A Biologically Inspired Event Memory Architecture for Machine Learning
Abstract
Our proposal substitutes the discriminant cland a dictionary of templates that are incrementally updated with new examples, through interactions with the world employment reinforcement learning (RL). This dictionary is in fact a content addressable memory (CAM) that grows with the number of different templates that the vision system encounters when roaming the environment. The DPCN is a self-organizing system that creates proto-objects from the input. The goal of this proposal is to develop the WM and the CAM to enable learning the object features and their relevance for the task goals and subsequent organization, storage and improvement directly through experience and rewards from the environment. We will be using inspiration from biology in several key aspects of the architecture: the WM will be inspired by the cortical column organization to search the memory storage in the CAM; the organization of the memory will be by content and sequence; the incremental learning by interaction with the environment reminds us the way humans learn through the perception action reward cycle (PARC) from the flow of time events to be stored. We willdevelop event memories that are more biological plausible i.e. content addressable (CAMs) for static and temporal patterns and that can be searched by temporal order. There are no methodologies for addressing CAMs using sequence information, and we believe that for realistic machine learning applications involving time functionals (video or time series) the capability of implementing sequences of actions autonomously will free the executive module from interacting on a sample by sample basis with the world. Another aspect is the control of the size of the object CAM (OCAM) which will be necessary even when one uses a functional CAM in an Hilbert space. We will be creating an infrastructure of orthogonal clusters using the NICE framework that guarantees very small cross talk (virtually none), but this is deemed insufficient. In fact, we are seeking ways to mimic the aggregation of information that occurs in chunking, and also that occurs when we distill properties across many different instances of similar objects. Finally, the proposal also presents an approach to substitute the conventional classifier to recognize objects with a memory-based architecture that is inspired by Perceptual decision making using attention, working memory (WM) and Long-Term Memory (LTM). The central tenet of the present application researched by Co-PI Keil is that dedicated neurobiological subsystems for perception, attentive selection, working memory, action preparation/motivation, etc. are inseparably linkedamong each other, but also with systems dedicated to memory formation. We plan to create a ML memory-based architecture that mimics the function of WM and LTM, more specifically, we will create a WM as an internal, sparse and low dimensional representation of the world where a certain number of proto-objects selected by the DPCN will persist for an undetermined but transient time, depending upon the time it takes to find a reasonable solution for the task goal. A concurrent search motivated by the call trees implementation in the cellular automata created by the cortical columns will be implemented. Each object will have its own channel, i.e. it will access independently an object CAM (OCAM) to improve the quality of the object class. Moreover, we plan to extract dynamic relationships amongst objects in the WM to capture object affordances (i.e. how the objects are manipulated by the agent), which will also be internally stored in an action CAM (ACAM) and linked to the OCAM.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2021
- Source ID
- N000142112324
Entities
People
- José PrÃncipe
Organizations
- Office of Naval Research
- United States Navy
- University of Florida