TERN: A Delta-rule Memory with a Seasonal Reference and Online Adaptation for Epidemic Forecasting
Shunya Nagashima, Yuta Funayama
cs.LG · cs.AI
Quick verdict
Delta-rule memory forecaster wins (uncontaminated) on three influenza benchmarks, but gains lean on seasonal/online add-ons and lack significance tests.
Summary
The paper proposes TERN, a forecaster for weekly regional influenza counts built around a linear-attention fast-weight memory updated by an erase-then-delta rule: channel-wise exponential decay with learned time constants (initialized over 1-20 week time scales), gated delta writes, and an erase address computed from local epidemic-phase features (growth, curvature, relative growth). An explicit seasonal reference (week-of-year embedding or two-season climatology with lead-time-dependent shrinkage) and online adaptation (refitting, blending with seasonal naive, Polyak averaging) complete the model. On the three Cola-GNN influenza benchmarks (Japan-Prefectures, US-Regions, US-States; leads 3/5/10/15 weeks), the window-regime TERN is best on Japan but behind reported EpiGNN on the US datasets, while the full-history regime attains RMSE 838/698/199, tying the seasonal naive on Japan, beating climatology on US-Regions, and matching reported EpiGNN on US-States. Ablations credit the delta-rule memory 5/3/22% RMSE over a softmax mixer and show the seasonal/online components dominate where used. Two benchmark findings stand out: the seasonal naive beats every published GNN at long leads yet validation cannot select it, and zero-shot Chronos-2/TimesFM-3 results on the US data are contaminated because their pretraining corpus contains the CDC FluView series. Limitations: single-disease scope, per-dataset component selection, and no statistical testing.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- TERN is, to the authors' knowledge, the first delta-rule linear-attention forecaster for epidemic surveillance series, differing from Kimi Delta Attention and Gated DeltaNet-2 by a single term.
- In the full-history regime TERN is the best uncontaminated model on all three Cola-GNN influenza benchmarks (RMSE 838/698/199), tying the seasonal naive on Japan and beating the two-season climatology on US-Regions.
- Under a controlled comparison holding context, seasonal reference and online adaptation fixed, the delta-rule memory contributes 5/3/22% RMSE reduction over softmax attention on the three datasets.
- The seasonal naive outperforms every published graph model at lead times of 10-15 weeks on Japan and US-Regions, but no validation-based selection procedure can choose it.
- Chronos-2 and TimesFM-3 zero-shot results on the US datasets are contaminated because the GiftEvalPretrain pretraining corpus contains the underlying CDC FluView series.
Strengths
Transparent contamination analysis of foundation models
Section 4 identifies that the GiftEvalPretrain corpus used by Chronos-2 and TimesFM-3 contains the CDC FluView ILINet/WHO-NREVSS series underlying US-Regions and US-States, flags those rows as contaminated rather than ranked, and supports the claim with evidence (TimesFM-3 RMSE on US-Regions is flat across leads at 318-355, unlike on Japan which is absent from the corpus).
Actionable benchmark finding about seasonal baselines
Finding (i) shows the 52-week seasonal naive beats every published GNN at h=10 and 15 on Japan and US-Regions, yet is unselectable by validation (Japan validation RMSE 1862 vs test 839) - a concrete diagnostic flaw in current epidemic-forecasting evaluation practice.
Controlled isolation of the memory contribution, with honest caveats
Precise lineage and fair comparator handling
Reproducibility package
Weaknesses
Per-dataset component selection without a stated procedure
No statistical testing or variance reporting
Mild overclaiming versus per-lead behavior
The memory alone does not win under the standard protocol
Narrow evaluation scope
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is TERN and what problem does it address?
TERN is a neural network for forecasting weekly regional influenza surveillance counts several weeks ahead. Its core is a linear-attention fast-weight memory updated by a delta rule with an explicit erase step whose gates read simple epidemic-phase features (growth rate, curvature, relative growth). The idea is to keep the shape of past seasons while discarding associations that become misleading once a wave peaks, and it adds an explicit seasonal reference plus online adaptation as new observations arrive.
How does TERN compare to a simple seasonal naive forecast?
On the three Cola-GNN influenza benchmarks, TERN's full-history mode ties the 52-week seasonal naive on Japan-Prefectures (RMSE 838 vs 839, with better correlation 0.926 vs 0.913), beats a two-season climatology on US-Regions (698 vs 727), and matches the best published graph model on US-States (199 vs 200). Its advantage concentrates at short lead times of 3-5 weeks; at longer leads it can be slightly worse than the seasonal naive, which the paper shows no published graph model beats at 10-15 weeks on two datasets.
Why are the Chronos-2 and TimesFM-3 results not ranked on the US datasets?
Because the GiftEvalPretrain corpus used to pretrain both foundation models contains the CDC FluView ILINet and WHO/NREVSS series that underlie the US-Regions and US-States benchmarks, their 'zero-shot' forecasts have effectively seen the test series during training. On the Japanese data, which is absent from that corpus, both models perform far less impressively, supporting the contamination explanation.
Is TERN's code and data available?
Yes. The paper states that code and configurations are publicly available on GitHub (github.com/Neurogica/TERN), and the evaluation uses the three public Cola-GNN influenza benchmark datasets (Japan-Prefectures, US-Regions, US-States) with a standard chronological 50/20/30 split.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 2 (forecast traces and learned time-constant histograms) is available only through its caption and in-text description, so the qualitative peak-timing claims cannot be independently checked.
- Table 1 reports averages over lead times; per-lead values are quoted only sporadically in the text, so claims such as the naive's dominance at h=10/15 could not be fully verified.
- Details of the refit stage and any supplementary material beyond the extracted main text were not available.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.