acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Gating Is Just Counting: Scalar-Mediated Energy-Identity Collapse in Federated Continual Learning under Concept Drift

Abstract

When adaptive inference systems (early exits, token pruning, gated adaptation) report energy savings, what should the savings be attributed to: the adaptive policy, or merely the operating point it induces? Under event-driven binary observation with a constant per-event cost (INA219 stage constants on Raspberry Pi 3B), we show that a controller's policy enters the energy identity only through a single scalar, the trigger count. By a Le Cam two-point argument the policy is unidentifiable from this channel, and gated adaptation stays within 2c_ad of the best count-matched open-loop schedule—on hardware (-0.27 J, Wilcoxon p = 7.1e-4; paired, n = 20). The collapse carries falsifiable content: across all 680 arrival orderings the feasible grid (V in 4, 8, 16, Theta = V.c_ad^harness >= 1, c_ad^harness = 0.2896 J) returns N_fire = floor[A/Theta] exactly; at V = 2 (Theta = 0.5792 < 1) 134/680 (19.7%) deviate at (A,T) = (3,17), diluting to 1.67% at the hardware window (A,T) = (5,300). The policy can matter; it cannot bypass one scalar. The claim is falsified precisely where Theta < 1. Given the event stream, the ledger is auditable without knowing the policy (against under-reporting and replay), and it locates where adaptation can still pay: state-dependent costs (Theorem 2(iii)).

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
federated continual learning, concept drift, energy accounting, edge inference, identifiability, adaptive computation, hardware measurement

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_gqg2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates how much policy information is retained by energy observations under fixed event costs and binary arrivals. It derives a trigger count relation from queue conservation and compares gated adaptation with schedules having matched counts. It also develops a phase-level energy ledger and auditing procedure and reports embedded measurements of the counting regime and several cases where adaptation helps or fails.

Critical strengths
  • Separating better scheduling from fewer updates is a useful attribution question, and work-matched energy comparisons are valuable.
  • The paper distinguishes metered, replayed, simulated, and predicted evidence and reports important negative results and cost variability.
  • The counting identity has an intuitive induction proof for the stated nondecaying threshold queue when the maximum arrival does not exceed the service quantum.
Critical weaknesses
  • The scope is much narrower than the title suggests. Constant event cost makes energy depend on count by construction; the specific queue additionally determines count from arrivals. This does not establish that adaptive gating generally reduces to counting. Timing and sample selection can affect accuracy or future work even with constant event cost. An oracle matched using future arrival totals is not a deployable information-matched baseline.
  • The regret guarantees are unsupported as stated. Theorem 3 lacks a fully specified loss/feedback model, comparator, learning update, drift budget, and detection assumptions. Its appendix proof invokes “standard” analysis and substitutes energy constants. Theorem 4 offers structural analogies to a cited result rather than a valid reduction of decision processes. Bounded reward alone is insufficient.
  • The appendix incorrectly equates the accumulating threshold gate with Q_next=(Q+a-Theta)^+. For Q=0, a=1, Theta=2, these yield one and zero respectively. The Le Cam proof also contains alpha+(1-alpha)=1/2 and confuses balanced classification success with test power. Identical distributions imply power equals size, not an unconditional power bound of one half.
  • Table 4 gives paired means of 0.712 J and 0.587 J, implying a gate-minus-oracle mean difference of +0.125 J, whereas Theorem 2(ii) states an arithmetic mean difference of -0.13 J. A negative Hodges–Lehmann estimate can coexist with a positive mean, but does not reconcile this mean-sign contradiction.
  • A central measured difference within 2c_ad does not validate a deterministic trajectory bound under the reported event-cost variability. Nonsignificant queue-dependence tests do not establish constant cost. Exact matched counts already give equal idealized energy; the extra residual needs clarification. Reported intervals also explicitly exclude Type B measurement uncertainty.
  • Hashes and sum consistency validate aspects of a ledger, not automatically its physical truth against a malicious reporter. Injected export faults cover a limited threat model. Hardware energy and simulated learning outcomes must be linked carefully before interpreting them as joint end-to-end operating points.
Questions for authors

Q1. What are the comparator, loss/feedback model, learning update, and drift assumptions for Theorem 3? Can you provide its complete proof and show that the reduction in Theorem 4 preserves every required assumption?

