{"id":"10be4738-df91-4c1f-8a11-00b1a26c1bcf","arxiv_id":"2501.09684","paper_version":1,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A perspective arguing that systems neuroscience should pursue a brain-wide functional architecture through multi-modal, interpretable data mining, but contains no new data or methods.","lead":"This paper is a perspective, not a data study, arguing that new brain-wide recording tools should push neuroscience from local circuits to a global 'functional architecture' spanning areas, tasks, and modalities. It inventories the technical hurdles around modular dynamics, multi-dataset alignment, interpretability, and multi-modal fusion, and calls for new mathematical models to meet them.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4 concedes that nonlinear latent dynamics are unidentifiable up to invertible transformation; this undercuts the claim that new data can reveal the brain's functional architecture, since modular structure may be an arbitrary reparameterization. No resolution is proposed.","rationale":"The reader's weakest assumption correctly noted that Section 4 concedes nonlinear latent dynamics are unidentifiable up to an invertible transformation, but it bundled this with a separate concern about dataset breadth. I focus on the identifiability issue alone because it is more fundamental: even with idealized, perfectly balanced multi-task and multi-modal data, the invertible transformation symmetry persists, so the latent modular structure cannot be uniquely recovered from observational recordings. The paper acknowledges this but does not propose how to break the symmetry, leaving the central opportunity claim without a key justification. This is load-bearing because the functional architecture is defined in terms of latent variables and their dynamics; if those latents are only identifiable up to an arbitrary transformation, then 'finding' the architecture is not well-posed without extra constraints. The paper's appeals to sparsity, geometry, or independent representations are not sufficient by themselves, as nonlinear ICA counterexamples show. Given that this is a perspective piece rather than a research result, the appropriate verdict remains UNVERDICTED—the paper's argument is coherent but its central claim is not demonstrated, and the identified unidentifiability reinforces the reader's cautious assessment. No author-level judgment is implied; the concern is purely about the logical support for the claim.","tokens_in":7501,"tokens_out":4960,"duration_ms":51400,"concrete_test":"Simulate a two-module latent system: z_t = (z^1_t, z^2_t) with linear dynamics g block-diagonal, and a nonlinear mixing Φ(z) (e.g., an MLP). Generate observations x_t. Then choose a smooth invertible h that mixes the two latent coordinates (e.g., h(z) = (z^1 + ε z^2, z^2) with mild coupling) and construct the transformed dynamics g̃ = h∘g∘h^{-1} and emission Φ̃ = Φ∘h^{-1}. Verify that the generated observations x_t are identical (up to numerical precision) while the transformed system has non-block-diagonal dynamics. This directly demonstrates the Section 4 ambiguity in a simple, reproducible case and shows that data alone cannot determine the functional architecture without extra constraints.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 states that for nonlinear dynamics z_t = g(z_{t-1}) and nonlinear emissions x_t = Φ(z_t), any invertible h yields another solution (Φ∘h^{-1}, h∘g∘h^{-1}, h(z_t)) with identical observations. This means the latent variables—and hence any functional architecture defined on them—are determined only up to an arbitrary invertible transformation. If the architecture is the modular structure of the latent dynamics, the same data are consistent with infinitely many architectures differing in the number of modules, their interaction graph, and their boundaries. The paper's central claim is that new multi-area, multi-task, multi-modal data will allow discovery of this architecture, but it does not show that adding modalities, tasks, or brain regions removes the h-induced equivalence. It only lists identifiability as a challenge and vaguely appeals to sparsity and geometry; nonlinear sparsity and independence are well known not to guarantee identifiability in general. Thus, even with perfect, unlimited observational recordings, the functional architecture is not uniquely recoverable absent additional assumptions or interventions. The claim that the field is 'approaching the data needed to find the functional architecture' is therefore unsupported: the data may be necessary, but they are not sufficient. This is the single most load-bearing weakness because the entire opportunity statement depends on data-driven discovery being feasible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a perspective/position