CAREER: Sparse Associative Deep Learning using Neural Mimicry in Multimodal Machine Learning

点击翻译按钮获取中文摘要

Artificial intelligence has made incredible progress in the past several years. AI technology is now successfully being used in voice assistants, photo recognition technology, chatbots, search engines, and self-driving cars. While current AI is very good at matching specific patterns for specific tasks, research has shown that it cannot generalize to different tasks and has no real understanding of what it is doing. Thus, radical new directions need to be explored to achieve a truly intelligent machine. This project explores a new kind of AI framework, one that mimics how the human brain senses and understands the world. This new AI system learns much like an infant would, by simply observing the world and learning through exploration. This project also utilizes a new type of computer chip that communicates information in the same way that neurons in the brain communicate. Ultimately, the project will create a new kind of AI by mimicking certain functions of the human brain. This research can inform new methods and approaches to creating an AI that better understands the world in which we live. Furthermore, the project attracts and supports the education of students interested in the interdisciplinary field of human and machine intelligence.This project develops a new multimodal machine learning paradigm that is principally different from the traditional deep learning methods used in the state-of-the-art today. This research is inspired by breakthroughs in computational and theoretical neuroscience that incorporate ideas not explored by current feed-forward deep learning architectures. Rather than using massive labeled datasets, the algorithms learn much like an infant learns, i.e., by unsupervised observation and exploration of the world through different sensory inputs. The project addresses three primary research challenges: (1) the algorithms will robustly learn the structure of the world, (2) the model will learn heterogenous associations from repeated stimuli, and (3) given the same fundamental architecture, the model will learn how to predict the future. Furthermore, the framework described in the project mimics the hierarchical architecture, sparsity, top-down, and feedback functions of the mammalian brain. This model is built upon recent advances in neuromorphic software and hardware that enhance the functionality, energy use, and speed of the underlying algorithms. Given that neuromorphic approaches are under active development, this project has the unique opportunity to provide algorithms and functionality in software and silicon.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.