Paano managinip ang computer? Scientists show how Google ‘dreams’

 

 

Artificial neural networks (ANNs) are computer systems designed to mimic the functions of biological brains. Google employs ANNs to help with image searches and speech recognition, for example.

Recently, scientists at Google were asked to see how accurate the company's ANNs were in understanding the visual data fed into them.
 

Surprisingly, the ANNs were found to be capable of “dreaming”—coming up with detailed images even after being presented with a random noise image.

 

According to an article written by Google software engineers Alexander Mordvintsev and Mike Tyka, and software engineer intern Christopher Olah, an ANN is trained by showing it millions pictures of the target object. The 10-30 stacked layers of the ANN are made up of artificial neurons, which each layer getting progressively more and more complex when it comes to analysis.

Going bananas
 

If you want your ANN to be able to identify a banana, for example, you show it pictures of different kinds of bananas. Each layer analyzes the images, with the first layer maybe looking for edges, the next looking at the color, and so on. The final layer decides what constitutes a banana. If the ANN makes a mistake, the team adjusts its parameters.

 

What’s interesting is that the team found that that ANNs could generate images aside from analyzing them. According to them, this presents an opportunity to check if the ANN “correctly learned the right features” that make up the object in question.

So to find out if the network knows what a banana is, ask it to generate a picture of a banana.

Levels of understanding
 

The team fed the ANN an image or photo and asked it to enhance whatever it found. The complexity of the generated images depended on which level of the ANN the team chose to use. “For example, lower layers tend to produce strokes or simple ornament-like patterns, because those layers are sensitive to basic features such as edges and their orientations,” they said.

 

One network thought that dumbbells had arms, probably because majority of the images of dumbbells also had someone lifting it.

 

“This technique gives us a qualitative sense of the level of abstraction that a particular layer has achieved in its understanding of images,” the team said, calling the technique Inceptionism. Processed images and their results can be found in Google’s Inceptionism gallery.

Click here to visit Google's Inceptionism gallery.

 

Computer dreams

The team then decided to repeatedly apply the algorithm on the network’s own outputs, zooming in a little bit after each iteration. They described the results as “an endless stream of new impressions, exploring the set of things the network knows about.” Using a random-noise image as a base means that the result is something that comes solely from the neural network. The team called the products the “dreams” of the ANN.

 

“The techniques presented here help us understand and visualize how neural networks are able to carry out difficult classification tasks, improve network architecture, and check what the network has learned during training. It also makes us wonder whether neural networks could become a tool for artists—a new way to remix visual concepts—or perhaps even shed a little light on the roots of the creative process in general,” they said. — Bea Montenegro/TJD, GMA News