Developing next-generation learning algorithms that draw inspiration from biological nervous systems to revolutionize intelligent systems
Researchers in APL’s Intelligent Systems Center (ISC) are extracting and translating design principles of neural connectivity and function from multiple species to create the next generation of robust, efficient intelligent systems operating in the real world.
Connectomes—snapshots of brain connectivity, extrapolated from detailed maps of neural connections in biological brains—can inform novel artificial recurrent neural network architectures, as demonstrated by the ISC’s work modeling the heading direction network of the fruit fly. ISC researchers have successfully tested fruit-fly-inspired algorithms on robotic platforms.
The ISC plays a key role hosting large-scale neuroscience datasets for the community, focusing in particular on nanoscale connectomes. Teams from the ISC also develop open-source analysis tools for extracting patterns from these datasets, such as repeated structural patterns, or motifs. In addition to extracting insight from the neuron-synapse connectome of fruit flies, ISC researchers extrapolate insights about neural connectivity from a range of species—including humans—and scale these insights to improve network robustness and performance.
Brian Robinson discusses “Informing generative replay for continual learning with long-term memory formation in the fruit fly.”
Building on the ISC’s expertise in sleep research, researchers at APL are translating the benefits of biological sleep—ubiquitous across intelligent organisms—to artificial neural networks. This has led to improvements to approaches to overcome catastrophic forgetting, which is critical to enabling improved continual learning in classification networks. APL has also developed an approach for efficient generative replay for continual learning based on fruit-fly memory consolidation, and is looking into replay mechanisms for acquisition of semantic knowledge through associative learning. Additionally, APL has built complex, robust, and adaptive swarming robotic controllers based on neural dynamics.
