Since the epic battle amid Google AlphaGo and legendary 9 dan Go master Lee Sedol, deep learning has become a word that popularized around the globe. The historic match also raised a question that has been discussed for decades, “Will artificial intelligence eventually replace humans?”
As always, AlphaGo is just a recent example that we are all familiar with. We will use AlphaGo as an example, and let you have a basic understanding of neural networks by reading this article. For neural network gurus, you may read our upcoming article about future prediction of AI’s potential.
As we have all heard of, AlphaGo is a program that leverages artificial neural network to make decisions. Neural network is in fact morphing the structure of a human brain, which consists of a fundamental building block: neurons. Neurons have their own weight and they will deliver bias. Weight is somewhat like a function, which causes neuron’s output differs in accordance to the input value. As inter-connected neurons are aligned into layers, the initial output of the 1st neuron becomes the input of 2nd layer of neurons. Undergoing the process of “weighing”, the neurons deliver output value known as “bias”. After penetrating through all neural layers, the final output makes up the decision and result of the network.
Therefore, as you may observe, the onus is on the initial output. If we are not obtaining an accurate input, no matter how meticulously the layers are designed, the output is always considered useless. Meanwhile, layers in between the first and last neuron, known as hidden layers, are carrying an essential role too. If weights are not precise, the result will be disastrous as we can all predict.
For the first problem, the accuracy of data of the initial neuron or the initial layer, greatly depends on the method and technology for obtaining raw data. AlphaGo is relatively simple for this process, as the chess are placed at limited boxes at fixed position. Yet, other neural network’s application, such as analyzing stock market by massive reading of analyst reports and financial news, might be much more troublesome. As programs have to apply the skill of natural language processing, a branch of neural network, so as to understand those metaphors used by the authors and refine useful information, the workload exponentially increases.
Therefore, the solution for the first problem is by utilizing advanced camera lens or sensors to perceive the surroundings. Google’s driverless car make use of LIDAR, a self-rotating radar that transmits ad receives pulsed lasers to have a clear understanding of the road condition. This clearly makes Google cars safer as they are able to detect the environment and react to the potential dangers at a rapid speed.
The second problem, which is the inaccurate weights and therefore causing inaccurate bias, can be resolved using a neural network training method called backward propagation. By training only two layers at a time, programmers are then able to acknowledge the malfunctioned neurons, so as to fix the problem by finely adjusting the weights. By having the generated output reversely imported into the second layer, if the two layers are able to create an output that is exactly the same as or very close to the initial output, it implies that these two layers are well-trained. All layers are theoretically able to attain the highest degree of accuracy, it is just the matter of money. Since outrageous number of calculations involved during the training process at every second, GPUs are often adopted to speed up working progress, as they are able to parallel process data, without the need of waiting for the prior result.
Will AI replace human in the long term? We will reveal it at in our next article.