Artificial General Intelligence: Efficient Reinforcement Learning to Achieve Human-Like Reasoning

Overview

Objectives/Goals Artificial neural networks suffer from a phenomenon known as "catastrophic forgetting", which prevents artificial intelligence agents from learning multiple tasks. This project explores various network structures and training methods to deal with this barrier to achieving general intelligence. Methods/Materials Used Python programming language with Theano and Keras libraries. Various unconventional artificial neural network structures were assessed on tasks such as object recognition and cursor navigation based on training speed and resilience. Results The traditional sequential models demonstrated catastrophic forgetting to a large extent, achieving over 90% accuracy on one task but under 10% on the other. Meanwhile, a sequential model trained simultaneously on both tasks reached only 50-60% accuracy. However, progressive or 2-column neural networks improved about 20% on this benchmark. There was a significant increase in performance and training efficiency. Conclusions/Discussion Progressive neural networks are an efficient way to improve AI performance on multiple tasks. They can be easily extended to encompass more than two tasks, through the use of multiple columns, or subnetworks. Choosing the output (since PNNs create multiple outputs) is a trivial task that may be left to a separate machine learning model. With learning models such as the PNN, the issue of catastrophic forgetting is no longer a major obstacle in achieving general intelligence.

Competition history

• CSEF 2017 Mathematical Sciences · Entry S1511

Resources

• CSEF project summary ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Mathematics projects
• CSEF Mathematics projects
• Every CSEF 2017 project