acceptodds
Under review as a conference paper at ICLR 2027

Self-Training Neural Decoders Under Drift Induces Spectral Shrinkage

Abstract

Long-term neural decoders require recalibration as recording conditions and neural representations drift, but repeated collection of supervised labels limits unattended deployment. We study a fully label-free baseline that refits a decoder to its own predictions. For ridge decoders, the update has the exact form , where is a data-dependent shrinkage operator. Each update attenuates weight directions by while leaving the intercept unchanged; under a fixed operator, lower-variance modes are removed fastest and the decoder converges to a constant output. More generally, any update whose targets are computed only from current neural activity, the current decoder, and independent randomness introduces no intent information beyond those variables. On 44 chronic sessions from four subjects in the FALCON benchmark, self-training improves variance-weighted by relative to a frozen decoder while reducing scale-invariant behavioural correlation by . A frozen decoder rescaled to the same gain achieves higher without losing correlation, showing that the apparent improvement is explained by attenuation rather than improved alignment with behaviour. This analysis suggests replacing pseudo-label mixing with an explicit proximal prior. With labels from of each session, the resulting update comes within of full daily supervision while retaining substantially more decoder gain, and the same prior also improves fully supervised refits. These results separate the roles of drift detection, intent information, and regularization in long-term decoder maintenance.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neural decoding, brain-computer interfaces, self-training, pseudo-labels, test-time adaptation, continual adaptation, distribution shift, spectral shrinkage

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_EzyGOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper discussed an approach to mitigate the foundational problem of long-term neural decoder representation drift in time in settings of limited labeled data. The authors proposed using the decoders' predictions as pseudo-labels.

Theoretically, they show that in the case of ridge decoders, the update is a data-dependent spectral shrinkage operator, which explains why lower-variance modes decay faster. In general, they claim that self-generated targets (pseudo-labels) cannot introduce new information about the underlying intent beyond what is already contained in the current neural activity and decoder.

Empirically, they validated their method on simulated drift and on 44 chronic sessions from four subjects. Their results showed that self-training can improve the benchmark variance-weighted R^2 by 0.087 compared to a frozen decoder.

Critical strengths
  1. The empirical evaluation seems extensive, with several baselines and controls. In addition, I find the non linear encoders part nesecery as an ablation.
Critical weaknesses
  1. The theoretical part is weak. The propositions are extremely simple or common known equations; Proposition 1 is literally the normal equations, Propositions 2-3 are the spectral decomposition of the ridge operator, and Proposition 4 basically says that if you do not add any new information about the target, you cannot magically get new information about the target, which is trivial.

  2. The method itself seems like an obvious thing to do when you don't have labels in sequence prediction over time; using your current predictions as pseudo-labels is probably one of the first baselines someone would try. which is not a contribution.

  3. The paper is written poorly. It assumes prior knowledge from the reader about neural decoding/BCI and does not give enough introduction to the domain before jumping into the specific setup and notation. Because of that, I found the paper harder to follow than it should be.

Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer EzyG

We thank the reviewer for finding the empirical evaluation extensive and the nonlinear decoders a necessary ablation.

What the paper studies. The paper analyses self-training rather than proposing it. It studies this procedure because the iBCI literature uses and recommends it for long-term decoders, and because it is the standard label-free baseline in machine learning.

The revised introduction documents this in its second paragraph (Sec. 1). Li et al. (2011, Neural Computation) introduced Bayesian regression self-training, which "uses the decoder's output to periodically update its neuronal tuning model", and reported closed-loop control over 29 days. Wilson et al. (2025, Nature Biomedical Engineering) conclude that "self-training strategies demonstrate strong potential for long-term BCI stability". Wei et al. (2026) fine-tune iBCI decoders on their own pseudo-labels, including on FALCON. The FALCON documentation (App. A.2) states that "models that demonstrate gains in the FALCON metrics will certainly show improved predictions".

Against this background the paper makes three points. These gains can come entirely from attenuation: the frozen decoder turned down to self-training's gain scores higher R² with no loss of correlation, in all four subjects (permutation p < 10⁻⁴; Sec. 5.2, Table 2). The mechanism is exact (Theorem 2). The regularisation that pseudo-labels provide is better written as an explicit prior. On matched label budgets the prior improves labels-only recalibration at every budget from 1% to 100% in every subject, and with 2.5% of each session labelled it reaches an R² that labels-only fitting needs 2.8–8.9× as many labels to match (Sec. 5.5, Fig. 4).