Q2. Why does Appendix A equate the accumulating gate with Q_next=(Q+a-Theta)^+? These differ at Q=0, a=1, Theta=2. Which recurrence does the code implement, and which claims require correction?

Q3. Why do Table 4's paired means imply a +0.125 J gate-minus-oracle difference while the text states -0.13 J? Please reconcile the paired data, sign convention, arithmetic mean, and Hodges–Lehmann estimate.

Q4. With variable measured event costs, is 2c_ad an ideal-model, probabilistic, or measured pathwise bound? How are residual costs, systematic uncertainty, and timing dependence incorporated?

Q5. Does oracle count matching require the realized future arrival total A(T)? What happens against deployable open-loop baselines with the same causal information and learning-quality requirement?

Q6. Which sensors, logs, and hash anchors are trusted? How does the audit distinguish an internally consistent ledger from a malicious but internally consistent account of physical energy?

Reviewer_ZsuCOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

When adaptive systems such as early exits or gated adaptation report energy savings, it is unclear whether the savings come from the adaptive policy or only from how many actions it takes. The authors study this on federated learning nodes that keep learning from data streams whose distribution changes, using a Raspberry Pi with a hardware power sensor. A detector flags drift events and a gate decides when flags become adaptation triggers, each costing a roughly constant amount of energy. They prove that under constant cost per trigger, a policy affects total energy only through its trigger count. Two policies with the same count then cannot be told apart from energy measurements, and the gated controller stays within twice the per-trigger cost of a fixed schedule with the same count. Exhaustive enumeration of arrival orderings and hardware runs support the claim where the gate threshold is at least one event, and it fails below that. The authors also build a per-phase energy ledger with audit checks that detects under-reporting of energy.

Critical strengths

None

Critical weaknesses

The paper appears to be AI written and is very hard to understand. Even the abstract is full of symbols, such as "Nfire ≡ ⌊A(T)/Θ⌋", and numbers with no plain explanation. Many terms are invented and not explained simply. Examples are "scalar-mediated energy-identity collapse," "calibers," "evidence levels M/R/S/P," "FUS-4," and "EBDA." Phrases like "a one-dimensional mediation, not an absence" and "the circularity objection is itself the finding" are confusing and explain nothing. The main idea is simple: energy depends on how many times the system triggers. It could be said in two sentences. Instead, the reader has to piece it together across the whole paper.

I interpreted the text with another AI and found several issues. The main result is a tautology, which means it is true by assumption. It follows directly from assuming that each trigger costs a fixed amount. Some issues are left unresolved, such as a predicted effect that could not be detected. Some conclusions are not statistically significant, such as an energy saving whose confidence interval includes zero.

Questions for authors

None

Reviewer_qMZ6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper looks into what is the observation channel for early exits, token pruning and gated adaptation solutions when they report energy saving under federated learning paradigm. The options include event-driven binary signals, benefit of adaptive inference be attributed to the policy itself, or only to the number of events it happens to trigger.

Critical strengths
  • I like how they focused on the problem on a specific class of hardware (Pi Class)
Critical weaknesses
  • I am not sure if the Assumption 2 is fair. "debt queue does not decay between events" -> is not a valid assumption for. most real world cases where these Pi Class devices are most important.
  • Assumption 4 "Per-step reward is bounded" is not valid for most deep learning algorithm. What is the basis of this assumption?
  • The paper proposes 5 theorem and one lemma but it is unclear on what is the need for them or why they are significant in solving the proposed question.
  • The paper started with early exits, token pruning and gated adaptation but the experiments are mostly dominated by training costs. I understand that this is a continual federated learning paradigm but inference is the key for these solutions.
  • It is not clear what model architectures and algorithms are used for the evaluation
Questions for authors
  1. Can you please provide the reasoning and justifications behind the assumptions?
  2. What are the justifications or necessity for the theorems.
  3. Can you provide the energy benefit analysis during inference?
  4. What is the motivation of this analysis for continual federated learning?
  5. Can you provide more details on the evaluation setup in terms for model architecture and adaptive algorithm?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.