Foundation of Artificial Neural Networks

Explore the McCulloch Pitts neural model, a unidirectional two-layer network with excitatory and inhibitory inputs, binary activation, and adjustable weights and threshold.

Explore the foundation of artificial neural networks: a two-input, one-output network that realizes and, or, and not logic using a weight vector and a threshold in a fully connected structure.

Explore a multi-layer neural network example with XOR to show hidden layers, thresholding, and weight calculations, and discuss learning challenges and future formal algorithms.

Explore the Hebb learning rule in a single-layer feedforward neural network, where weights initialize to zero and adjust automatically during supervised training using input-target pairs.

Demonstrates Hebb nets learning model with binary input and output, showing bias and weights, training and testing phases, weight updates, and a shift to bipolar outputs.

Demonstrates Hebb nets learning with binary inputs and bipolar outputs, using zero-initialized weights and input-target weight updates, testing patterns and showing three learned patterns with one failure.

Learn how the perceptron, a universal feedforward neural network, trains with an iterative Hebbian rule on binary or bipolar inputs, updating weights via the bias and a threshold.

Explore a perceptron example with bias, threshold, and learning rate, using Hebbian updates to adjust weights and classify input patterns until convergence.

Adaline presents the adaptive linear neuron model in a two-layer network with a bias, trained via delta rule or LMS, using bipolar inputs and testing with bipolar step activation.

Learn Madaline, a multi-layer adaptive linear neural network using bipolar step activation, with input-to-hidden weights updated via a learning rate and bias terms.

Explore backpropagation networks, a multi-layer, fully connected, feedforward neural model trained by supervised learning, using generalized delta rule to propagate errors and adjust weights.

Explore the backpropagation neural network architecture, including feed-forward input through input, hidden, and output layers, error backpropagation with the generalized delta rule, and weight updates.

Explore the activation functions used in backpropagation networks, including binary sigmoid, bipolar sigmoid, and hyperbolic tangent, and trace the training algorithm from feed forward to weight updates with error backpropagation.

Explore a backpropagation network with a bipolar sigmoid activation, using a minus one and one input to illustrate feedforward, backpropagation, and weight updates across a two-neuron hidden layer.

This lecture explains auto associative networks, a pattern association model and content addressable memory that learn patterns via weight matrices, retrieve them from noisy inputs, and illustrate capacity n-1.

Explore hetero associative networks that perform pattern association, learn a weight matrix via hebb's rule, and retrieve target values from varied input and output sizes using bipolar activation.

Explore the Kohonen self-organizing map, an unsupervised, competitive neural network that clusters inputs using a distance-based winner-takes-all mechanism and topological preservation.

Explore a self organizing map as unsupervised learning. Use winner-takes-all with Euclidean distance to form clusters, updating the winning unit and weights within a radius on a learning rate schedule.

Explore learning vector quantization, a supervised alternative to self-organizing maps, using a single winning unit, Euclidean or Manhattan distance, and weight updates toward the target or away, without topology preservation.

Explore learning vector quantization, a supervised two-cluster classifier using five input patterns, weight initialization, Euclidean distance to pick a winning unit, and supervised weight updates with learning rate over iterations.