- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes Neural Predictive Coding (NPC), a self-supervised objective for binned spiking activity that follows the contrastive predictive coding template. A per bin encoder and a causal GRU over a short history window (300 ms, or 500 ms for navigation) produce a 128 dimensional state, and horizon specific affine heads contrast the true future population embedding at 100, 200, and 400 ms against negatives drawn from temporally separated bins of the same sequence and from other sequences matched on total population count. Frozen states are evaluated across 26 recordings covering motor control (macaque), navigation (rat CA1), vision (mouse, Allen Visual Behavior Neuropixels), and interval timing (macaque). The evaluation has three endpoints: Poisson forecasts of neurons withheld from pretraining, task probes under label budgets, and fitted state evolution, where per horizon affine transitions in a 16 dimensional projection are read out through a fixed emission map. Additional studies cover adaptation to source disjoint mice, common trunk objective controls on five recordings, and trained versus random encoder controls on seven recordings. Comparators are a causal adaptation of CEBRA-Time and NDT2 trained from scratch with a local causal adapter. NPC generally exceeds CEBRA-Time, results against NDT2 are mixed (better neural forecasting in vision and timing, worse in motor and navigation, and worse task decoding in three of four functions), fitted transitions improve over static readouts, and pretraining beats a 250 update scratch control but loses to a longer scratch control from 10 percent query coverage onward. The objective controls suggest that the true past future pairing and multiple horizons carry most of the benefit, while the structured negative sampler contributes little. The authors frame all within recording comparisons as exploratory, and no main benchmark comparison reaches statistical significance.
Critical strengths
1.Careful multi endpoint evaluation with unusually candid reporting. The paper separately tests three properties of a frozen representation (forecasting of withheld neurons, task accessibility under label budgets, and readout fixed state evolution) across four brain functions, three species, and 26 recordings, and it adds mechanism oriented controls (shuffled futures, direct future count prediction, temporal contrast, trained versus random encoders). Post hoc decisions, a failed pre specified transfer criterion, and unfavorable comparisons are all disclosed rather than hidden. This matters because it makes the mixed and negative findings credible, and the evaluation protocol itself is a reusable contribution for assessing self supervised objectives on neural population data.
2.Two findings with real value for the field. First, in the common trunk diagnostic, contrasting true futures beats shuffled future training and direct Poisson prediction of future counts on both endpoints in all five recordings (Tables 2 and 7), which identifies the past future association itself, rather than architecture or added features, as the source of utility. Second, predictive training can reduce task accessibility: trained NPC decodes image identity worse than its own random initialization (Table 5) while forecasting neural activity better (Table 6), and random NDT2 encoders already decode image identity at 0.94 and above. This is a concrete and non obvious caution for objective design in neural foundation models.
Critical weaknesses
1.Limited novelty and no consistent advantage over the strongest baseline. NPC maps almost one to one onto CPC (van den Oord et al., 2018): per step encoder, GRU context, step specific linear predictors, and a contrastive loss. Its main neural specific element, the count matched and within sequence negative sampler, adds essentially nothing to forecasting relative to global negatives (plus 0.003 bits per spike, 2 of 5 recordings; Table 2). Against NDT2, the only strong comparator, NPC is lower on 400 ms forecasting in motor and navigation and lower on task decoding in three of four functions, with a large gap in vision (0.62 versus roughly 0.96 balanced accuracy; Table 4). The CEBRA-Time comparator is a local causal adaptation whose tuning is not described, so wins against it carry limited weight. Expected baselines are absent: DCA (explicitly omitted) or nonlinear predictive information methods, a Poisson future count forecasting model in the main benchmark (it appears only in the five recording diagnostic), raw history decoders for the task probes, and Neural Latents Benchmark standard metrics even though two of its datasets already appear in Appendix E.
