Salesforce AI Research Proposes 'DeepTime,' A Deep Time-Index Based Model Trained Via A Meta-Learning Formulation To Automatically Learn A Representation Function From Time-Series Data

Abstraction: Deep time-index meta-learning model for non-stationary time-series forecasting

Key points:

• DeepTime extends traditional time-index models by substituting the hand-specified function with a deep neural network, allowing the model to learn the time-to-value mapping from data
• Addresses non-stationarity via a locally stationary distribution assumption: a long non-stationary series is split into shorter windows (tasks) treated as stationary
• Uses a meta-learning formulation with inner/outer loops; only the final ridge regressor layer is adapted in the inner loop, making training fast and avoiding overfitting
• Ridge regression enables an exact one-step closed-form solution rather than an iterative approximation, improving prediction accuracy
• Achieves state-of-the-art on 20 of 24 benchmark settings (MSE metric) across six real-world datasets spanning diverse domains and forecast horizons
• Outperforms all benchmarks in memory and runtime efficiency, potentially reducing carbon footprint for forecasting deployments

Connections: Salesforce · Time Series Forecasting · Meta Learning · Deep Learning

Source: https://www.marktechpost.com/2022/10/17/salesforce-ai-research-proposes-deeptime-a-deep-time-index-based-model-trained-via-a-meta-learning-formulation-to-automatically-learn-a-representation-function-from-time-series-data/