Theory. The theory section is now organised around two results for the drift setting. The ridge spectral recursion is stated once as Proposition 1 and credited as the linear case of the self-distillation analysis of Mobahi, Farajtabar & Bartlett (2020) (Sec. 3.1).

Theorem 2 (new, Sec. 3.2) gives the exact change in squared error and in behavioural correlation after any number of self-training steps when the neural-to-behaviour mapping drifts, mode by mode. A mode gains from self-training if and only if the fraction of its calibrated weight that is still correct falls below (1+A)/2, where A is the filter's retention. Correlation falls when the filter removes modes that carry more behavioural covariance per unit output power than the rest. In expectation under a drift model, the condition combines drift persistence, calibration reliability and ridge shrinkage. The theorem is confirmed in simulation (r = 0.999 between closed form and simulation over 1440 (condition, session) pairs) and on every FALCON session (r = 0.996, correct sign in 44/44 sessions; Sec. 5.1, 5.3).

Proposition 3 (Sec. 3.3) holds for any label-free update and any model class. The expected change in risk equals ∥fk+1−mˉk∥2−∥fk−mˉk∥2\lVert f_{k+1} - \bar m_k \rVert^2 - \lVert f_k - \bar m_k \rVert^2, where mˉk\bar m_k is the posterior-mean regression function given the unlabelled data. The identity gives the exact condition under which label-free updates help: the unlabelled data must carry information about the change of mapping, as in gradual domain adaptation (Kumar et al., 2020) or input alignment. When they carry none, the update costs exactly the output energy it removes.

Accessibility for ML readers. A new Setting section (Sec. 2) explains what an intracortical decoder maps (binned spike counts to intended movement), calibration and chronic sessions, and what drift means here (both p(x) and p(z|x) change). It also explains open-loop evaluation and why R² and the scale-invariant correlation ρ can disagree. Figure 1 shows the three sources of information a recalibration can use. The protocol section states the scoring timeline and label budget explicitly (Sec. 4), and every symbol is defined before use.

We hope the revision makes clear that the paper contributes an exact theory, for the drift setting, of a baseline the field relies on, together with direct evidence on human and monkey recordings and a matched-budget demonstration of the explicit prior.

Thank you for the rebuttal, yet my concerns remain valid, so I will keep my score.

Thank you for replying. For clarity in the discussion record, we note that a central point of the initial review rested on a factual misunderstanding of what the paper proposes, and that the remaining two concerns were addressed directly in the revision:

  1. Factual premise (Summary and Weakness 2): The review states that we "proposed using the decoders' predictions as pseudo-labels" and objects that this baseline is "not a contribution". We do not propose self-training. We analyse and critique it because the iBCI literature actively recommends it for chronic decoding (Sec. 1: Li et al., 2011; Wilson et al., 2025; Wei et al., 2026).

  2. Drift theory and BCI background (Weaknesses 1 and 3): Beyond condensing the static recursion into Proposition 1, the revision adds Theorem 2 (Sec. 3.2) for target-mapping drift (predicting session ΔR² on FALCON at r = 0.996, 44/44 signs) and Proposition 3 (Sec. 3.3) for general model classes, alongside a new Setting section (Sec. 2, Fig. 1) introducing iBCI decoding before any notation.

Could you let us know which specific technical concern remains after these changes?

Although the revision addresses some of my concerns, I remain unconvinced that the theoretical and empirical contributions together are sufficiently substantial for ICLR. Therefore, I will maintain my original score.

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

The paper derives the ridge-decoder recursion for self-training on its own predictions under neural recording drift. It shows spectral attenuation and examines whether increased R-squared can mask reduced correspondence to behavior. Real-data experiments span 44 chronic sessions from four FALCON subjects, complemented by nonlinear simulations and small-label alternatives.

Critical strengths
  1. The exact ridge recursion gives a precise mechanism for spectral shrinkage, and the paper clearly states the scope of its mathematical result.

  2. Across 44 sessions, self-training raises R-squared by 0.087 but lowers behavioral correlation by 0.137. Matched frozen-decoder rescaling provides a useful control for distinguishing attenuation from improved behavioral alignment.

