Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SNN-Based Online Learning of Concepts and Action Laws in an Open World

Christel Grimaud, Dominique Longin, Andreas Herzig

TL;DR

This work tackles open-world continual learning of concepts and action laws for autonomous agents by embedding a semantic memory in a spiking neural network. It introduces two concept types—object concepts and action concepts—learned online through STDP-inspired rules, boosted recruitment of rare features, and selective forgetting to preserve general knowledge while adapting to change. The architecture combines interface, object-concept neurons (O-neurons), and action-concept neurons (A-neurons) with a three-compartment learning scheme to predict action outcomes and guide decisions, demonstrated through a sequence of open-world experiments including room transfers and environmental changes. The results show rapid generalization to new situations, robust retention of general rules, and fast adaptation to new or relocated features, highlighting the potential of SNN-based semantic memory for real-time, open-world autonomy and planning extensions in the future.

Abstract

We present the architecture of a fully autonomous, bio-inspired cognitive agent built around a spiking neural network (SNN) implementing the agent's semantic memory. This agent explores its universe and learns concepts of objects/situations and of its own actions in a one-shot manner. While object/situation concepts are unary, action concepts are triples made up of an initial situation, a motor activity, and an outcome. They embody the agent's knowledge of its universe's action laws. Both kinds of concepts have different degrees of generality. To make decisions the agent queries its semantic memory for the expected outcomes of envisaged actions and chooses the action to take on the basis of these predictions. Our experiments show that the agent handles new situations by appealing to previously learned general concepts and rapidly modifies its concepts to adapt to environment changes.

SNN-Based Online Learning of Concepts and Action Laws in an Open World

TL;DR

This work tackles open-world continual learning of concepts and action laws for autonomous agents by embedding a semantic memory in a spiking neural network. It introduces two concept types—object concepts and action concepts—learned online through STDP-inspired rules, boosted recruitment of rare features, and selective forgetting to preserve general knowledge while adapting to change. The architecture combines interface, object-concept neurons (O-neurons), and action-concept neurons (A-neurons) with a three-compartment learning scheme to predict action outcomes and guide decisions, demonstrated through a sequence of open-world experiments including room transfers and environmental changes. The results show rapid generalization to new situations, robust retention of general rules, and fast adaptation to new or relocated features, highlighting the potential of SNN-based semantic memory for real-time, open-world autonomy and planning extensions in the future.

Abstract

We present the architecture of a fully autonomous, bio-inspired cognitive agent built around a spiking neural network (SNN) implementing the agent's semantic memory. This agent explores its universe and learns concepts of objects/situations and of its own actions in a one-shot manner. While object/situation concepts are unary, action concepts are triples made up of an initial situation, a motor activity, and an outcome. They embody the agent's knowledge of its universe's action laws. Both kinds of concepts have different degrees of generality. To make decisions the agent queries its semantic memory for the expected outcomes of envisaged actions and chooses the action to take on the basis of these predictions. Our experiments show that the agent handles new situations by appealing to previously learned general concepts and rapidly modifies its concepts to adapt to environment changes.

Paper Structure

This paper contains 34 sections, 3 equations, 5 figures, 5 tables, 27 algorithms.

Table of Contents

  1. Introduction
  2. Related works
  3. The agent and its universe
  4. The universe
  5. The agent
  6. Concepts and action laws
  7. Implementing the agent's semantic memory in the neural network: overview
  8. Spiking Neural Networks
  9. The neural network's architecture
  10. The neural network's functioning
  11. General notions
  12. Connections.
  13. Spike thresholds.
  14. Last spikes.
  15. Synapses' Growth Rates.
  16. ...and 19 more sections

Figures (5)

  • Figure 1: The agent's accessible world. A: in the first phase, Room 1 only; B: after opening the door, Rooms 1 and 2. C : after the sound feature is moved downwards (from row 3 to row -1) in Room 2.
  • Figure 2: Schema of the agent
  • Figure 3: Schema of the SNN. O-neurons #1 and #2 respectively encode the concepts [#0, Cold, SouthWall, WestWall, OK] and [#5, SouthWall, KO]. A-neuron #1 encodes the concept [[#0, Cold, SouthWall, WestWall, OK], [E, Orth], [#5, SouthWall, KO]]. A-neuron #3 encodes the concept [[OK], [Diag], [OK]]. For graphical readability, connections weights are omitted.
  • Figure 4: Actions' mean outcomes with door closed; Green: OK, Red: KO, Blue: Failure, Grey: no data.
  • Figure 5: Prediction of walls in arrival boxes given Sound as information about the depart location. Black dotted line: door opens; Red dotted line: sound changes location; Blue: NorthWall predictions considering a north move; Orange: SouthWall predictions considering a south move (percentages over 50 trials).