A Framework for Universal Generalization via Memory Based Computation
Abstract
The recipe for the recent successes in data science and decision making applications has been simple: collect a large dataset to train a deep neural network. While this methodology has led to impressive progress, its fundamental limits are not surprising from a theoretical standpoint: most machine learning systems are designed assuming that train and test distribution are either identical or that the test distribution is known. Going beyond these assumptions requires re-imagining how we manage data for decision making. Our core insight is that problems such as poor generalization and catastrophic forgetting are intimately tied with the practice of prematurely condensing agentÕs experience into summary statistics (or features). An alternative is to avoid compressing data ahead of time, and instead combine data on-the-fly for the environment or the task encountered by the agent. In this vein, a particularly powerful and important form of data selection comes from the use of memory. By retrieving only relevant examples from an episodic buffer, selective memory retrieval can improve generalizationÑat least in principle. In practice, the optimal use of memory is an unsolved problem. We believe that inspiration from biological memory systems may hold the key to unlocking the full potential of selective memory retrieval in machine learning systems. In this proposal, we will first articulate a set of high-level computational principles for the design of memory systems (both artificial and biological), with the computational principles coming from knowledge about how the brain encodes and retrieves information in memory. We will then describe how these principles, common to biological and artificial systems, can be leveraged to tackle a range of challenging machine learning tasks, that are elementary to people given our memory representations. Finally, we will show how these same principles can be leveraged to understand how biological memory systems represent and retrieve naturalistic inputs. This work is critical to the research thrust on Neuro-Inspired Distributed Deep Learning (NIDDL). Our methods will help learn to transform data into understanding, enable autonomous agents to be persistently aware of their operating environment and optimize their operation appropriately, and be a major step toward the integration of artificial intelligence into a wide variety of real world systems. The end result will be scientific understanding of workings of biological memory, why is it organized the way it is, its role in decision making and practical algorithms for generalization to new tasks, lifelong learning without catastrophic forgetting and transfer across sensory modalities.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 28, 2023
- Source ID
- W911NF2310277
Entities
People
- Priyanka Agrawal
Organizations
- Army Contracting Command
- Massachusetts Institute of Technology
- United States Army