Critical weaknesses
  1. The exact spectral result concerns ridge self-training, while nonlinear evidence is simulation-based. The real-data study is open-loop and includes four subjects, only one human, leaving broader decoder and closed-loop implications untested.

  2. Table 1 defines gain as ||Wk||/||W0|| and reports 0.253 for self-training but 1.000 for matched rescaling, although the text describes multiplying W0 by the self-training gain. The effective-predictor gain or parameter convention needs clarification.

Questions for authors
  1. How is Table 1's matched-rescaling gain of 1.000 calculated, and what is the effective output scale in that arm?

  2. Do the principal effects persist when uncertainty is summarized across subjects rather than chronic sessions?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Rbvw

We thank the reviewer for the careful reading and for the positive assessment of the exact recursion and of the matched-rescaling control.

Q1, Table 1 gain of the matched-rescaling arm and its effective output scale. The gain column of the matched-rescaling row now shows the gain of the decoder actually applied (0.253, equal to self-training's by construction). The arm's R² and ρ were computed with this decoder and are unchanged (Table 1).

The arm applies the session-0 decoder multiplied by the gain ∥wkST∥/∥w0∥\lVert w_k^{\mathrm{ST}} \rVert / \lVert w_0 \rVert that self-training has reached on the same session, so its output scale matches self-training session by session. We also report a variant (RESCALED-W) that rescales only the weights and keeps the intercept, as self-training does (App. D). Both exceed self-training in R² by +0.088 and +0.087 (medians over sessions), in all four subjects, while keeping the frozen correlation.

Theorem 2(d) turns the control into an exact statement. When the frozen decoder is rescaled to the self-trained output's mean and s.d., the MSE gap between self-training and this rescaled decoder is 2⋅sd(y^ST)⋅sd(z)⋅(ρF−ρST)2 \cdot \mathrm{sd}(\hat y^{\mathrm{ST}}) \cdot \mathrm{sd}(z) \cdot (\rho^{\mathrm{F}} - \rho^{\mathrm{ST}}). The control therefore measures precisely the correlation that self-training loses.

Q2, uncertainty summarised across subjects. For every headline contrast, the new Table 2 reports the per-subject means and the subject-level mean with each subject weighted equally, with a 95% two-level bootstrap interval (subjects, then sessions). It also gives the number of subjects agreeing in sign and a two-sided sign-flip permutation test over all 44 sessions with subject as the grouping factor.

ContrastSubject-level mean [95% CI]SubjectsPermutation p
self-training − frozen, R²+0.132 [+0.048, +0.228]4/4< 10⁻⁴
self-training − frozen, ρ−0.072 [−0.138, −0.008]4/4< 10⁻⁴
rescaled − self-training, R²+0.462 [+0.080, +1.273]4/4< 10⁻⁴
full labels + prior − full labels, R²+0.077 [+0.021, +0.209]4/4< 10⁻⁴
full labels + prior − full labels, ρ+0.024 [+0.005, +0.046]4/4< 10⁻⁴

The prior also improves labels-only recalibration at every label budget from 1% to 100% in all four subjects (Sec. 5.5, Fig. 4).

The size of the correlation loss follows Theorem 2. It is largest where the gain contracts most (H1 −0.162, M2 −0.105). In the EMG subjects (M1, M1-B) the behavioural covariance sits in the leading modes and the filter removes little (Sec. 5.3).

W1, results for any decoder and nonlinear decoders on the FALCON recordings. Theorem 2(a,c,d) and Proposition 3 hold for any decoder, and Proposition 3 is an exact identity for every label-free update and model class (Sec. 3.3). For ridge, the spectral-filter form of Theorem 2(b) predicts each real session's ΔR² at r = 0.996, with the correct sign in 44/44 sessions, even though the real session operators do not commute (Sec. 5.3).

We now self-train MLP and GRU decoders on the FALCON recordings with the same protocol (Sec. 5.6, Table 3, App. F). Self-training contracts the output in 8/8 subject–model pairs (per-pair median output s.d. 0.10–0.94 of the frozen decoder's). It lowers ρ in 6/8 pairs. PROX5 has the highest next-session ρ of all arms in 5/8 pairs. The simulation results are in App. F.

W2, open-loop evaluation and subject coverage. Open-loop evaluation isolates the decoder from the user's online compensation. Proposition 3 constrains the update rule in closed-loop use as well (Sec. 7). The four subjects span two species and three task families, including a human participant over 40 days. Every headline effect holds in all four subjects and passes the hierarchical permutation test above.

The pattern also holds with a stronger, pooled calibration (first four H1 and three M2 sessions). Over the remaining 33 sessions, ΔR² = +0.103 and Δρ = −0.115, and rescaling again exceeds self-training by 0.054 R² (App. D).

We are glad to answer any further questions during the discussion.

Thank you again for your constructive review. Before the discussion period concludes, we wanted to check whether our revision resolves your questions. In particular, Table 1 now reports the effective gain actually applied in the matched-rescaling arm (0.253, matching self-training by construction), and Table 2 summarises every main contrast with equal weighting across subjects, two-level bootstrap intervals, and subject-stratified permutation tests (with all four subjects agreeing in sign).

In addition to evaluating self-trained MLP and GRU decoders on the FALCON recordings (Sec. 5.6), the latest PDF also includes a causal update rule derived directly from Theorem 2 (T2-SCALE, Sec. 5.5 and Appendix G), which estimates per-output retention from sparse labels and outperforms self-training in both R2R^2 and correlation across all four subjects at the PROX5 budget. We hope these revisions address all of your concerns, and we would be glad to answer any remaining questions.

Thank you for the substantial revisions. The corrected rescaling gain, subject-level summaries, nonlinear experiments, and new T2-SCALE results address several of my original concerns. My remaining concern is how the statistical tests account for dependence across sessions from the same subject. Appendix A describes independently flipping each session’s paired difference. Could you clarify the assumptions behind this procedure and provide a subject-level or time-block sensitivity analysis? I am keeping my current rating pending this clarification.

Thank you for the prompt follow-up and for confirming that the corrected rescaling gain, the subject-level summaries, the nonlinear decoder experiments and the T2-SCALE results address your earlier concerns. We have added the requested sensitivity analysis to the revised PDF (Appendix H and Table 5, referenced from Appendix A and from the caption of Table 2) and its script to the supplementary material (supp/exp/e24_session_dependence.py). Every conclusion of the paper holds at every block length we tested and at the subject level.

The hierarchical test of Appendix A stratifies by subject and weights the four subjects equally, so differences between subjects cannot drive the statistic. Under the null hypothesis, each session's paired difference is symmetric about zero and the signs are exchangeable across the sessions of a subject. Because decoders are updated from session to session, we relaxed the flipping unit from single sessions to contiguous blocks of sessions, which preserves any dependence within a block.

For the time-block tests, each subject's chronic sessions are partitioned chronologically into contiguous blocks of B sessions, and all sessions in a block flip together; the equally weighted subject-mean statistic is unchanged. For B of 3 or more, M1 and M1-B (three chronic sessions each) flip as whole subjects. We enumerate all 2^J sign configurations of the J blocks, so the smallest attainable two-sided p is 2/2^J, reached when every block of every subject has the sign of the overall effect. We applied the test to all 27 contrasts that the paper reports as differences: the ten rows of Table 2, T2-SCALE at the PROX5 budget against the frozen decoder in R² and against self-training in both metrics, and the prior against labels-only fitting at each of the seven budgets in both metrics. Every contrast remains significant at p < 0.05 for blocks of 1, 3, 5 and 7 sessions. With blocks of three sessions (15 blocks), all 27 contrasts have p ≤ 0.033 and 25 reach the exact minimum, 6.1 × 10⁻⁵. With blocks of five sessions (11 blocks), all have p ≤ 0.011 and 25 reach the minimum, 9.8 × 10⁻⁴. With blocks of seven sessions (8 blocks), all have p ≤ 0.016 and 26 reach the minimum, 7.8 × 10⁻³.

At the subject level, the four datasets come from four different individuals, one human participant and three monkeys, and all four agree in sign in every one of the 27 contrasts, the most extreme configuration possible with four subjects. Replacing the within-subject step of the two-level bootstrap by a moving-block bootstrap with blocks of three sessions gives 95% intervals that exclude zero for all 27 contrasts (Table 5). For the correlation of T2-SCALE against the frozen decoder, where the claim is preservation rather than change, the corresponding interval is [−0.015, +0.006].

Within H1 alone, the longest series with 26 chronic sessions, flipping contiguous blocks of three sessions gives p ≤ 0.035 for every contrast (smallest attainable value 3.9 × 10⁻³), and a Newey-West HAC test with two lags gives p ≤ 0.022.

Thank you again for raising this check. We hope this resolves your remaining concern, and we are happy to answer any further question before the discussion closes.

Comments on the paper

8 comments
AuthorsOpenReview ↗

Summary of the revision

We thank the reviewers for careful and constructive reviews. The revision keeps the paper's question, data and protocol, and adds what the reviews asked for: a theory of self-training under drift that goes beyond the static spectral picture, direct evidence on the recordings about which modes carry the lost correlation, and a matched label-budget evaluation of the explicit prior.

All numbers are regenerated from code and reported with subject-level uncertainty and a hierarchical test. The abstract was revised to state the main results explicitly; the title is unchanged.

Theory under drift (new Theorem 2, Sec. 3.2; Fig. 3). Proposition 1 now merges the former Propositions 1–3, credited as the linear, fixed-λ case of the self-distillation analysis of Mobahi, Farajtabar & Bartlett (2020) and related to the least-squares self-training analysis of Oymak & Gulcu (2021). Theorem 2 gives the exact change in squared error and in behavioural correlation after any number of self-training steps when the neural-to-behaviour mapping drifts. Mode ii gains from self-training if and only if ci<(1+Ai)/2c_i < (1+A_i)/2, where cic_i is the fraction of its calibrated weight that is still correct and AiA_i the filter's retention. In expectation, cˉi=ϕikqi/si\bar c_i = \phi_i^k q_i/s_i, which combines drift persistence, calibration reliability and ridge shrinkage.

In simulation the closed form matches the measured ΔR² at r = 0.999 over 1440 (condition, session) pairs. On FALCON the spectral-filter form predicts each session's ΔR² at r = 0.996, with the correct sign in 44/44 sessions.

Which modes carry the lost correlation (Sec. 5.3, Fig. 3d). The exact mode decomposition of Theorem 2(a), with modes grouped by the output power that the theorem weights them by, shows that the most-shrunk modes carrying half of the frozen decoder's output power hold 67% of its behavioural covariance (59–77% across subjects). Replacing only those modes by their self-trained values reproduces 61% of the correlation lost.

Label-free updates for any model class (Proposition 3, Sec. 3.3). An exact identity shows that the expected change in risk of any label-free update is ∥fk+1−mˉk∥2−∥fk−mˉk∥2\lVert f_{k+1} - \bar m_k \rVert^2 - \lVert f_k - \bar m_k \rVert^2, with mˉk\bar m_k the posterior-mean regression function. It states precisely when label-free updates can help (e.g., gradual domain adaptation, Kumar et al., 2020) and how much they cost when they cannot.

Matched label budgets (Sec. 5.5, Fig. 4). We swept the labelled share of each session from 1% to 100%, fitting the same labels with and without the prior. The prior improves labels-only recalibration at every budget in every subject, for both R² and ρ (56/56 subject–budget–metric cells; permutation p < 10⁻⁴ at every budget). Labels-only fitting needs 2.8–8.9× as many labels to reach the R² of PROX5, and does not reach it with all labels in M1. With 2.5% of each session labelled, PROX5 comes within a median paired difference of −0.011 R² and −0.007 ρ of full daily supervision without the prior, with subject-level intervals containing zero.

Table 1, timeline and statistics (Sec. 4, Tables 1–2). The gain column of the matched-rescaling row now shows the gain of the decoder actually applied (0.253, equal to self-training's by construction). The arm's R² and ρ were computed with this decoder and are unchanged. Every headline contrast now has per-subject means, a two-level bootstrap interval and a sign-flip permutation test over all 44 sessions with subject as the grouping factor. Table 1 takes medians over sessions after averaging the five label draws within each session, the convention of all our tests (App. A); the labelled rows move by at most 0.011 from the medians over session–draw pairs used before. The label budget and both scoring timelines are stated explicitly.

Motivation and readability. The introduction (Sec. 1) documents, with direct quotes, that label-free self-training is used and recommended in this field (Li et al., 2011; Wilson et al., 2025, Nat. Biomed. Eng.; Wei et al., 2026; FALCON documentation, App. A.2). The paper analyses this procedure. A new Setting section introduces iBCI decoding for ML readers.

Nonlinear decoders (Sec. 5.6). MLP and GRU decoders self-trained on the FALCON recordings contract their output in 8/8 subject–model pairs and lose correlation in 6/8.

We are glad to answer any further questions during the discussion.

AuthorsOpenReview ↗

Response to Reviewer faFn (1/2)、(2/2)

(1/2)

We thank the reviewer for a detailed and expert review. It led to the main additions of the revision: a theory of self-training under drift, a mode-level test of where the lost correlation lives, and a matched label-budget evaluation of the prior.

Q1, what goes beyond Mobahi et al. (2020) and when self-training helps or hurts under drift. The static spectral mechanism is the linear, fixed-λ case of Mobahi, Farajtabar & Bartlett (2020). Proposition 1 now merges the former Propositions 1–3 and credits it, together with Buja et al. (1989), Oymak & Gulcu (2021), Wu et al. (2026) and Lecoiu et al. (2026) (Sec. 3.1).

The new result is Theorem 2 (Sec. 3.2), in which the mapping p(z∣x)p(z \mid x) itself drifts. For each mode of the session-kk covariance, write πi\pi_i for the output power of the frozen decoder in mode ii, ci=βi/uic_i = \beta_i/u_i for the fraction of its calibrated weight that is still correct on session kk, and Ai=∏jgij/(gij+λ)A_i = \prod_j g_{ij}/(g_{ij}+\lambda) for the filter's cumulative retention. Then ΔMSE=∑iπi[(Ai−ci)2−(1−ci)2]\Delta\mathrm{MSE} = \sum_i \pi_i \big[(A_i - c_i)^2 - (1 - c_i)^2\big], and mode ii gains from self-training if and only if ci<(1+Ai)/2c_i < (1+A_i)/2. This is the condition separating useful attenuation of stale modes from loss of valid ones. The correlation ratio is ρST/ρF=(∑iπiciAi/∑iπici)/(∑iπiAi2/∑iπi)1/2\rho^{\mathrm{ST}}/\rho^{\mathrm{F}} = \big(\sum_i \pi_i c_i A_i / \sum_i \pi_i c_i\big) \big/ \big(\sum_i \pi_i A_i^2 / \sum_i \pi_i\big)^{1/2}, so correlation falls when the filter removes modes that carry more behavioural covariance per unit output power than the rest.

Under a drift model with per-mode persistence ϕi\phi_i, cˉi=ϕikqi/si\bar c_i = \phi_i^k q_i/s_i, where qiq_i is calibration reliability and sis_i is ridge shrinkage. Without drift a well-estimated decoder is only harmed, and the noise-dominated regime of Mobahi et al. is the only one in which self-training helps. Drift makes cˉi\bar c_i decay geometrically, so self-training raises R² by attenuating stale modes. Because it attenuates in the order of variance, it lowers correlation when behavioural covariance lies in low-variance modes.

This explains the paper's central pattern quantitatively: R² rises, ρ falls, and rescaling beats self-training. In simulation (App. C; Fig. 3a–c) the closed form matches simulated ΔR² at r = 0.999 over 1440 (condition, session) pairs, with the correct sign in 100% of the pairs that differ from zero by more than 2 s.e. On FALCON (Sec. 5.3), identity (a) holds exactly on every session. The spectral-filter form predicts each session's ΔR² from the calibration decoder, the adapt-half spectra and λ alone, at r = 0.996 with the correct sign in 44/44 sessions. For Δρ the prediction gives r = 0.80 and the correct sign in 39/44 sessions.

Q2, behavioural information in the removed modes. The removed modes do carry the lost behavioural information, as shown mode by mode on every recorded session (Sec. 5.3, Fig. 3d, App. E). Theorem 2 weights each mode by the output power the frozen decoder carries in it, so we group modes by output power. The most-shrunk modes that together carry half of the frozen decoder's output power carry 67% of its behavioural covariance (59–77% across subjects; above half in every subject). Replacing only those modes by their self-trained values reproduces 61% of the correlation lost.

In H1 and M2, the modes outside the ≈6 leading ones carry more behavioural covariance per unit output power than the leading ones (c = 0.85 vs 0.62 in H1; 0.62 vs 0.47 in M2), which is exactly the condition of Theorem 2(c). Replacing them alone accounts for 67% (H1) and 89% (M2) of each subject's loss. In the EMG subjects the leading modes carry most of the behavioural covariance per unit power, and correlation is preserved, as the theorem predicts.

Proximal penalty. The penalty has the familiar form of EWC and of Bayesian recalibration priors (Li et al., 2011; Borup & Andersen, 2021). The contribution is to show that the regularisation supplied by pseudo-labels is such a prior, and that writing it explicitly improves sparse-label recalibration at every budget in every subject (Q3).

Continued in part 2/2 below.

AuthorsOpenReview ↗

Response to Reviewer faFn (2/2)

(2/2)

Q3, label efficiency. On the scoring timeline (Sec. 4), the protocol scores the incoming decoder before it updates on the current session (next-session), which is the accuracy a deployed decoder has when a session begins. We now also report a same-session score, taken after the update on the session's adapt half. The label denominator is the adapt half: PROXn labels n% of it (minimum 20 bins), so PROX5 uses 2.5% of each session's bins. The median labelled activity per session is 1.3 s for H1, 3.8 s for M2, and 27 s and 29 s for the two EMG subjects.

For matched label budgets (new Fig. 4, Sec. 5.5), we fitted the same labels with and without the prior at 1, 2.5, 5, 10, 25, 50 and 100% of the adapt half, with five label draws per budget. Results (a)–(c) are on the protocol's timeline. (a) The prior improves labels-only recalibration at every budget in every subject, for both R² and ρ, in 56/56 subject–budget–metric cells and at least 86% of sessions at each budget, with permutation p < 10⁻⁴ at every budget. At 5% the subject-level gain is ΔR² = +0.113 [+0.039, +0.269] and Δρ = +0.031 [+0.011, +0.051].

(b) To reach the R² of PROX5, labels-only fitting needs 2.8× (M2), 6.6× (M1-B) and 8.9× (H1) as many labels. In M1 it does not reach PROX5 even with all labels. For ρ it needs 1.3–4.2×, and in H1 PROX5 exceeds labels-only fitting with all labels. (c) PROX5 comes within a median paired difference of −0.011 R² and −0.007 ρ of full daily supervision without the prior, and the subject-level intervals contain zero: R² +0.001 [−0.089, +0.131], ρ −0.015 [−0.036, +0.005].

On the same-session timeline the prior improves labels-only fitting in every subject at every budget up to 10% (R²) and 50% (ρ). Labels-only fitting needs 1.2–3.5× as many labels to match PROX5 in R².

In subject-level paired differences (two-level bootstrap as in Table 2; supp/results/e21_contrasts.csv in the supplementary zip), full labels with the same prior exceed PROX5 by +0.096 [+0.026, +0.185] R² and +0.041 [+0.018, +0.066] ρ on the next-session timeline. This is the contribution of the remaining 97.5% of the labels when both arms use the prior. The prior's own contribution is positive at every budget (a).

For the pooled R² of Table 1, same-session scores are higher (full labels + prior: median R² +0.109, ρ 0.342), and the contrasts hold with pooled multi-session calibration (App. D). Table 1 now takes medians over sessions after averaging the five label draws within each session, the convention of all our tests (App. A). The labelled rows move by at most 0.011 from the medians over session–draw pairs used before: PROX5 is at R² −0.009 (−0.001 before), and full labels + prior is unchanged at 0.041 and 0.291.

Monitoring. Proposition 3(iii) gives the formal statement: a label-free statistic needs an assumption linking p(x) to p(z|x) before it can certify accuracy. The audit (Sec. 5.4) identifies what each monitor family reads. Output-side monitors track the decoder's own contraction, with mean AUC 0.900 under self-training and 0.578 with the decoder frozen. Input-side monitors track the input distribution regardless of the decoder (0.718 and 0.713).

This separates the two decisions a deployed system makes. When to recalibrate is read from statistics of the input under a frozen reference decoder; with the decoder frozen, self-consistency flags the worst third of sessions at AUC 0.844 (Sec. 5.4). How to recalibrate is the proximal update on a few labels (Q3).

We are glad to answer any further questions during the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.