A Novel Predictive-Coding-Inspired Variational RNN Model for Online Prediction and Recognition.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Neuroscience

Background:

• Variational Bayes Recurrent Neural Networks (RNNs) face challenges in learning meaningful latent representations and transferring future observations.
• Existing models struggle to dynamically adapt stochasticity in latent states for complex temporal patterns.

Purpose of the Study:

• Introduce PV-RNN, a novel variational RNN inspired by predictive coding, to address limitations in current models.
• Enable latent variables to learn meaningful representations and improve inference model's ability to transfer future observations.
• Investigate the impact of a 'meta-prior' parameter on model behavior and performance.

Main Methods:

• Developed PV-RNN with adaptive vectors and prediction error-driven backpropagation for information conveyance.
• Introduced error regression for predicting unseen sequences, inspired by predictive coding mechanisms.
• Optimized the model by maximizing a lower bound on marginal likelihood, balancing prediction errors and KL divergence using a meta-prior.

Main Results:

• PV-RNN dynamically adjusts latent state stochasticity to extract probabilistic structures from temporal data.
• High meta-prior values led to imitation of data randomness via deterministic chaos; low values resulted in random process behavior.
• Intermediate meta-prior values yielded optimal performance, capturing latent structures with good generalization.

Conclusions:

• PV-RNN demonstrates effective capture of latent probabilistic structures in sequential data.
• The meta-prior parameter allows for controlled exploration of stochastic dynamics, offering theoretical insights and practical benefits.
• PV-RNN with error regression achieved superior prediction performance on a robot imitation task compared to standard variational Bayes models.