- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The paper proposes a discrete generative model of spontaneous spiking activity recorded on high-density microelectrode arrays, where each recording observes only a small and varying subset of the array, with a potentially different subset in each recording. Rather than modeling a fixed set of channels or sorted units, the authors place every recording on the full array footprint, so that recordings with different observed subsets share one common spatial canvas.
The model is "discrete" in the sense that a VQ-VAE maps patches of activity to a finite set of symbols, and a masked transformer prior (MaskGIT) generates sequences of these symbols. The idea is that the symbols correspond to spatiotemporal activity motifs, which the authors claim are shared, at least in part, across recordings and datasets.
The structure of the paper and the lack of a problem formalization make the contributions hard to identify (even if some contributions are stated explicitely by the authors, I'm focusing on the contributions that I am really able to identify in the manuscript), in my opinion, but I would describe them as twofold: (1) a discrete generative model adapted to extremely sparse array-wide spike volumes, combining a residual VQ-VAE with a blank token for empty patches and a prior factorized into where activity occurs and which motif appears there, applied to organoid and ex vivo hippocampal recordings; (2) evidence that the learned motifs are reused across recordings and across the two tissue types, obtained with a single shared model that has no learned per-recording parameters. To support (1), the authors report better reconstruction and completion accuracy than a flat-codebook MaskGIT baseline, while acknowledging that a static per-recording site map and supervised inpainters perform better on completion. To support (2), they report that recording identity explains only 9% of motif usage entropy and that motif overlap across tissue types is comparable to overlap within them, evaluated on held-out segments of the training recordings.
Critical strengths
-
The problem setting is relevant and under-explored: modeling array-wide activity across recordings whose observed electrodes differ.
-
The idea of encoding population activity as a shared motif vocabulary is appealing and I think could lead to new kind of latent structure/interpretability/understanding of neuronal computation.
-
The reporting is transparent. The paper discloses negative results (simple baselines beating the method), and provides the associated code.
Critical weaknesses
I found the paper difficult to follow: the presentation is dense, key definitions are scattered across the main text and appendices, and the flow of ideas is often interrupted by forward references and defensive statements. Some of my comments below may therefore reflect misunderstandings rather than actual flaws, and I welcome corrections from the authors. More broadly, I believe the current presentation substantially hinders a thorough technical assessment, and a clearer structure would allow readers to evaluate the contributions on their merits.
-
The problem is never formalized.
- There is no clear probabilistic statement of the task in a standard generative modeling framework. Equation 3 is a good start, but there is no derivation of this generative model and assumptions are not properly discussed.
-
Generalization and scalability are not tested.
- The claim of no per-assay parameters is a design property. The random assay code makes generalization to unseen recordings impossible (but I may be wrong on this, and if this is the case, I’m curious on how the author would do it?)
-
The pipeline is complex without comprehensive ablation.
- Six stages and roughly twenty loss terms and they are not individually ablated; yet they are not classical in the VQ-VAE or the neural generative field.
-
The presentation is hard to follow and defensive.
-
The introduction concedes limitations before stating the problem, uses undefined terms, discusses evaluation before the idea, and points mainly to Sections 5 and 6, making Sections 2 to 4 feel secondary. Section 2 mixes baseline justification, method claims, and provenance. It omits the motif-detection literature (seqNMF, PPSeq, SPADE, Cell Assembly Detection).
-
Figure 1 is a cluttered block diagram whose caption does not explain its encoding (e.g., the blue boxes). Figure 2 panel A message is not clear and the panel B right part is under-discussed.
-
-
Limited accessibility for a general machine learning audience.
-
The paper assumes substantial familiarity with HD-MEAs and neural recording practice. Key concepts such as how the array works, why only about 1,000 of 26,400 electrodes can be read at once, what electrode routing means, what spike sorting produces, and why recordings differ in their observed electrodes are mentioned only briefly or scattered across Sections 1, 3, and the appendices. Yet these points motivate the whole approach. For an ICLR audience, a short self-contained primer would make the problem, and the relevance of the proposed design, much easier to grasp. It could cover the recording hardware, the path from voltage to binary spike volumes, and the two tissue preparations, ideally with a simple illustrative figure. In its current form, the paper fit with a general machine learning venue is weakened by this missing background.
-
The experiments do not establish why the method should be preferred over alternatives. On the paper own completion tasks, a static per-assay site map, which carries no clip-specific information, outperforms all learned models in every setting. The directly supervised convolutional baselines also outperform the proposed method on most completion settings. The only learned baseline the method consistently beats is MaskGIT-flat, which the authors constructed themselves and which differs from the proposed model in several components at once (tokenizer architecture, codebook structure, blank-token routing, and auxiliary losses). As a result, none of the proposed design choices is isolated by a controlled comparison, and it remains unclear which of them, if any, drives the observed gains. The authors argue that their contribution is a reusable discrete representation rather than imputation accuracy, but the usefulness of this representation is never demonstrated.
Questions for authors
I find the core idea genuinely interesting: representing array-wide activity from recordings with different observed electrodes as a shared vocabulary of motifs could open new ways to compare preparations and study population structure. However, I believe the paper in its current form would require modifications too extensive to be addressed within a rebuttal for me to consider it publishable in the main track. In particular, the formalization needs to be deeply reworked: the generative task, what distribution is being modeled, and how the conditioning relates to the observed data should be stated explicitly before the method.
Formal task. Can you state the target distribution at the voxel level, for example p(X_hidden | X_visible, g_r), and explain how the token-level factorization of eq. 3 and the decoder together define it ? Is the decoder deterministic or sampled?
Usefulness of the representation. A downstream demonstration would substantially strengthen the claim that the motif vocabulary is useful. Examples include predicting tissue type or recording condition from motif usage, or comparing the learned motifs with those found by established methods such as seqNMF, PPSeq or SPADE.
Conditioning on the hidden region. The local descriptor l(X) is computed from the full clip, hidden region included, and its first feature is log mean firing density. How do completion accuracy and the generation families A to D change when l is computed from the visible region only, or omitted? Without this, it is unclear how much of the reported performance comes from information about the answer.
Per-recording code. With 31 fixed random codes and a learned mapper, how does your model differ in practice from one with a learned embedding table? What concrete benefit does the absence of per-recording parameters provide, given that the model cannot be applied to an unseen recording?