Artificial Neural Networks
In the past, no one could think of paintings drawn by computers. They only existed in Sci-Fi movies or in imagination. However, recent development in Artificial Intelligence (AI) have proved that computers can draw paintings: not simple paintings just composed of dots and lines, but beautiful and complex drawings like the below image:
Figure 1. Image from Inceptionism Gallery [1]
The image was created by Google’s technology called “Inceptionism” [1]. The Inceptionism technique created the image by recognizing features from an inputted picture and emphasizing them to the picture.
To understand the “Inceptionism,” it is necessary to understand Artificial Neural Network [2]. Artificial Neural Network is one of the machine learning algorithms, and it shows superior performance in classification task – e.g. whether a cat is in a picture or not. Thanks to its superior performance, it is widely used like in Google’s speech recognition system and web searching system. The neural network is composed of many layers, which contain many “nodes” and “edges.” One structural example of the network is described below:
Figure 2. Neural Network Structure [2]
Through feeding many images (and/or labelled with ground truth), Artificial Neural Network “learns” to modify weights between nodes. For instance, if node A is good for recognizing a cat ear and input image is a cat, then the network classify the input image as a cat through putting strong the weight from node A to the output layer. Therefore, the learned network has different weights between nodes – strong or weak. This concept is very similar to what we learned in class. Friendship networks, for instance, has strong ties or week ties between friends, and the network tells us much information like which friend groups that a person is close with. Similar to the Artificial Neural Network, the strong and week ties between nodes are effectively used to classify images.
Nowadays, Artificial Neural Network has advanced further and it is composed of many and many layers, such as GoogLeNet [3]. It is interesting to see how further these network will develop and bring innovations in our life.
Related Articles and Paper
[1] Link to Inceptionism: http://googleresearch.blogspot.com/2015/06/inceptionism-going-deeper-into-neural.html
[2] Link to a book “Neural Networks and Deep Learning”: http://neuralnetworksanddeeplearning.com/index.html
[3] Link to GoogLeNet paper: http://arxiv.org/abs/1409.4842