Wake-Sleep Energy Based Models for Continual Learning
Bibliographic record
Abstract
This paper introduces a novel approach for continually training Energy-Based Models (EBMs) on the classification problems in the challenging setting of class incremental learning. Despite the fact that EBMs offer longer retention of knowledge on prior tasks, training EBMs contrastively remains a challenge. Driven by biological plausibility, we leverage the observation that sleep in humans supports active system consolidation and propose a new approach for training EBMs, which we call Wake-Sleep Energy Based Models (WS-EBMs), which rely on wake-sleep cycles. Our training approach consists of short wake phases followed by long sleep phases. During the short wake phase, the free energy associated with ground truth labels is minimized, which conditions the model towards the correct solutions. This is followed by a long sleep phase, where the free energy of the whole system is minimized contrastively, which allows the model to push the energy of incorrect solutions further from the correct response. We provide a theoretical analysis of WS-EBM showing that it satisfies the sufficient condition for designing proper EBM loss. Our empirical evaluation confirms the plausibility of our approach and demonstrates favorable performance of WS-EBM compared to traditional EBM training as well as state-of-the-art class-incremental continual learning techniques. Furthermore, our proposed two-phase training strategy can be easily integrated with existing techniques resulting in substantial boosts in their performance. Finally, we also provide interesting insights justifying our approach by analyzing the orthogonality between the sequential task vectors, and flatness of the optimized energy surfaces, which may guide the design of class incremental continual learning strategies.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.005 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.003 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.003 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".