Communications of the ACM
nveiling the Hidden Layers of Deep Learning
Envisioning how neural networks learn.
Credit: Daniel Smilkov and Shan Carter
A new method for visualizing the mechanisms and hidden layers of neural networks could provide insights into deep learning.
It is established that an order exists to how the hidden layers function, in that from input to output, each layer manages information of increasing complexity.
The Tensor Flow project's goal is to shed light on these layers by enabling users to interact and experiment with them via an open source tool described as a neural network playground.
The playground employs blue and orange points scattered within a field to "teach" computers to find and echo patterns. By choosing different dot-arrangements of varying degrees of complexity, users can manipulate the learning system by adding new hidden layers and new neurons within each layer. Afterward, every time users hit the "play" button, they can observe as the background color gradient changes to approximate the arrangement of blue and orange dots.
As the pattern's complexity grows, additional neurons and layers help the computer to complete the task more successfully. Connections among neurons are imaged as either blue or orange lines, with blue signaling the output for each neuron is the same as its content, while orange means the output is the opposite of each neuron's values.
From Scientific American
View Full Article
Abstracts Copyright © 2016 Information Inc., Bethesda, Maryland, USA
No entries found