Research map / Label-efficient and robust learning / Test-time adaptation
Training for the Future: A Simple Gradient Interpolation Loss to Generalize Along Time
Anshul Nasery, Soumyadeep Thakur, Vihari Piratla, Abir De, Sunita Sarawagi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GI instead of DPerturb | Temporal generalization models · 2-Moons | misclassification error (%): 5.6 ± 2.6 -> 3.5 ± 1.4 | Improved |
| Use GI instead of Generative Model | Temporal generalization models · Rot-MNIST | misclassification error (%): 13.5 ± 3.2 -> 7.7 ± 1.3 | Improved |
| Use GI instead of DPerturb | Temporal generalization models · M5-Hob | MAE: 0.28 ± 0.09 -> 0.09 ± 0.03 | Improved |
| Use Deep ODE instead of GI | Temporal generalization models · 2-Moons | misclassification error (%): 3.5 ± 1.4 -> 2.2 ± 0.8 | Improved |
| Use GI instead of FTPL | Online-learning comparison · Rot-MNIST | error: 11.5 ± 3.1 -> 7.7 ± 1.3 | Improved |
| Use GI instead of FTRL | Online-learning comparison · Elec2 | error: 20.7 ± 0.8 -> 16.9 ± 0.7 | Improved |
| Use GI instead of FTPL | Online-learning comparison · Shuttle | error: 0.75 ± 0.12 -> 0.21 ± 0.05 | Improved |
| Use GI instead of FTRL | Online-learning comparison · House | error: 10.3 ± 0.3 -> 9.6 ± 0.02 | Improved |
| Use GI instead of CIDA | Temporal generalization training · Rot-MNIST | training time (seconds): 257.3 -> 207.1 | Improved |
| Use GI instead of ERM | Temporal generalization training · 2-Moons | training time (seconds): 6.2 -> 7.1 | Hurt |
| Use one shared adversarial δ per minibatch rather than a separate δ per example (ablation) | GI fine-tuning · Reported across datasets | performance | Improved |
Closest papers by meaning
- TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.86)
- Timer: Generative Pre-trained Transformers Are Large Time Series Models (ICML 2024 · similarity 0.85)
- TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting (ICLR 2024 · similarity 0.85)
- Grokking as the transition from lazy to rich training dynamics (ICLR 2024 · similarity 0.85)
- Let Me Grok for You: Accelerating Grokking via Embedding Transfer from a Weaker Model (ICLR 2025 · similarity 0.85)
- Grokking in Linear Estimators -- A Solvable Model that Groks without Understanding (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