paper arguing that advances in large-scale neural recording, behavioral monitoring, and multi-modal imaging have created an opportunity to move from 'local' systems neuroscience to a 'global' account of the brain's functional architecture. The author proposes that by synthesizing data across brain areas, tasks, and modalities, and by using interpretable mathematical models with explicit modular or latent-state structure, the field can discover the distributed subsystems of brain computation and their interactions. The paper surveys relevant modeling approaches (e.g., dLDS, manifold methods, shared/private latent-variable models, multi-view fusion) and lists challenges: data alignment, task diversity, interpretability versus expressivity, and the non-identifiability of nonlinear latent dynamics. It concludes with a call for interpretable AI, data geometry, sparsity, and independent representations as guiding themes.","tokens_in":7766,"tokens_out":3857,"duration_ms":44520,"significance":"If the central claim is correct, the paper identifies a timely and practically important research direction: moving from anatomy-centric descriptions to data-driven maps of functional circuits. The manuscript has the merit of naming concrete technical hurdles—cross-session and cross-animal alignment, shared-versus-private latent structure, and the expressivity/interpretability trade-off—and of pointing to a set of recent methods as initial steps. It is also candid in explicitly acknowledging the non-identifiability of nonlinear latent dynamics, which many similar programmatic papers omit. However, the paper is an opinion piece without new data, derivations, or simulations, and its central assertion that current or imminent datasets will allow discovery of the functional architecture is not backed by a formal argument. The significance therefore rests on whether the identified obstacles can be resolved, especially the identifiability problem, which the manuscript does not address beyond listing it as a challenge.","major_comments":[{"comment":"The paper's central claim that new multi-area, multi-task, multi-modal data will allow discovery of the brain's functional architecture is not reconciled with the non-identifiability result stated in Section 4. There, the author correctly notes that for nonlinear latent dynamics z_t = g(z_{t-1}) and nonlinear emissions x_t = Φ(z_t), any invertible h yields an equally valid solution (Φ∘h, h^{-1}∘g∘h, h^{-1}(z_t)). Since the proposed 'functional architecture' is defined on latent states, the number of modules, their boundaries, and their interaction graph are all invariant under this reparameterization. The paper acknowledges that 'theoretical advances are needed' but does not explain what those advances would be; the concluding appeal to sparsity, independence, and geometry is not connected to any identifiability theorem. Without either restricting the model class (e.g., identifiable nonlinear ICA with auxiliary variables), invoking interventions or perturbations, or explicitly reframing the goal as discovering an equivalence class of architectures, the abstract's claim that the data are 'needed to find' the architecture overstates what observational recordings alone can deliver.","section":"Section 4"},{"comment":"The argument that synthesizing data across tasks will differentiate between co-active systems is presented only through an intuitive example (reaction task versus decision task). The manuscript does not state what conditions on the task set, behavioral coverage, or neural-population overlap would make such differentiation possible, nor does it show that any existing or planned dataset satisfies those conditions. As written, the claim that multi-task data will reveal shared versus private systems is an assertion, not a demonstrated research conclusion. The authors should either temper the claim or provide concrete criteria—for example, the number and type of tasks needed to decorrelate system engagement—so that the opportunity statement is falsifiable.","section":"Section 3"}],"minor_comments":[{"comment":"Typographical errors: 'moleclular' should be 'molecular', and 'inmodality' should be 'in modality'.","section":"Abstract"},{"comment":"The notation for the transformed latent variable is confusing: 'gz(t) = h^{-1}(z(t))' should be written as a new latent variable, e.g., \\tilde z(t) = h^{-1}(z(t)), with corresponding definitions \\tilde Φ = Φ∘h and \\tilde g = h^{-1}∘g∘h. As printed, 'gz(t)' looks like another dynamical map rather than the transformed state.","section":"Section 4"},{"comment":"In the sentence defining multi-modal readouts, 'x2 = f2(x2)' should read 'x2 = f2(z)'.","section":"Section 5"},{"comment":"'Sythesizing' should be 'Synthesizing'.","section":"Section 5"},{"comment":"The paper uses 'functional architecture' as a central term but never defines it operationally. Even a working definition—e.g., a partition of latent variables into modules with specified interactions—would help the reader understand what kind of evidence would confirm or refute a proposed architecture.