2.The evidence is exploratory by construction, and several key analyses have design problems, so method level conclusions are not supported. (a) There is no independent evaluation: within recording comparisons use validation splits that informed result driven tuning applied only to NPC (the motor encoder variant and the navigation checkpoint and learning rate schedule; Appendix B), and the transfer cohort had been inspected before (Appendix G). (b) The significance tests cannot detect effects by design. With n equals 4 animals, an exact two sided paired sign flip test has a minimum attainable p of 0.125, so no four animal contrast can reach 0.05 even before Holm correction, and with n equals 6 the minimum of 0.031 cannot survive Holm over four or eight tests. The null results in Figure 2b, Table 4, and Section 5.3 are therefore guaranteed and carry no information. The single significant result (Section 5.4, p equals 0.0469) is exactly the floor attainable when all eight mice agree, in a post hoc emphasized condition. (c) The incremental readout test behind the first claim of the abstract appears mis specified: in Table 6 the history plus state readout scores below the state only readout in 11 of 14 cells, and seven increments are negative, which suggests the shared regularization is dominated by the high dimensional history block rather than measuring value beyond history. (d) The static reference in the evolution analysis assumes the future equals the present; NPC's static vision forecast is negative (minus 0.175 bits per spike), so the 0.434 gain and the 60 percent advantage over CEBRA-Time largely reflect a degenerate reference and the differing intrinsic timescales of the two representations, and the per horizon transitions are fitted independently, so dynamical consistency is untested. (e) The transfer control is weak: a 5,250 update scratch model wins from 10 percent query coverage onward, the pre specified 10 and 25 percent criterion failed, and the arm that would show whether pretraining helps at convergence (longer adaptation or full fine tuning) is missing.
Questions for authors
1.In the incremental readout, are history and state features penalized by a single shared regularization strength? Please report state only, history only, and combined readouts for the main benchmark with block specific penalties. If NPC alone clearly beats history alone and the combined model at least matches both, the forecasting claim becomes considerably stronger.
- Can you freeze all configurations and evaluate on recordings not used in development, with a comparable tuning budget for the baselines? Additional Allen Visual Behavior mice would also allow the sign flip tests to reach significance in principle.
3.Please add: DCA or a nonlinear predictive information baseline; the Poisson future count GRU in the main benchmark, matched in wall clock time or trained to convergence given the 20 to 29 fold training time difference reported in Appendix E; raw history decoders for the task probes; and NDT2 with a 128 dimensional projection to control state dimensionality.
4.Under a common constant rate reference, how does the evolved forecast compare with a direct horizon specific readout from the same 16 dimensional projection, which upper bounds any linear transition? Do fitted transitions compose, that is, does the 200 ms transition approximate the 100 ms transition applied twice? Without these, it is unclear what evolved versus static establishes beyond differences in representation timescale.
5.Please add transfer arms with 5,250 adaptation updates and with full fine tuning, and a variant where negatives and rate strata are restricted to the same fraction, so that the study can speak to data efficiency rather than only restricted query sampling.
The paper is a careful, honest, and labor intensive empirical study, and its evaluation protocol and controls are genuinely useful. However, the method is an established idea transplanted to spiking data with its main novel component shown to contribute little, there is no consistent or statistically supported advantage over the strongest baseline, all evidence is exploratory with result driven tuning applied only to the proposed method, and the transfer result rests on a post hoc emphasis after the pre specified criterion failed. For the applications to neuroscience and cognitive science area, the paper also delivers no new neuroscientific finding. It is above clear rejection because of its rigor and transparency, and I would raise my score if a confirmatory evaluation on independent cohorts with frozen configurations showed NPC matching or beating strong baselines on neural forecasting with the readout issues fixed.
5 comments
Response to Reviewer nKdY (2/4)
Q1: Does NPC add predictive value after correcting the shared readout penalty?
A: Combining NPC’s learned state with recent spike-count summaries improves mean forecasting scores over either alone in all four brain functions. We replaced the shared penalty with separately calibrated penalties for history and state. We refitted all 234 model–recording–seed combinations with frozen encoders, matched observations and targets, and inner-only selection that also permits either block to be excluded.
