Fundamentals of Neural Networks Dr. Satinder Bal Gupta Dr. Satinder Bal Gupta, VCE, Rohtak
Fundamentals of Neural Networks Dr. Satinder Bal Gupta, VCE, Rohtak
Dr. Satinder Bal Gupta, VCE, Rohtak Introduction Dr. Satinder Bal Gupta, VCE, Rohtak
Dr. Satinder Bal Gupta, VCE, Rohtak Why Neural Networks ? Dr. Satinder Bal Gupta, VCE, Rohtak
Dr. Satinder Bal Gupta, VCE, Rohtak History Dr. Satinder Bal Gupta, VCE, Rohtak
Advantages and disadvantages of neural networks One major advantage of neural networks is that they complement symbolic AI. For one, neural networks are based upon the brain, and for two, they are based on a totally different philosophy from symbolic AI. For this reason, neural networks have shown many interesting practical applications which are unique to neural networks. Another major advantage of neural networks is their easy implementation of parallelism since, for example, each neuron can work independently. Generally, developing parallel algorithms for given problems or models (e.g., search, sort, matrix multiplication, etc.) is not easy. Dr. Satinder Bal Gupta, VCE, Rohtak
Advantages of neural networks (cont.) Other advantages are: Learning capability. Neural networks can learn by adjusting their weights. Robustness. For example, neural networks can deal with certain amount of noise in the input. Even if part of a neural network is damaged (perhaps similar to partial brain damage), often it can still perform tasks to a certain extent, unlike some engineering systems, like a computer. Generalization. A neural network can deal with new patterns which are similar to learned patterns. Nonlinearity. Nonlinear problems are hard to solve mathematically. Neural networks can deal with any problems that can be represented as patterns. Dr. Satinder Bal Gupta, VCE, Rohtak
Disadvantages of neural networks First, they have not been able to mimic the human brain or intelligence. Second, after we successfully train a neural network to perform its goal, its weights have no direct meaning to us. That is, we cannot extract any underlying rules which may be implied from the neural network. A big gap remains between neural networks and symbolic AI. Perhaps this situation is essentially the same for the brain - the brain performs at a high level of intelligence, but when we examine it at the physiological level, we see only electrochemical signals passing throughout the natural neural network. A breakthrough for connecting the micro- and macroscopic phenomena in either area, artificial or natural neural networks, may solve the problem for the other. A solution for either area, however, appears unlikely to come in the near future. Third, computation often takes a long time, and sometimes it does not even converge. A counter-argument against this common problem of long time training is that even though it may take a month of continuous training, once it is successful, it can be copied to other systems easily and the benefit can be significant. Fourth, scaling up a neural network is not a simple matter. For example, suppose that we trained a neural network for 100 input neurons. When we want to extend this to a neural network of 101 input neurons, normally we have to start over an entire training session for the new network. Dr. Satinder Bal Gupta, VCE, Rohtak
Biological Neuron Model Dr. Satinder Bal Gupta, VCE, Rohtak
Information Flow in a Neural Cell Dr. Satinder Bal Gupta, VCE, Rohtak
Artificial Neuron Model Dr. Satinder Bal Gupta, VCE, Rohtak
Dr. Satinder Bal Gupta, VCE, Rohtak Functions Dr. Satinder Bal Gupta, VCE, Rohtak
Model of Artificial Neuron Dr. Satinder Bal Gupta, VCE, Rohtak
Artificial Neuron–Basic Elements Dr. Satinder Bal Gupta, VCE, Rohtak
Dr. Satinder Bal Gupta, VCE, Rohtak Basic Elements (Cont.) Dr. Satinder Bal Gupta, VCE, Rohtak
Dr. Satinder Bal Gupta, VCE, Rohtak Example Dr. Satinder Bal Gupta, VCE, Rohtak
Neural Network Architectures Dr. Satinder Bal Gupta, VCE, Rohtak
Single Layer Feed-forward Network Dr. Satinder Bal Gupta, VCE, Rohtak
Multi Layer Feed-forward Network Dr. Satinder Bal Gupta, VCE, Rohtak
Dr. Satinder Bal Gupta, VCE, Rohtak Recurrent Networks Dr. Satinder Bal Gupta, VCE, Rohtak
Learning Methods in Neural Networks Dr. Satinder Bal Gupta, VCE, Rohtak
Classification of Learning Algorithms Dr. Satinder Bal Gupta, VCE, Rohtak
Learning Methods (Cont.) Dr. Satinder Bal Gupta, VCE, Rohtak
Dr. Satinder Bal Gupta, VCE, Rohtak Hebbian Learning Dr. Satinder Bal Gupta, VCE, Rohtak
Gradient decent Learning Dr. Satinder Bal Gupta, VCE, Rohtak
Competitive and stochastic Learning Dr. Satinder Bal Gupta, VCE, Rohtak
Applications of Neural Networks Dr. Satinder Bal Gupta, VCE, Rohtak