Related Experiment Video
Updated: Aug 14, 2025
03:14
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
659
From eye-blinks to state construction: Diagnostic benchmarks for online representation learning.
Banafsheh Rafiee1, Zaheer Abbas2, Sina Ghiassian1
1Department of Computing Science and the Alberta Machine Intelligence Institute (Amii), University of Alberta, Edmonton, AB, Canada.
Adaptive Behavior
|January 9, 2023
Summary
We introduce new problems for online prediction learning, inspired by animal conditioning, to improve recurrent neural network capabilities. These challenges facilitate research into scalable representation learning for continual learning agents.
Area of Science:
- Artificial Intelligence
- Computational Neuroscience
- Machine Learning
Background:
- Classical conditioning demonstrates animals' ability to form long temporal associations for multi-step prediction.
- Recurrent neural networks (RNNs) can learn temporal associations by constructing internal state representations.
- Current RNN training methods are computationally expensive for online, continual learning scenarios.
Purpose of the Study:
- To propose novel diagnostic prediction problems for evaluating online prediction learning.
- To facilitate research on enabling agents to replicate animal-like temporal association abilities.
- To highlight limitations of current recurrent learning methods in continual learning settings.
Main Methods:
- Development of three new diagnostic prediction problems based on classical conditioning paradigms.
- Utilizing recurrent neural networks as the agent architecture for learning temporal associations.
- Focusing on the challenges of online prediction (continual learning at each time step).
Main Results:
- The proposed problems effectively test the learning capabilities inspired by animal conditioning.
- The problems reveal limitations in current recurrent learning methods for online prediction.
- The problems are designed to be non-trivial yet amenable to analysis in a small-compute regime.
Conclusions:
- The new problems accelerate research towards scalable online representation learning methods.
- Further research using these problems can isolate and address specific issues in recurrent learning.
- This work aims to bridge the gap between biological learning and artificial agent capabilities.
