Backpropagation: A supervised learning neural network method

Aug. 5, 2017, 9:53 a.m. By: Vishakha Jha


Backpropagation is another name given to finding the gradient of the cost function in a neural network. It is short for the backward propagation of errors. This algorithm is the classical feed-forward artificial neural network. It is a supervised learning method for multilayer feed-forward which is still used to inculcate large deep learning networks. It can also be used with gradient-based optimizer.

An artificial neuron network (ANN) is basically a computational model constructed on the structure and functions of biological neural networks. They are regarded as tools were the critical association of patterns are found among input and output. And Feed-forward neural networks are based on the concept of data processing of one or more neural cells.

The basic concept behind this approach is to generate an expected output signal from a particular function by transforming internal weightings of input signals. The system is trained in such a way that the difference between the envisioned output and system's output is used to modify its internal state. The basic reason behind using backpropagation is due to the complicated means of solving optimal weights because of which we use gradient descent.

It is an approach towards training the weights in a multilayer feed-forward neural network. A network structure is referred as one or more layers in which each and every layer is completely connected to the other layer.

Backpropagation can be used for both classification and regression problems.

For classification, we let y=0 or 1 represent the two class labels called one hot encoding. The best results are achieved when the network consists of one neuron in the output layer for each class value.

For regression problems, we consider scaling our outputs so that they lie in the [0, 1] range.

Backpropagation is a productive and well-ordered algorithm applied for computation of gradients but has certain flaws associated. The gradient of a node is the product of the error term of this node and the activation of the node in another layer. This continuous multiplication can lead to a lot of incorrect values and lead to change the weights of hidden layers slowly. There are two ways to which one can fix this problem. Fixing the weights or by finding better activation function with a good stable derivate.

For certain situations the Back propagation is a preferred option:

  1. When a great amount of input or output data is accessible, but there are discrepancies regarding how to relate it to the output.

  2. Outputs are considered non-numeric or perplexed.

  3. When it is simple to produce a number of instances corresponding to the correct behaviour.

  4. The problem that corresponds to high level of complexity

Presently the algorithm has a great number of practical applications in the area of artificial intelligence, including optical character recognition, NLP and image processing. It is usually termed as a type of supervised machine learning because it demands a known and expected resultant value corresponding to its input to calculate loss function gradient. Along with Bayesian filters and decision trees, the algorithm has turned out to be crucial in terms of predictive analysis.

Backpropagation Neural Network - How it Works

Video Source: RimstarOrg