Software Art Studio Assignment 10

Part 1: READ

1. A Beginner’s Guide to Neural Networks by Nehal Udyavar

Part 1: I think the training of perceptron is done by using an evolutionary algorithm. Part 2: The Sigmoid Function is putting the whole thing to a machine way, it defines whether "not sure" will become true or false by formula. But in human mind, although we are uncertain on something we should do or not, we are affected by hormone. We propose differently on the samething as we affected by memories and feeling.

2. Another video explaining neural networks:

Part 2: RESEARCH

StyleTransfer

Examples: Her Muse

Reference: TensorFlow Library.

My Understanding

So basically style transfer is the process of analysing a style from painting(s)(as they are more likely to have a specific, recognizable texture) and applying it to another image(sequence).
The use of neural network: it's been trained to build up a model for the style(analysis). From the ml5.js, it is a prebuilt model that allows the user to make use of it, which is very convinient as training these models consumes a lot of time and resources. The ready made model, I believe is based on VGG(Visual Geometic Group) model. What it does is inputting partial information of large areas at first. As the input data goes through iterations, the areas scale up and the number of nodes stacks up, finally, to the smaller area.
Process: For now, we have the model trained online. We can extract the style representation within the intermediate layers. The first few layers are not so "intelligent", they only represent low-level features like edge, textures, colour. These are still done by going pixel to pixel data. Then, in the deeper layer, the number of nodes increases, and they started to compare themselves to the network. They can come up with a higher level of content in style. The network has been trained to turn image pixels into features of style.
The content of the image is represented by the intermediate feature maps(processed nodes). The style of the image is represented by the means and correlations of different feature maps. A gram matrix is needed to be put in each location of the image by taking the outer products of the feature map and averaging the outer product over all locations. We are able to preview a specific feature map, but all it returns is just the specification of values, e.g. min, max, mean. After all, it is only a component to construct the style, there is no purpose for it to be readable.
The last part is how we apply the style to the output image. After the analysis, there will be mostly the high-frequency contents(e.g. grass, ripple). But we don't want our style only to apply to these areas. therefore, we need control over our weighting, specific term: total variation regularization. I will describe it as smoothing the amount between heavy weighted and light-weighted contents. Opposite to path-finding.
Development: While we use this library in machine learning, I think it is a traditional way to use the neural network. It is relative slow because it is a single process of generation, input-process-output, flow in a single direction. The neural network can be applied to the GAN system. The image transformation net as G and the VGG network as D. G is the faker to propose an image and D is the style extractor to validate and train G. That will continuously both of them develop continuously and return a faster system of machine learning.

Final Project Development

Website

This week, I have registered the domain for the final project and also my own website: JASONSART.INFO.

PHP

Also, I have created a database for the final project by using MySQL. I am still figuring out how to use PHPMyAdmin to manage tables. I have tested PHP from one of my websites to connect to MySQL.