Competitive learning is an established branch of the general theme of unsupervised learning.
Abstract
Competitive learning is an established branch of the general theme of unsupervised learning.The elementary principles of competitive learning are:
1. Start with a set of units that are all the same except for some randomly distributed parameter which makes each of them slightly differently to a set of input patterns.
2. Limit the ‘strength’ of each unit.
3. Allow the units to compete in some way for the right to respond to a given subset of inputs.
Applying these three principles yields a learning paradigm in which individual units learn to specialize on sets similar patterns and thus become ‘feature detectors’.
The Problem
The EEG signal is registration of electrical activity in the brain. The activity us registered by placing electrodes on the scalp and measuring the voltage on them. The electrical activity is do firing great number of neurons in the brain. Temple firing always appears in the measures, but different patterns appear as a result of awakening state and other parameters.
The parts of the EEG that comes from external evoking belongs to the part of the brain activity that is called evoked potential EP
Different amplitudes and frequencies characterize slow potentials that are usually interest. Another characterize is that they have a very low signal to noise ratio SNR. This makes it very hard to identify them from signal measurements.
Solution
Our solution based on Competitive Learning Algorithm. The solution applied to wide range of problems were required recognizing of weak signals on the noise background.
A typical architecture of a competitive learning system appears in Fig. (1). The system consists of a set hierarchically layered neurons in which each layer is connected via excitatory connections to the following layer. Within a layer, the neurons are divided into sets of inhibitory clusters in which all neurons within a cluster inhibit all other neurons in the cluster, resulting in a competition among the neurons to respond to the pattern appearing on the previous layer; the stronger a neuron respond to an input pattern, the more it inhibits the other neurons if it’s cluster.
There are many variations of the competitive learning scheme. We have selected a single layer structure, where the output neurons are fully connected to the input nodes and the non-linearity is implemented in the learning-phase only. The advantage of using this structure lies in enhanced analysis capabilities of the converged network, as the weights actually converge to the embedded signal patterns and thus form a Pattern Identification Network. The general network structure is depicted in Fig. (2). For neuron j to be the winning neuron, its net internal activity level vj for a specified input pattern xj must be the largest among all neurons in the network. The output signal yj of a winning neuron j is set equal to one, and all other neuron outputs that lose the competition are set equal zero.
This is an output result of running our algorithm:
Acknowledgements
We would like to thank our supervisor, Igor Makienko , for his patience and for guiding us through the project. We would also like to thank Johanan Erez and the rest of the laboratory staff for their help and support.