ADAPTive Input Training for Many-to-One Pre-Training on Time-Series Classification

2026-04-09 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

ADAPT (ADAPTive Input Training) is a novel pre-training paradigm designed for time-series data, addressing the challenge of generalizing self-supervised models across diverse datasets. Traditional one-to-many pre-training methods struggle when multiple datasets are introduced, hindering the development of foundation models for time-series. ADAPT overcomes this by efficiently aligning the physical properties of time-series data, enabling mixed-batch pre-training despite significant discrepancies in input sizes and channel dimensions. The model was trained on 162 time-series classification datasets, achieving new benchmark performance for classification. This capability to simultaneously train on a wide range of time-series datasets represents a crucial step towards building generalist foundation models in this domain.

Key takeaway

For research scientists developing foundation models for time-series data, ADAPT offers a critical solution to the generalization problem across diverse datasets. You should consider implementing ADAPT's approach to align physical data properties, allowing for more robust mixed-batch pre-training. This method can significantly enhance model performance and scalability, accelerating the development of truly generalist time-series AI.

Key insights

ADAPT enables mixed-batch pre-training for time-series data by aligning physical properties across diverse datasets.

Principles

• Align physical properties for diverse time-series data.
• Mixed-batch pre-training improves generalization.

Method

ADAPT efficiently aligns physical properties of time-series data to facilitate mixed-batch pre-training, accommodating varied input sizes and channel dimensions.

In practice

• Pre-train on 162 time-series datasets.
• Improve classification benchmarks.

Topics

• ADAPTive Input Training
• Time-Series Classification
• Self-Supervised Pre-training
• Foundation Models
• Many-to-One Pre-training

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Pre-Training Isn’t Bitter Enough — V-pretraining learns optimal self-supervised tasks using downstream feedback, making pre-training more adaptive and effective.
• Self-Adaptive Scale Handling for Forecasting Time Series with Scale Heterogeneity — Self-Adaptive Scale-handling (AS) improves time series forecasting by adaptively managing scale heterogeneity without losing signal or amplifying errors.
• Towards a Unified Generative Model for Scarce Time Series with Domain Experts — TimeMoDE unifies Diffusion Transformers and Mixture-of-Experts for robust time series generation in data-scarce settings.
• Bootstrapping-based Regularisation for Reducing Individual Prediction Instability in Clinical Risk Prediction Models — A novel regularisation method improves clinical prediction model stability without sacrificing interpretability, crucial for healthcare adoption.
• How are you handling training data when public datasets don't match your use case? [D] — Public datasets often fail to meet specific ML project needs, necessitating alternative data sourcing and generation strategies.
• Announcing Together AI and Adaption Partnership — Optimizing training data before fine-tuning open models significantly enhances model quality and development speed.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.