The additional 400-ms scores are in bits/spike. H denotes the specified recent-count summary, Z the learned state, and “Selected” the inner-selected standalone or combined readout.
| Function | Model | H | Z | H+Z | Selected |
|---|---|---|---|---|---|
| Motor | NPC | 0.0592 | 0.0783 | 0.0815 | 0.0815 |
| Motor | CEBRA-Time | 0.0592 | 0.0556 | 0.0695 | 0.0695 |
| Motor | NDT2 | 0.0592 | 0.0805 | 0.0818 | 0.0818 |
| Navigation | NPC | 0.1699 | 0.2312 | 0.2332 | 0.2330 |
| Navigation | CEBRA-Time | 0.1699 | 0.2187 | 0.2344 | 0.2344 |
| Navigation | NDT2 | 0.1699 | 0.2553 | 0.2559 | 0.2549 |
| Vision | NPC | 0.2395 | 0.3034 | 0.3218 | 0.3218 |
| Vision | CEBRA-Time | 0.2395 | 0.1532 | 0.2600 | 0.2600 |
| Vision | NDT2 | 0.2395 | 0.2420 | 0.2848 | 0.2848 |
| Timing | NPC | 0.2251 | 0.2960 | 0.2992 | 0.2988 |
| Timing | CEBRA-Time | 0.2251 | 0.2914 | 0.3035 | 0.3021 |
| Timing | NDT2 | 0.2251 | 0.2608 | 0.2630 | 0.2621 |
The gains hold across all four function means. NPC’s increments over H / Z are +0.0223 / +0.0033 (motor), +0.0633 / +0.0021 (navigation), +0.0823 / +0.0184 (vision), and +0.0742 / +0.0032 (timing), calculated before rounding. Combined-below-state cases decrease from 52/78 to 15/78; four combined-below-history cases remain. These complete results distinguish function-level gains from individual outcomes.
Selected fits passed convergence checks; some unselected candidates did not, and penalty-boundary selections are retained in the audit. All methods use uniform fresh fits because archived scores were not uniformly reproduced. The historical fourteen-cell random-feature screen remains separate and unrepaired.
Q2: Is there an independent evaluation with frozen configurations?
A: We have established an independent-cohort evaluation with configurations fixed for NPC and both baselines. The protocol specifies 12 previously unused mice, matched training-update budgets, and selection using inner calibration only. Recorded coverage at the cutoff is 12 of 36 model–recording evaluations. The matching applies to update counts.
Response to Reviewer nKdY (3/4)
Q3a: How does NPC compare with a predictive-information baseline?
A: NPC exceeds DCA’s state-only and combined forecasting means on the six visual development mice. The additional comparison with DCA [5] uses common targets, anchors and independently regularized readouts, in bits/spike:
| Model | H | Z | H+Z / selected |
|---|---|---|---|
| Original frozen NPC | 0.2395 | 0.3034 | 0.3218 |
| DCA | 0.2395 | 0.2883 | 0.2970 |
NPC uses a 128-dimensional state; DCA uses 240 projection-history features. NPC scores average three training seeds, and training budgets differ.
Q3b: How does NPC compare with count prediction at matched training time?
A: NPC’s combined readout exceeds both standalone readouts in all six mice at matched training time. Both methods start from the same encoder weights and share observations and measured training-step time on the same GPU; calibration time is excluded. This additional visual comparison uses one training seed.
| Model | H | Z | H+Z / selected |
|---|---|---|---|
| Fresh NPC | 0.2395 | 0.2832 | 0.3061 |
| Future-count GRU | 0.2395 | 0.3013 | 0.3120 |
NPC exceeds count prediction in three mice; the count GRU has the higher cohort mean, 0.3120 versus 0.3061. These are fixed-time results, with convergence untested. Fresh NPC’s 0.3061 and the original frozen model’s 0.3218 come from separate training runs; the submitted five-recording comparison retains its fixed-update scope.
Q3c: What do raw-history task decoders reveal?
