Oct 01, 2021 19:00:00

How is 'deep learning' done, which lets a computer learn things?

'Deep learning, ' a machine learning method that uses neural networks that automatically extract features from data, has rapidly developed in recent years and is useful for the development of image recognition AI and autonomous driving technology. IEEE Spectrum published by the Institute of Electrical and Electronics Engineers explains the mechanism of such deep learning.

How Deep Learning Works
https://spectrum.ieee.org/what-is-deep-learning/step-4

Deep learning is performed on a neural network that connects a large number of artificial neurons. This neural network is divided into 'input layer (INPUT LAYER)', 'hidden layer (HIDDEN LAYERS)' and 'output layer (OUTPUT LAYER)'. IEEE Spectrum explains the mechanism of deep learning by learning how to recognize handwritten numbers in a neural network with two hidden layers.

First, in this example, the area where the numbers are handwritten is divided into multiple squares, and one artificial neuron in the input layer is assigned to each square. These artificial neurons 'fire' when the filled area of each square exceeds a certain value, and pass an electrical signal to the artificial neurons in the hidden layer. When learning is complete, an appropriate 'weight' is multiplied when the electrical signal is passed to the hidden layer, and an artificial neuron that exceeds a certain amount of passed electrical signal fires. By repeating this, the artificial neurons in the output layer corresponding to the handwritten numbers are finally fired. In other words, by adjusting the weight when each signal is passed, it is possible to improve the accuracy of recognizing handwritten numbers.

If learning is not progressing, the handwritten number '3' may be recognized as '4'. When the input value and the output value are different, the weight is set so that the correct output result can be obtained in the order of 'output layer', 'second hidden layer', 'first hidden layer', and 'input layer'. By adjusting, learning will proceed.

For example, if you read a handwritten '3' and output a '4', it means that the artificial neuron that outputs the '4' received the most electrical signals. In this case, the weight of the entire neural network needs to be adjusted so that the artificial neuron that outputs '3' instead of '4' receives the most electrical signals.

To make '3' receive the most electrical signals, increase the weight when passing the signal from the artificial neuron in the second hidden layer to the artificial neuron that outputs '3'.

Then the same weight adjustment is done for the first hidden layer. For example, in this example, the weight of the electrical signal passed to the artificial neuron is adjusted to be small in order to suppress the firing of the artificial neuron located at the top of the second hidden layer.

Finally, the weights are adjusted in the input layer as well. Now, when the handwritten '3' displayed on the left side of the figure below is recognized, the neuron that outputs the '3' receives the most electrical signal.

In deep learning, learning is advanced by repeating the above weight adjustments in response to a huge number of inputs, and deliverables such as highly accurate handwriting recognition models are created.