Accelerating Training with Neuron Interaction and Nowcasting Networks

TL;DR

NiNo tackles accelerating training by combining a periodic weight-nowcasting model with neural-graph representations of neurons. It extends prior WNNs by encoding neuron connectivity with graph neural networks to perform multi-step forecasting $\hat{\bm{\uptheta}}_{t+k}=\bm{\uptheta}_t+f_k^{\phi}(\cdot)$ while keeping the base optimizer (e.g., Adam) as the primary updater. Empirically, NiNo achieves speedups up to ~50% on vision and language benchmarks, generalizes to larger Transformer-style models, and introduces low-dimensional graph embeddings that can illuminate training dynamics. The work demonstrates the value of permutation-aware, structure-exploiting parameter prediction for scalable optimization.

Abstract

Neural network training can be accelerated when a learnable update rule is used in lieu of classic adaptive optimizers (e.g. Adam). However, learnable update rules can be costly and unstable to train and use. Recently, Jang et al. (2023) proposed a simpler approach to accelerate training based on weight nowcaster networks (WNNs). In their approach, Adam is used for most of the optimization steps and periodically, only every few steps, a WNN nowcasts (predicts near future) parameters. We improve WNNs by proposing neuron interaction and nowcasting (NiNo) networks. In contrast to WNNs, NiNo leverages neuron connectivity and graph neural networks to more accurately nowcast parameters. We further show that in some networks, such as Transformers, modeling neuron connectivity accurately is challenging. We address this and other limitations, which allows NiNo to accelerate Adam training by up to 50% in vision and language tasks.

Figures (2)

• Figure 1: Adam without and with nowcasting using our NiNo vs WNN jang2023learning on a language task that NiNo and WNN have not seen during their training.
• Figure :