Yufan Zhang Rose Hills
Analyzing Neural Networks in the Context of Concept Learning Tasks
Neural network models have traditionally been viewed as a black box, with tremendous capabilities in a variety of domains, yet with inexplicable inner workings. Past attempts at analyzing neural networks include analyzing model results and learned weights in an effort to design explainable artificial intelligence, as well as early efforts to determine the full capabilities of neural network models. My project seeks to continue in this tradition by investigating the logic-building aspect of neural network models. More specifically, I aim to investigate whether a novel program synthesis neural network model builds internal logical structure during the course of a simple rule-learning task, and whether the model’s logic-building process shares similarities to humans, demonstrated in biases such as a preference for simplicity or brevity.