Investigating the Generality of Deep Learning
Master thesis
Permanent lenke
http://hdl.handle.net/11250/2408428Utgivelsesdato
2016Metadata
Vis full innførselSamlinger
Sammendrag
This thesis investigates how general the knowledge stored in deep-Q-networks are. This general knowledge can be used to reduce the training time of deep neural networks. Recent advances in the field of deep reinforcement learning have yielded more general solutions than previously possible. Deep architectures are computationally expensive to train, and general knowledge can be used to kick-start the training, effectively reducing training time.
We know that the low-level features in deep convolutional neural networks trained on image recognition tasks tend to be of a somewhat general nature. To investigate if this is the case for deep reinforcement learning, deep-Q-networks were implemented and trained on two similar Atari 2600 games; Pong and Breakout. First, the low-level features between two networks were visually compared. Second, the differences between the low-level features were quantified. Third, the first convolutional layer of a fully trained base network was transferred to a target network before training. This could determine if the general features in the base network would give a cutback in training time for the target network.
The results were mixed. Visually, there were few similarities between the two tasks, and many filters resembled task-specific features. Nevertheless, the quantified difference showed that there were indeed similarities. Using Breakout as base network and Pong as target network resulted in faster convergence and a possible cutback in training time. However, using Pong as base task and Breakout as target task did not. This may be due to the variation in difficulty between the two tasks.