NeurIPS 2023 Poster Session 2 (Wednesday Morning)

AI Behavioral Science Expert

Understanding the underlying dynamics of RNN training is critical for practical applications, especially as AI systems become more integrated into decision-making processes. The fluctuating nature of training losses illuminates potential behavioral patterns in automated processes. Improved techniques like generalized teacher forcing can be pivotal in establishing more reliable AI outputs.

AI Data Scientist Expert

This exploration into RNN training highlights the complex interplay of structure and performance in neural networks. The relationship between training dynamics and empirical data allows for the recalibration of models in real-time, thus optimizing their functionality in diverse environments. Such findings are invaluable as industries increasingly rely on sophisticated AI systems.

RNN (Recurrent Neural Network)

RNNs face challenges like loss divergence that affect their training dynamics.

Generalized Teacher Forcing

This method has been shown to reduce variance in loss jumps during model training.

Dynamical Systems

Analyzing dynamical systems provides insights into training behaviors across RNNs.