Deep learning is the magic behind the breakthrough success of ML in AI systems in this decade. It is this emerging area of computer science that offers the most promising approach for machine learning and it is revolutionizing artificial intelligence.
Deep learning requires a neural network having multiple layers —?each layer doing mathematical transformations, and feeding into the next layer. The output from the last layer is the decision of the network for a given input. The layers between the input and output layer are called hidden layers.
A deep learning neural network is a massive collection of perceptron’s interconnected in layers. The weights and bias of each perceptron in the network influence the nature of the output decision of the entire network. In a perfectly tuned neural network, all the values of weights and bias of all the perceptron are such that the output decision is always correct (as expected) for all possible inputs. How are the weights and bias configured? This happens iteratively during the training of the network —?called deep learning.
The diagram shows a deep neural network designed for deep learning with multiple layers. Data inputs enter the network into the input layer. Each layer has multiple perceptrons. They transform the inputs and generate outputs which feed into the inputs of the next layer. The interconnection between the layers and the mathematical function of each perceptron is determined by the deep learning network designers. The inputs get successively transformed layer by layer and eventually generate an output decision?—?a value between 0 and 1, indicating the confidence level (as a probability) of the decision for the input data. For example, if the input image is that of a cat and the network has to identify it as a cat, the confidence level should be as close as possible to 1. A lower value indicates that the network is not yet well optimized to identify the cat as a cat.
During the training phase of the neural network, the output is compared with the desired output. Deviations (errors) are back propagated through the network, adjusting and tuning the weights and biases of all the perceptrons in the network using the “cost function.” Learning happens at every tuning of the network parameters. Training the network requires a vast number of cases where the desired output is known. At the conclusion of the training, all the weights and biases of all the perceptrons have been tuned to their final values, and the network is able to deliver the correct decision for all the cases. This is equivalent to having trained a specialist with lots of cases so that they have learned to take the correct decision in all situations. Now the neural network is ready for deployment.
The learning parameters are stored in the weights and biases of individual perceptrons in the network. A large number of perceptrons in the network results in a higher resolution in decision making —?making them more valuable. Most neural networks have thousands of perceptrons. Modern neural networks are usually made up of approximately 10-20 layers and contain around 100 million programmable parameters. Since the algorithmic logic for decision making is spread out in the weights and biases of thousands of perceptrons, it is almost impossible to reconstruct the logic or rationale used for the decision making —?making the AI system based on deep learning neural networks a black box.
Amazon CEO, Jeff Bezos says: “We are now solving problems with machine learning and artificial intelligence that were … in the realm of science fiction for the last several decades. And natural language understanding, machine vision problems, it really is an amazing renaissance.” Bezos calls AI an “enabling layer” that will “improve every business.”