szymonlyszkowski / soft-computing

Soft-Computing

This project contains solutions for soft-computing subject labs.

##Task 0 Linear Neuron

Main purpose of this task was artificial neuron model implementation. Simple neuron is activated by identity function and is it trained using Delta principle. Task should prove that neural network is trained correctly for equation set with exactly one solution.

##Task 1 MADALINE Network

Objective of this exercise was to implement Multiple Adaptive Linear (MADALINE) network which is able to recognize letter patterns. Letters are implemented as arrays of 0 and 1 which corresponds respectively to white and black pixel. Each neuron has fixed weights which are not trained during network execution. Result of MADALINE network execution are neuron outputs. Neuron with highest output value which can be <0,1> is chosen as recognized letter. Package training_set_letters/letters.py contains patterns for letters wchich can be recognized.

##Task 2 Kohonen Network

Objective of this exercise was to implement Kohonen network which will be able to compress images. Implementation handles gray-scale images. Images should be squares and frame used to scan image have dimensions: 4x4, 8x8, 16x16. Such image array is encoded to list of tuples where elements are: (frame id, neuron_decoding). Using such object to as input to neural network it is possible to obtain compressed image.

##Task 3 Multilayer Perceptron

Objective of this exercise was to implement Perceptron consisting of 3 layers: input(copying), hidden, output. Method used to train network was backward propagation of error. Perceptron is supposed to recognize 4 vectors: [1,0,0,0],[0,1,0,0],[0,0,1,0],[0,0,0,1].