case? (4) If you increase the learning rate and train again, does it go more rapidly? (5) Can you think of ways to improve this network so that performs better?

Java demonstrations of neural net concepts

Applets for neural networks and artificial life