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"This is a perspective paper, so I evaluated it on the clarity and force of its argument rather than on novel technical results. The main concern is that the non-identifiability issue in Section 4 undercuts the paper's headline claim, and the revision needs to address it substantively rather than merely list it as an open challenge. I also note that a substantial fraction of the cited methodological references (dLDS, GRAFT, SIBBLINGS, CREIMBO, and others) are by the author and collaborators; this is not inappropriate for a perspective describing the author's research program, but the editor may wish to ensure the presentation does not read as primarily an advertisement for those methods."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is a perspective, not a research preprint. There are no new equations, datasets, or algorithms. What is new is the packaging: a clear agenda for moving systems neuroscience from local, single-task, single-modality analyses to brain-wide, multi-task, multi-modal synthesis in order to discover the brain's 'functional architecture.' The paper does a genuinely good job of structuring the challenges — modular dynamics, cross-task and cross-animal alignment, multi-modal fusion, interpretability — and of pointing to concrete existing ingredients (dLDS, GRAFT, SIBBLINGS, shared/private latent models, butterfly architectures). It is honest enough to state in Section 4 that nonlinear latent dynamics are identifiable only up to an invertible transformation. That honesty is a strength.\n\nThe soft spots are real but typical of the genre. The central claim — that the field is 'approaching the data needed' to find the architecture — is asserted, not demonstrated. The paper never shows that current datasets span enough conditions to separate shared from private systems, nor that adding modalities or tasks removes the h-induced unidentifiability it names. The stress-test note is right that the identifiability gap is load-bearing for the strong version of the thesis: if modular structure is just a reparameterization of the latent variables, the same data can instantiate many architectures. The paper lists this as a challenge and moves on; a sharper version would discuss what assumptions (sparsity? independence? interventions?) could pin the architecture down, or at least argue why those assumptions are biologically plausible. The citation practice is heavy on the author's own work, but those are the relevant methods in this area, so I do not read it as a flaw.\n\nWho is this for? Someone writing a grant or a review on brain-wide recording analysis will find it a useful framing document. It is not a methods paper and should not be cited as one. I would bring it to a reading group as a discussion prompt, not as a technical anchor.\n\nRecommendation: send it to peer review as a perspective/opinion piece. A serious referee should ask for a more explicit treatment of identifiability and a more concrete proposal for tests or benchmarks. With revisions along those lines it could be a genuinely useful agenda piece, though it will never be a result.","headline":"A coherent perspective that sells a brain-wide functional-architecture agenda, but the central promise is asserted rather than demonstrated, and the paper's own identifiability concession is never squared with that promise.","tokens_in":8248,"tokens_out":2871,"would_cite":false,"duration_ms":29708,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"New brain-wide recordings finally make the functional architecture of the brain discoverable.","keywords":["functional architecture","systems neuroscience","neural data mining","latent variable models","multi-modal data fusion","brain-wide recordings","interpretability","modular dynamics"],"falsifier":"Collect a large-scale dataset that combines calcium or voltage imaging across several brain areas with detailed behavior in the same animals across multiple distinct tasks, then fit a shared/private latent-dynamics model; if the inferred modules do not stabilize across task sets or cannot be aligned across animals, the central premise that the functional architecture is discoverable from current observational data fails.","tokens_in":7311,"feed_emoji":"🧠","tokens_out":8052,"duration_ms":79031,"temperature":0.7,"pith_summary":"Systems neuroscience has been local by necessity: small populations, single tasks, single brain areas, single imaging modalities. The paper argues that new recording and behavioral-monitoring technologies now provide the data to move to global neuroscience, in which the brain is treated as an interconnected set of distributed subsystems. Its central thesis is that the field is finally approaching the data needed to find the functional architecture of the brain—the 'software' running on the anatomical 'hardware'—and that doing so requires interpretable mathematical models that synthesize data across areas, tasks, and modalities. The payoff would be a roadmap of the general principles behind the brain's flexibility, robustness, and efficiency, and baselines for understanding neurodegenerative and psychiatric disorders.","feed_headline":"Data now exist to map the brain's functional software","feed_subtitle":"That would let us trace how distributed circuits produce behavior and how disease disrupts them.","key_machinery":"The load-bearing machinery is the latent-variable modeling framework for multi-dataset neural data. In this view, observable activity $x(t)$ is generated from low-dimensional latents $z(t)$ through a (possibly nonlinear) readout $x(t)=\\Phi(z(t))$, and the goal is to decompose the latent space into modular operators and into shared versus private components across views. One concrete instance, decomposed linear dynamical systems (dLDS), represents dynamics as paths on a manifold whose tangent spaces are spanned by a sparse set of linear operators $\\{G_k\\}$, each operator corresponding to a distinct interaction module. The same shared/private separation underlies multi-modal fusion, where the latent state splits as $z \\to \\{z_s, z_1, z_2\\}$. The paper uses this framework to state both the promise and the challenge: modularity must be built into the model, and without it, nonlinear latent dynamics are unidentifiable up to invertible transformations.","core_discovery":"The paper's central claim is that the brain's functional architecture can be discovered from data that now exist. While anatomically defined regions provide a hardware map, functional architecture refers to where and how information spreads and transforms across widely distributed circuits; the paper cites evidence that activity is 'everywhere' and that no strict anatomical boundary predicts function. To find that architecture, the author argues, models must explicitly learn modular dynamics rather than treating all recorded units as one shared state space, and must separate shared from private information when fusing data from different tasks, animals, and modalities. The paper also flags the central mathematical obstacle, in Section 4: when latent dynamics are nonlinear, models are unidentifiable up to an arbitrary invertible transformation, so interpretable discovery requires additional structure or constraints.","pith_inferences":["The paper's own unidentifiability caveat suggests a testable criterion: if a functional architecture is real, the modules recovered under different inductive biases (sparsity, independence, geometry) should converge; if they diverge, the data alone do not pin down the architecture.","A next step the paper leaves implicit is causal validation: inferred modules from observational recordings could be tested with targeted perturbations (e.g., optogenetics or lesions) to see whether they are functionally necessary.","If the multi-task RNN results generalize, training recurrent networks on the same task battery as animals should produce internal modules whose structure parallels the brain's, providing a fast testbed for model interpretability before invasive experiments.","The 'everything is everywhere' evidence implies that a functional atlas based on latent dynamics might generalize across individuals better than an anatomical atlas, which would be a direct test when comparing healthy and diseased populations."],"forward_implications":["Brain-wide recordings across multiple tasks will reveal functional modules that are invisible in single-task studies, because distinct systems are recruited together in one task and separately in another.","Models that explicitly separate shared and private information across modalities will prevent erroneous scientific conclusions about shared brain function caused by information leakage.","Data geometry—curvature and tangent-space structure of neural manifolds—will become a primary language for linking dynamics, behavior, and brain-wide recordings.","Interpretable models, rather than black-box ANN predictors, are required because scientific discovery needs extrapolation beyond the training domain, not just accurate interpolation.","If the program succeeds, disorders such as neurodegeneration and psychiatric disease can be understood as changes in the functional architecture, not just as localized region-specific activity changes."],"supporting_citations":[{"why":"Establishes that advances in neural recording are reshaping how neural data are analyzed.","marker":"[37]"},{"why":"Provides a volumetric two-photon imaging method capable of brain-wide cellular-resolution recording.","marker":"[35]"},{"why":"Reports distributed coding of choice, action, and engagement across the mouse brain, supporting the 'everything is everywhere' view.","marker":"[36]"},{"why":"Introduces decomposed linear dynamical systems, the paper's primary example of modular latent dynamics.","marker":"[25]"},{"why":"Shows that recurrent networks trained on multiple tasks develop shared dynamical motifs, suggesting modules appear only across tasks.","marker":"[13]"},{"why":"Presents a butterfly-architecture method for separating shared and private geometry in multi-view data, addressing information leakage.","marker":"[22]"},{"why":"Demonstrates joint encoding of neural and behavioral time series in a dynamical latent-state model.","marker":"[34]"},{"why":"Provides markerless 3D pose estimation across species and behaviors, enabling multi-behavior monitoring.","marker":"[28]"}],"fun_headline_variants":["Brain's functional software now mappable from data","Global neuroscience: data reveal brain's functional map","From local to global: data unlock brain's functional architecture","Functional brain architecture discoverable with current data","Beyond anatomy: data can trace brain's functional circuits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The program depends on the assumption that functional modules leave stable, separable signatures across the tasks, animals, and recording modalities available today, even though the paper itself concedes that nonlinear latent dynamics can be transformed arbitrarily without changing the data it produces.","fun_headline_variants_meta":{"raw":{"variants":["Brain's functional software now mappable from data","Global neuroscience: data reveal brain's functional map","From local to global: data unlock brain's functional architecture","Functional brain architecture discoverable with current data","Beyond anatomy: data can trace brain's functional circuits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00027,"raw_usage":{"total_tokens":1627,"prompt_tokens":949,"completion_tokens":678,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":604}},"tokens_in":565,"tokens_out":678,"duration_ms":7662,"temperature":1.0,"reasoning_tokens":604,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:45:07.044057+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a large-scale dataset that combines calcium or voltage imaging across several brain areas with detailed behavior in the same animals across multiple distinct tasks, then fit a shared/private latent-dynamics model; if the inferred modules do not stabilize across task sets or cannot be aligned across animals, the central premise that the functional architecture is discoverable from current observational data fails.","supporting_citations":[{"cited_title":"How advances in neural recording affect data analysis","cited_arxiv_id":null,"evidence_quote":"Establishes that advances in neural recording are reshaping how neural data are analyzed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides a volumetric two-photon imaging method capable of brain-wide cellular-resolution recording."},{"cited_title":"Distributed cod- ing of choice, action and engagement across the mouse brain","cited_arxiv_id":null,"evidence_quote":"Reports distributed coding of choice, action, and engagement across the mouse brain, supporting the 'everything is everywhere' view."},{"cited_title":"Decomposed lin- ear dynamical systems (dlds) for learning the latent components of neural dynamics","cited_arxiv_id":null,"evidence_quote":"Introduces decomposed linear dynamical systems, the paper's primary example of modular latent dynamics."},{"cited_title":"Flexible multitask computation in recurrent networks utilizes shared dynamical motifs","cited_arxiv_id":null,"evidence_quote":"Shows that recurrent networks trained on multiple tasks develop shared dynamical motifs, suggesting modules appear only across tasks."},{"cited_title":"Unsupervised discovery of the shared and private geometry in multi-view data","cited_arxiv_id":null,"evidence_quote":"Presents a butterfly-architecture method for separating shared and private geometry in multi-view data, addressing information leakage."},{"cited_title":"Mod- eling behaviorally relevant neural dynamics enabled by preferential subspace identification","cited_arxiv_id":null,"evidence_quote":"Demonstrates joint encoding of neural and behavioral time series in a dynamical latent-state model."},{"cited_title":"Using deeplabcut for 3d markerless pose esti- mation across species and behaviors","cited_arxiv_id":null,"evidence_quote":"Provides markerless 3D pose estimation across species and behaviors, enabling multi-behavior monitoring."}],"review_version":1}