Neural NetworksOverviewIntroductionBiological Neural NetworksSlide 5Artificial Neural NetworksSlide 7Slide 8ArchitecturesFeed-Forward NetworksFeedback NetworksLearning TechniquesSupervised LearningUnsupervised LearningAdvantagesApplicationsExamplesOptical Character RecognitionEndReferencesNeural NetworksSteven LeOverview•Introduction•Architectures•Learning Techniques•Advantages•ApplicationsIntroduction•A Neural Network is data processing model which is composed of a large number of processing elements which individually handle one piece of a larger problem •Two main types of neural networksBiological Neural Networks•The human brain is a neural network•Nervous system is composed of neurons•Signals travel into the neuron via dendrites•Signals are sentout via the axon•Signals coming into the dendrite can be either exhibitive or inhibitive•Synapses may add resistance before adding•A Threshold determines if the neuron is excited enough to send a signal out through the axonArtificial Neural Networks•Try to simulate how biological neural networks process information•Acquires knowledge through learning•Knowledge is stored within inter-neuron connection strengths known as synaptic weights.Model of an Artificial Neuron•Synaptic weights are multiplied with an input to give the weighted input•Activation function computes the values of every input and if they exceed the threshold, the neuron will fire•Output, like the biological version, can either be -1 or 1 (alternatively 0 or 1)ArchitecturesFeed-Forward Networks•Signals only travel in one direction•Output of a layer doesn’t affect the same layerFeedback Networks•Signals travel any direction and can loop•Node states are always changing until an equilibrium is reached•Remains at restuntil new input isintroduced or newequilibrium is neededLearning Techniques•Before they are used, neural networks go through a learning phase in which they acquire knowledgeSupervised Learning•Incorporates an external teacher so that each output unit is told what its desired response to input signals should be•The aim is to determine a set of weights which minimizes the error between actual and desired outputsUnsupervised Learning•Uses no external teacher and is based upon only local information. •Also known as Self-Organization, the output unit is trained to respond to clusters of pattern within the input•No pre-set categoriesAdvantages•Parallelism: Neurons act independently•Adaptive learning•Self-organization•Fault tolerance•Interacting with noisy dataApplications•Since neural networks are best at identifying patterns or trends in data, they are well suited for prediction or forecasting•Some applications include: targeted marketing, voice recognition, financial forecasting, data validation, and credit evaluationExamples•A company has a database of 1million potential customers. 20,000 (2%) response is the goal•Contact 100,000. Use this subset to train the neural network•Present the other 900,000 to the neural network which will classify 2% of them as buyersOptical Character RecognitionEndReferences•Null, Linda. Computer Organization and
View Full Document