What
This is an article that summarizes what I noticed and researched when learning machine learning using Chainer. This time, we will study the basics of neural networks.
It is written based on my understanding, so it may be incorrect. I will correct any mistakes, please forgive me.
Content
A neural network refers to a computational graph created by connecting differentiable transformations. Here are some other terms you should know about the set. Node: Input or (final) output The calculation result in the middle is also a node (it seems) Edge: A line connecting nodes. Just make the state transition easier to see? It seems that machine learning using a neural network with a large number of nodes is called deep learning. (Finally came out)
Layer Input layer-> intermediate layer or hidden layer (calculation process)-> output layer.
The number of nodes in the middle layer is arbitrary. The sense of the designer is questioned. Parameters that you decide are called hyperparameters. There are several types of neural networks. (Full join, convolution, recursive type, etc.)
When an input is given, each layer of the neural network is calculated in order, and the calculation to the output is called forward propagation.
Linear transformation is possible when the input / output relational expression can be expressed using a matrix.
If you want to handle non-linear input / output, treat the neural network as non-linear by inserting one process of non-linear conversion of the linear conversion result. The function that linearizes → non-linearizes is called the activation function. There seem to be various activation functions.
So far this time, I will continue next time
Comment I often distinguish between experimental and theoretical systems, but I think it is absolutely necessary to have someone who can understand both in the not too distant future. Machine learning When I said that I would study and master it independently, I received support. Let's do our best
Recommended Posts