A: The raw-history control supports separate evaluation of neural forecasting and task decoding. All six visual mice are complete. Linear probes use the same ordered 300-ms observations and labeled anchors; NPC/random scores average three seeds within mouse. Balanced accuracy is:
| Label budget | Raw history | NPC | Random NPC |
|---|---|---|---|
| 1% | 0.8358 | 0.3182 | 0.2766 |
| 5% | 0.9877 | 0.4435 | 0.4927 |
| 10% | 0.9926 | 0.4914 | 0.5498 |
| 25% | 0.9939 | 0.5272 | 0.5912 |
| 100% | 0.9950 | 0.5496 | 0.6290 |
Raw history has the higher mean at every budget, and trained NPC is below random NPC in five mice at full labels. Together with the submitted trained/random forecasting controls, this shows how predictive training changes linear task accessibility. These probes differ from the submission’s inner-selected probe families; small label budgets can omit classes. The result concerns linear accessibility, with other brain functions still untested by this control.
Q3d: Does NPC’s advantage persist when NDT2’s readout dimension is matched?
A: NPC retains higher visual and timing forecasting means when NDT2 is projected to 128 dimensions. The additional state-only comparison is:
| Function | NDT2 PCA256 | NDT2 PCA128 | NPC native128 |
|---|---|---|---|
| Motor | 0.080471 | 0.080319 | 0.078288 |
| Navigation | 0.255260 | 0.254826 | 0.231167 |
| Vision | 0.241988 | 0.240015 | 0.303389 |
| Timing | 0.260806 | 0.260388 | 0.296013 |
The function-level ordering is preserved: NPC leads in vision and timing; NDT2 leads in motor and navigation. This matches readout dimension, while encoder capacities differ. The numerical audit records tolerance exceedances and two flagged fits per arm. The largest final-score discrepancy is 0.0000184 bits/spike.
Response to Reviewer nKdY (1/4)
Thank you for highlighting the careful evaluation and the findings on true future pairing and task accessibility. We address i) contribution and scientific value, ii) strength of evidence, iii) readout and baseline controls, and iv) dynamics and transfer.
W1: What does NPC contribute, and where does it offer an advantage?
A: We contribute a compact predictive model for neural forecasting and task decoding, a reusable evaluation framework across four brain functions, and controlled evidence showing what predictive training adds. Building on CPC [1], these contributions are:
i) NPC uses one compact representation for both neural forecasting and task decoding. Within each recording, NPC learns to predict future population representations. We then freeze the encoder and use its state to forecast activity in neurons excluded from pretraining and decode task variables. The encoder stays fixed as the readout changes.
ii) Our evaluation compares forecasting, task decoding and fitted state evolution under matched input conditions across four brain functions. It covers 26 recordings and 18 animals spanning motor, navigation, vision, and timing. NPC achieves higher task-decoding means than CEBRA-Time [2] in all four functions and higher forecasting means in motor and vision. It exceeds NDT2 [3] in vision and timing forecasting. Visual NPC uses 0.11–0.12 million parameters versus approximately 3.32 million for the evaluated NDT2 model. After correcting readout regularization, NPC also adds forecasting value beyond recent-count summaries in all four function means (Q1).
iii) Matched controls show that true future pairing and multiple prediction horizons improve neural forecasting. Trained NPC forecasts better than its random initialization in all seven screened recordings. True future pairing improves forecasting and decoding in all five diagnostic recordings, and multiple prediction horizons improve forecasting in all five. The visual controls also show that forecasting gains and linear task accessibility respond differently to training, giving a concrete guide for choosing neural learning objectives.
Structured negatives improve forecasting in two and decoding in four of the five diagnostic recordings. The baseline comparisons use causal CEBRA-Time and locally adapted, scratch-trained NDT2; broader pretrained-model and standardized NLB [4] comparisons remain open.
W2: What do the development results and statistical tests establish?
A: We report animal-level effects and distinguish development evidence from independent confirmation. Seeds and sessions are averaged within animals. With four/six animals, the minimum two-sided exact p-values are 0.1250/0.0313, versus 0.0063 at the first step of each eight-contrast Holm family. Adjusted p-values for the primary forecasting and task comparisons range from 0.25 to 1.00. NPC received additional motor/navigation development tuning; the frozen-cohort follow-up is reported in Q2. The historical transfer p = 0.0469 is six-fraction Holm-adjusted, with a raw minimum of 0.0078, and retains its post-hoc interpretation.