Continual Sequence Modeling With Predictive Coding.

Area of Science:

• Robotics
• Machine Learning
• Artificial Intelligence

Background:

• Recurrent neural networks (RNNs) excel at sequential data but rely on batch training and backpropagation through time (BPTT).
• Lifelong learning in developmental robotics requires continuous adaptation in evolving environments.
• Existing methods often require storing past data, posing memory challenges.

Purpose of the Study:

• Investigate RNN designs and learning methods for continual learning in robotics.
• Evaluate algorithms with low memory requirements that do not store past information.
• Propose and validate a novel model for lifelong learning in embodied agents.

Main Methods:

• Evaluated various RNN architectures and learning algorithms in a continual learning setting.
• Utilized a generative modeling task involving sequential trajectory generation.
• Focused on methods with minimal memory footprint for parameter updates.

Main Results:

• Identified conceptors and predictive coding as effective approaches for this task.
• The proposed PC-Conceptors model, combining these two mechanisms, achieved superior performance.
• PC-Conceptors demonstrated lower average prediction error across 20 sequential trajectories.

Conclusions:

• PC-Conceptors offers a promising solution for lifelong learning in robotics.
• The model effectively handles sequential data with continuous adaptation.
• This approach advances the development of adaptive embodied AI systems.