DEEP LEARNING

Deep learning, a branch of artificial intelligence, has become prominent only recently, it is based on concepts that are familiar to chemical engineers. This article describes artificial neural networks the algorithms that enable deep learning. Artificial intelligence is the capability of a machine to imitate intelligent human behavior . Machine learning (ML) is a branch of AI that gives computers the ability to “learn” often from data without being explicitly programmed. Deep learning is a subfield of ML that uses algorithms called artificial neural networks (ANNs), which are inspired by the structure and function of the brain and are capable of self-learning. ANNs are trained to “learn” models and patterns rather than being explicitly told how to solve a problem.

The building block of an ANN is called the perceptron, which is an algorithm inspired by the biological neuron?. Although the perceptron was invented in 1957, ANNs remained in obscurity until just recently because they require extensive training, and the amount of training to get useful results exceeded the computer power and data sizes available. To appreciate the recent increase in computing power, consider that in 2012 the Google Brain project had to use a custom-made computer that consumed 600 kW of electricity and cost around $5,000,000. By 2014, Stanford AI Lab was getting more computing power by using three off-the-shelf graphics processing unit (GPU)-accelerated servers that each cost around $33,000 and consumed just 4 kW of electricity. Today, you can buy a specialized Neural Compute Stick that delivers more than 100 gigaflops of computing performance for $80.

The average human brain has approximately 100 billion neurons. A human neuron uses dendrites to collect inputs from other neurons, adds all the inputs, and if the resulting sum is greater than a threshold, it fires and produces an output. The fired output is then sent to other connected neurons. A perceptron is a mathematical model of a biological neuron?(6). Just like a real neuron, it receives inputs and computes an output. Each input has an associated weight. All the inputs are individually multiplied by their weights, added together, and passed into an activation function that determines whether the neuron should fire and produce an output.