Computational tools for understanding biological and artificial neural networks
- Benjamin Poole.
- [Stanford, California] : [Stanford University], 2018.
- Copyright notice
- Physical description
- 1 online resource.
Also available at
At the library
All items must be viewed on site
Request items at least 2 days before you visit to allow retrieval from off-site storage. You can request at most 5 items per day.
|3781 2018 P||In-library use|
- Neural networks are a core component of both biological and artificial intelligence. Despite advances in training artificial networks to solve complex tasks, they are often brittle in ways that biological networks are not, and our understanding of how they work remains incomplete. In this thesis, I take steps towards applying machine learning to better understand biological neural networks, and leverage ideas from biology to improve artificial neural networks. First, I present a new set of computational tools for processing, analyzing and understanding large-scale datasets in neuroscience. Applying these tools to calcium imaging from direction-selective neurons in the fruit fly visual system, I identify the algorithm the fly uses for computing motion, resolving a nearly 60 year old debate in fly visual neuroscience. In the second part of this thesis, I draw inspiration from neuroscience to advance the capabilities of artificial neural networks. By augmenting neural networks with higher-dimensional synapses, I drastically improve the ability of neural networks to acquire new information without forgetting old information when learning sequences of tasks. I also demonstrate how Poisson-like noise often found in biological systems can be incorporated into artificial neural networks to improve their sparsity, robustness, and ability to generalize. Finally, I demonstrate the reason neural networks are so expressive by analyzing signal propagation in random neural networks using tools from Riemannian geometry and mean field theory. Increasing the depth of even random neural networks exponentially increases their expressivity allowing them to disentangle highly complex manifolds in input space. The results presented in this thesis highlight the increasingly strong relationship between machine learning and neuroscience, both in terms of tools for understanding neural systems as well as inspiration for the capabilities and mechanisms needed for building artificial intelligence.
- Publication date
- Copyright date
- Submitted to the Department of Computer Science.
- Thesis Ph.D. Stanford University 2018.
Browse related items
Start at call number: