- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper uses a Poisson autoencoder with latent partition into stimulus-private, choice-private, and shared subspaces; causal convolutional encoder; subspace-specific temporal attention gates; The strongest defensible novelty is therefore the specific task-private/shared + learned temporal-gating construction. It is validated on synthetic data with known ground truth and barrel-cortex recordings from a mouse whisker task; main claims are highest held-out stimulus decoding, emergent alignment of stimulus readout with whisker contact, and stimulus-before-choice ordering.
Critical strengths
-
The scientific question is the best thing about this paper. Asking not just whether stimulus and choice are decodable, but from which subspace and when on each trial, is a meaningful step beyond dPCA-style methods. Nobody has done this within-region private/shared split with single-trial temporal readout weights simultaneously before, and the related work section makes a convincing case for that gap.
-
The ablation suite is genuinely thorough — layout search, scorer variants, label shuffles, loss term ablations, reconstruction weight sweep.
Critical weaknesses
-
In figure 3B related to real data results, the stimulus-private block decodes stimulus at 0.747 and choice at 0.690. The choice-private block does better (0.877 vs 0.637). The selectivity of the stimulus-private block is very weak. The authors acknowledge this but frame it as "expected" because choice is more decodable and the variables are correlated. However this seems a weakness in the core claim about private subspaces.
-
Synthetic benchmark comparison seems insufficient. For a method claiming to identify which subspace is read out and when, you would want to know whether the model correctly identifies choice-first ordering when ground truth is choice-first, or whether it always biases toward stimulus-first due to the loss weighting asymmetry. Also authors do not compare against TAME-GP or any external method with known ground-truth recovery, which would be the natural use of having a synthetic benchmark.
-
A simple causal baseline comparison would help clarify the results. Appendix F.1 already reveals that raw spikes carry the stimulus-before-choice ordering without any subspace model. A causal convolutional model with the identical attention pooling mechanism applied directly to raw spikes — no encoder, no decoder, no latent partition — would tell you how much of TSA's temporal performance comes from the attention gate architecture versus from the subspace decomposition. Without this, you cannot attribute the timing result to the scientific contribution of the paper.
-
Table 7 explicitly shows that the 23/23/24 layout was selected because it has "the highest cross-seed readout reproducibility of any layout... and the highest correlation between the stimulus readout and whisker contact." The paper states: "The balanced allocation is therefore justified on the grounds... a readout that is stable across initializations and aligned to the sensory event." Contact alignment (ρstim,contact) was directly used to choose which layout to report. It is then also the primary external validation signal claimed in Section 5.3 and Table 1. This is circular. The independence claim of the contact alignment result depends on contact having no role in model decisions, but it was used to choose the latent layout.
-
inconsistency: Main text Section 4 says "stimulus attention peaked before choice attention on 0.949 ± 0.006 of trials." Appendix Table 6 for the same experiment reports TSA (full) Stim-first = 0.694. These both describe synthetic data with 9 runs (3 dataset × 3 model seeds). That is a 0.255 difference on the key timing metric, could you please clarify this discrepancy?
Minor weakness:
-
Please number the equations so they can be referred to.
-
Please improve the caption of figure 2 for clarity. The legend used in second column also causes confusion, do you mean latent instead of neuron?
Questions for authors
- The denominator in the attention weight formula sums sigmoid scores over all T time bins including τ > t. This means α_i^r(t) requires knowing future scores at inference time. Can you explain how strict causality is maintained under this normalization?
-
The choice of 70-dimension latent with 14 neurons in real-data seems very unusual, have authors tried low dimensional latent and did not get better results?
-
Were you aware of “Multi-Integration of Labels Across Categories for Component Identification in Multi-trial Time Series” (ICML 2026), which also partitions multi-trial neural data by label categories with per-trial temporal traces? How would you distinguish TSA's contribution from this framework given that the shared/private category structure and per-trial temporal modeling are both present there?
-
there is no mechanism to capture task-irrelevant variance. All latent dimensions are partitioned into stimulus-private, choice-private, or shared. The reconstruction term forces the model to use all three blocks to explain the spikes, meaning task-irrelevant activity gets absorbed into whichever block has capacity. This contaminates the decomposition. Does the absence of a task-irrelevant subspace affect the interpretability of the shared block? neural activity in the barrel cortex is not fully explained by stimulus and choice variables alone. What happens to movement-related or other non-task variance in the current model?
-
why was TAME-GP not included as a baseline, given it is described as the closest prior work?
3 comments
W3. Apply the same causal attention directly to raw spikes. Done. As you expected, the gate explains most of the decoding and timing:
| main session (the same 20 seeds for both rows; 23/23/24 version) | stimulus acc. | choice acc. | correlation of stimulus attention with touch times |
|---|---|---|---|
| TSA | 0.74 | 0.90 | 0.58 |
| same attention on raw spikes | 0.72 | 0.90 | 0.50 |
On simulated data the raw-spike gate also recovers the per-trial timing. What TSA adds over it: a latent model that also reconstructs the spikes, and private blocks that prefer their own variable on the main session (General Response 5/5). A simpler model with two gates on one shared latent and no subspaces also decodes as well (General Response 3/5), so we do not claim a decoding or timing advantage for the partition. Its role is to make the block names mean something, clearly on the main session and only partly in new sessions (General Response 5/5).
W4. Contact alignment chose the layout and then validated it: circular. Yes, it was. Now:
- no test data or contact is used for any choice (General Response 2/5);
- contact alignment is tested on 17 new sessions where it was never used: stimulus attention follows touch in all 17;
- the layout is no longer argued from contact: all layouts of 36+ dimensions give the same results in the new sessions (General Response 3/5).
We also note that the choice attention follows touch about as well, so contact alignment is no longer presented as evidence for a stimulus subspace.
W5. Stimulus-first 0.949 (Section 4) vs 0.694 (Table 6). The two numbers came from different models and code versions (General Response 2/5). Both are removed; the redone benchmark gives 0.73–0.83 (DZq9 response, Q3).
Minor. Equations are numbered. The Figure 2 legend shows observed neurons grouped by the factor they load on; Figure 2 is replaced together with the old synthetic results, and the new figure labels this explicitly.
Q0. The normalizer sums over future bins; how is causality kept? It is not, for the normalized weight; our wording was wrong. What stays causal: each bin's gate value and latent use only past spikes, and the normalizer divides all bins of a trial by the same number, so it cannot change which bin is highest; peak times and orderings are unchanged. The paper's causality sentence is corrected.
Q1. A 70-d latent with 14 neurons is unusual; did you try low-dimensional latents? Yes. The private blocks can be small, but the whole latent must not be too small (17 new sessions):
| latent | stimulus acc. | run-to-run agreement, stimulus attention |
|---|---|---|
| 6/6/6 (18-d) | 0.69 | 0.31 |
| 12/12/12 (36-d) | 0.68 | 0.34 |
| 23/23/24 (70-d, submitted) | 0.67 | 0.43 |
| 12/12/46 and 2/2/66 (70-d) | 0.68 | 0.43–0.44 |
| 20/20/20 + 10-d null (revised) | 0.68 | 0.41 |
Decoding is the same at every size. With 18 or 36 dimensions, the attention curves differ more between training runs (18-d: lower in all 17 sessions; 36-d: in 14; on the main session the stimulus attention barely repeats, 0.12 vs 0.41). A likely reason: the spikes of each time bin are reconstructed from that bin's latent alone, so the latent must also hold recent history; with few dimensions, little room is left for the task signals.
Q2. How does TSA differ from MILCCI (ICML 2026)? Thank you; we will cite it. Both use per-trial labels and time courses, but:
| MILCCI | TSA | |
|---|---|---|
| labels are | inputs that select components | targets; a test trial is encoded without its labels |
| shared structure | none between categories | an explicit block shared by stimulus and choice |
| time course | smoothed over the whole trial | causal per-bin readout weights |
Q3. No task-irrelevant subspace: where does movement-related variance go? Thank you; we adopted your suggestion. TSA (revised) has a 10-d null block (20/20/20 + 10). It is used only to reconstruct the spikes, and during training we repeatedly remove from it the directions that separate the two stimuli and the two choices.
| 18 sessions | result |
|---|---|
| null block decodes stimulus / choice, time-averaged | 0.50 / 0.50 in every session (chance) |
| the same, pooled by the attention gates | 0.53–0.56 (new sessions), vs 0.60–0.77 for the task blocks |
| decoding, run-to-run agreement, effect of the split | unchanged vs 23/23/24 (General Response 3/5) |
Movement signals still mix into the task blocks: in the 23/23/24 version almost all of the latent's prediction of whisking amplitude (R² 0.41) went through the shared block, and removing licking and whisking from the spikes lowers choice decoding from 0.77 to 0.54. The null block holds mostly task-free activity, but it does not separate movement from the task signals. The paper states this limitation.
Thank you for a sharp review. You suspected the gate, not the partition, does much of the work; that is largely right, and the paper now says so (General Responses 3–5).
W1. The stimulus-private block decodes stimulus 0.747 and choice 0.690: weak selectivity. Agreed. We checked what the split really does, and why it cannot do more.
Paper change: "expected" is withdrawn; the paper reports the full block-by-variable table and where a preference holds.
W2a. Does TSA recover a choice-first order when that is the truth? Yes. On simulated data with the true order reversed, TSA reports stimulus-first on only 0.13 of trials (0.73–0.83 when the truth is stimulus-first). We also swapped the roles of stimulus and choice inside the model (the losses written for stimulus act on choice and vice versa). In a separate simulator (calibrated on all trials rather than training trials only), the recovered order still follows the true order (0.04 stimulus-first when the truth is choice-first, 0.94 when it is stimulus-first), so the order comes from the data, not from the losses. TSA does under-recover the order (about 0.8 where the truth is 1).
W2b / Q4. No comparison with TAME-GP. Added, with both our reimplementation and the authors' official code, tuned on training trials only.
TAME-GP is better at choice and at the order score. That score only asks which peak comes first, so it can be high even when the peaks are in the wrong place; TAME-GP's timing correlation of 0.10 shows it does not locate the events. TSA is much better at stimulus and at per-trial timing. We also corrected our description of TAME-GP (its private/shared split is across brain areas).