Artificial Neuroscientist: A Web Application for Visually Examining and Manipulating Neural Networks
CitationBryk, William. 2019. Artificial Neuroscientist: A Web Application for Visually Examining and Manipulating Neural Networks. Bachelor's thesis, Harvard College.
AbstractArtificial neural networks have shown remarkable success in recent years, shattering benchmarks on a diverse set of important tasks, and are now widely used in everyday technologies. However, the sheer size and nonlinear structure of neural networks make it difficult to understand their decision-making processes. Even as neural network development rapidly progresses, researchers have yet to fully explain how they work and often use trial and error to optimize their performance. It is especially difficult for students and those from other fields new to neural networks to participate in the neural network development process: to build and train them properly, and to optimize their many hyperparameters. In a similar vein, neuroscientists and psychologists have struggled to understand biological neural networks but have successfully developed techniques to better visualize and even intervene on their target systems' inner workings. Inspired by these fields, and in an attempt to bridge the gaps between the accessibility, interpretability and performance of artificial neural networks, I present a web application called “Artificial Neuroscientist Application”, or ANA. ANA enables users to build and train deep neural networks in the browser within minutes, visualize them in a dynamic 3D interface, and visually manipulate them by applying a variety of methods to probe their mechanics and functional idiosyncrasies. ANA does this with an intuitive plug-and-play design that requires no programming, lowering the barrier of entry for those eager to explore neural networks, but who lack the significant programming experience required to create them. ANA builds on previous work in neural network visualization, interactivity, and interpretability, combining some of the best elements of other tools into a new type of application designed to further demystify and democratize these algorithms.
Citable link to this pagehttps://nrs.harvard.edu/URN-3:HUL.INSTREPOS:37364595
- FAS Theses and Dissertations