In the recent years, deep learning, a machine learning algorithms, has seen tremendous success in the recent years. Companies such as Google and Facebook have invested a large amount of money into deep learning. Speech and object recognition software that uses deep learning is significantly more accurate compare to software that uses traditional methods for speech and object recognition. In fact, AlphaGo, a deep-reinforcement learning application that plays the board game GO, successfully beat Lee Sedol, the grandmaster in GO. This feat has long considered impossible with traditional methods. So how does deep learning work? Deep learning uses neural networks and backpropagation.
Figure 1: an artificial neural network
Credit to wikimedia
Although there are many variations of deep learning algorithms. every deep learning algorithm uses an artificial neural network, which is also known as neural nets. The artificial neural network consists of artificial neuron, input, and output nodes. An artificial neuron node simulates the our brain’s neurons, which are given inputs and produce outputs. An example of neural network is shown in figure 1, where the hidden layer represents the artificial neural networks. An input node is connected with multiple neuron nodes, and a neuron node is connected with multiple output nodes because neuron nodes rely on multiple sources of inputs and output nodes consider the judgement of multiple neuron nodes. This way, the neural network can accurately predict data if given the inputs. To create an artificial neural network, a developer first programs to create the structure for the neural nets. Then, with a large amount of correct inputs and outputs, the developer trains the neural network by using the algorithms backpropagation. Given inputs, if the outputs by the artificial neural network do not match with the correct outputs, the backprogation algorithm tweaks the calculation of artificial neurons, so the outputs from the neural network match with the correct outputs. After the backpropation algorithm runs for thousands of times, the neural network can predict outputs accurately.
There is a lot of potential for deep learning. Currently, researchers are using deep learning to build self-driving cars. Others are trying to combine deep learning with robotics. Also, with deep learning, computers have the potential to recognize diseases better than a doctor can. However, currently, a major problem with deep learning is that a large amount of data is needed to train the neural network.