A Novel Predictive-Coding-Inspired Variational RNN Model for Online Prediction and Recognition (1811.01339v3)

Abstract: This study introduces PV-RNN, a novel variational RNN inspired by the predictive-coding ideas. The model learns to extract the probabilistic structures hidden in fluctuating temporal patterns by dynamically changing the stochasticity of its latent states. Its architecture attempts to address two major concerns of variational Bayes RNNs: how can latent variables learn meaningful representations and how can the inference model transfer future observations to the latent variables. PV-RNN does both by introducing adaptive vectors mirroring the training data, whose values can then be adapted differently during evaluation. Moreover, prediction errors during backpropagation, rather than external inputs during the forward computation, are used to convey information to the network about the external data. For testing, we introduce error regression for predicting unseen sequences as inspired by predictive coding that leverages those mechanisms. The model introduces a weighting parameter, the meta-prior, to balance the optimization pressure placed on two terms of a lower bound on the marginal likelihood of the sequential data. We test the model on two datasets with probabilistic structures and show that with high values of the meta-prior the network develops deterministic chaos through which the data's randomness is imitated. For low values, the model behaves as a random process. The network performs best on intermediate values, and is able to capture the latent probabilistic structure with good generalization. Analyzing the meta-prior's impact on the network allows to precisely study the theoretical value and practical benefits of incorporating stochastic dynamics in our model. We demonstrate better prediction performance on a robot imitation task with our model using error regression compared to a standard variational Bayes model lacking such a procedure.

• The paper introduces a predictive-coding variational RNN that learns probabilistic temporal structures through a novel adaptive update mechanism.
• Adaptive error regression enables the model to dynamically update its internal state using prediction errors during testing.
• Empirical results show robust generalization in capturing stochastic sequences, outperforming conventional variational Bayes RNNs in real-world tasks.

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

The paper introduces a novel predictive-coding-inspired variational recurrent neural network (PV-RNN) designed to address challenges in learning the probabilistic structures in temporal sequences. This model leverages ideas from predictive coding to improve upon traditional variational Bayes RNNs, specifically targeting issues related to meaningful representation of latent variables and the incorporation of future observations into the model.

The PV-RNN introduces several key innovations. It utilizes adaptive vectors that dynamically change during evaluation to capture the underlying structure of the data. Unlike traditional methods that rely on external inputs, PV-RNN uses prediction errors during backpropagation as a mechanism to inform the network of external data, which aligns with predictive coding principles. Additionally, the model employs error regression, an online mechanism during testing that continually updates the internal state to enhance prediction accuracy.

The model's learning objective is to maximize a lower bound on the marginal likelihood of sequential data, composed of the expectation of prediction errors and the KL divergence between prior and approximate posterior distributions. Through empirical evaluations on datasets with probabilistic structures, the paper demonstrates that the PV-RNN can transition between deterministic chaos and random processes. The tunable parameter, meta-prior, balances the optimization pressure between these two dynamics. Optimal results are obtained at intermediate values of the meta-prior, striking a balance that facilitates generalization and the capture of latent probabilistic structures.

In simulations, the model exhibits robust generalization capabilities, accurately capturing stochastic sequences by differentiating between deterministic and probabilistic temporal patterns. Practical applications are showcased through a robot imitation task where PV-RNN outperforms conventional variational Bayes models in prediction accuracy, showcasing its effectiveness in real-world scenarios.

Theoretically, the work expands on the intersection of machine learning and computational neuroscience, exploring how neural architectures might leverage stochastic dynamics akin to probabilistic reasoning observed in biological systems. This research opens avenues for advancing AI systems in their mimicking of human-like prediction, recognition, and learning processes.

Future developments may explore the integration of goal-directed planning and the extension of PV-RNN to more complex, dynamic environments, potentially addressing challenges in cognitive robotics and adaptive systems. Moreover, further research might investigate the PV-RNN's scalability and efficiency, particularly in handling computational demands of error regression in real-time applications. Overall, the paper provides a significant step in enhancing RNN capabilities through the integration of predictive coding principles, with promising implications for both theoretical advancements and practical implementations in AI.