Pith. sign in

REVIEW 3 major objections 5 minor 63 references

NLP4Neuro: Sequence-to-sequence learning for neural population decoding

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read An off-the-shelf mixture-of-experts LLM, fine-tuned briefly, outperforms recurrent and other transformer models at decoding zebrafish tail movements from whole-brain calcium imaging.

desk verdict Useful empirical sweep of off-the-shelf LLMs for zebrafish whole-brain decoding, but the significance tests are invalid and the abstract overstates pre-training benefits. read the letter →

arxiv 2507.02264 v1 pith:I7UXFB4O submitted 2025-07-03 q-bio.NC cs.LG

Jacob J. Morra , Kaitlyn E. Fouke , Kexin Hang , Zichen He , Owen Traubert , Timothy W. Dunn , Eva A. Naumann This is my paper · ORCID
classification q-bio.NCcs.LG
keywords neuraldecodinglargelanguagemodelssequence-to-sequencelearningcalciumimagingzebrafishmixture-of-expertssaliencemappingtransfer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that off-the-shelf large language models, fine-tuned for only a few epochs, can decode behavior from whole-brain calcium imaging better than the recurrent networks traditionally used for neural decoding. In larval zebrafish, the mixture-of-experts transformer DeepSeek-c7b produced the lowest prediction error across most sequence lengths and the highest correlation with measured tail kinematics, and its salience maps pointed to neurons in the pretectum, a region already implicated in visuomotor control. The finding matters because it suggests that text-pretrained transformer weights carry a useful inductive bias for reading neural population activity, and that a modular pipeline of open-weights LLMs could help neuroscientists identify circuit-relevant neurons for later causal tests. The absolute accuracy remains modest, so the paper's case is about relative model comparison and interpretability rather than high-fidelity decoding.

What carries the argument

The central mechanism is sequence-to-sequence learning with a pre-trained transformer, treated as an off-the-shelf token-level map: each time frame's whole-brain fluorescence vector is a token, and the model is fine-tuned to emit s tail predictions for s input frames. DeepSeek-c7b contributes a mixture-of-experts decoder whose routing network computes softmax scores over experts and recruits the top-k experts (k = 2) per token; the paper credits this MoE routing, plus the pre-trained textual weights, for the decoding advantage. Salience mapping then uses input-gradient scores, summed over output frames and averaged over sliding windows, to attribute each neuron's contribution to predicted tail components.

What would settle it

Retrain the LSTM, reservoir computer, GPT-2, and BERT baselines with per-model hyperparameter search on the same held-out fish, and also re-run DeepSeek-c7b with shuffled input tokens; if a tuned baseline matches DeepSeek's RMSE and r, or if shuffled tokens produce the same pretectal salience pattern, the paper's central claims would be undercut.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a 7-billion-parameter mixture-of-experts LLM, DeepSeek-c7b, fine-tuned for 5 to 10 epochs on simultaneous calcium-imaging frames and tail-angle recordings, outperforms an LSTM, a reservoir computer, GPT-2, and BERT at sequence-to-sequence prediction of tail kinematics. For sequence lengths s = 5, 10, 15, 20, DeepSeek-c7b has the lowest root-mean-square error for all but s = 15 and the best Pearson correlation with ground-truth tail sums (r about 0.08 to 0.10); it captures the timing of tail bursts better than their amplitude, and its advantage over the best RNN grows with sequence length. The paper also reports that pre-trained weights learned from natural-language text improve fine-tuned predictions for DeepSeek-c7b and BERT-bu, while relative positional encoding helps other transformers but not DeepSeek-c7b. Finally, input-gradient salience averaged over sliding windows places the most behavior-predictive neurons in the pretectum and shows lateralization consistent with the known optomotor-response circuit.

Load-bearing premise

The central comparison assumes that one fixed fine-tuning recipe—a learning rate of 1e-6, batch size 16, and 5 to 10 epochs—is a fair test across model families; if, say, the LSTM would beat DeepSeek-c7b with its own tuned hyperparameters, the 'best decoder' conclusion does not hold.

Editorial extensions

If this is right

  • Newer open-weights LLMs can be dropped into the same two-stage pipeline without redesign, so neural-decoding benchmarks can track the pace of LLM releases.
  • If the transfer is real, neuroscience groups without large training data can begin decoding experiments from pre-trained text models rather than training recurrent networks from scratch.
  • Salience maps from DeepSeek-c7b nominate pretectal and right-lateralized neurons as candidate controllers of tail behavior, giving photostimulation experiments a shortlist to test.
  • Longer input contexts benefit the MoE model more than the LSTM, so the approach is particularly suited to behavioral readouts that require integrating neural activity over long timescales.
  • Because the model captures tail event timing better than amplitude under RMSE loss, changing the loss to penalize burst amplitude may improve decoding of high-amplitude short events.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Inference: The paper does not isolate whether DeepSeek's advantage comes from the MoE architecture or from its code-heavy pretraining corpus; a same-size MoE pretrained only on general text would separate these.
  • Inference: The salience maps imply a specific causal experiment not run here: photostimulate the top ranked pretectal and lateralized neurons and ask whether the predicted tail direction appears.
  • Inference: Because best test-set r is about 0.10, the practical near-term role for NLP4Neuro is hypothesis generation about circuits rather than high-fidelity neural prostheses; benchmarking on event-triggered epochs is an immediate way to test whether frame-averaged tail sums understate its decoding ability.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes NLP4Neuro, a two-stage pipeline that fine-tunes off-the-shelf pre-trained transformer-based LLMs (BERT-base-uncased, GPT-2, DeepSeek Coder-7b) for sequence-to-sequence decoding of tail kinematics from whole-brain calcium imaging in larval zebrafish, and compares them with two RNN baselines (reservoir computer and LSTM). In Experiment 1, the authors report that DeepSeek-c7b yields the lowest RMSE for sequence lengths s=5, 10, and 20 and the highest Pearson r for all s, with the best RNN (LSTM) competitive at short horizons. Experiment 2 compares pre-trained versus randomly initialized transformers, Experiment 3 evaluates five input embedding strategies, and Experiment 4 applies input-gradient salience mapping to identify neurons in the pretectum as important for tail predictions. The authors conclude that pre-trained LLMs, especially mixture-of-experts models, are effective and interpretable neural decoders.

Significance. If the empirical results survive scrutiny, the paper offers a practically useful and reproducible demonstration that modern MoE LLMs can be fine-tuned for brain-wide neural decoding, and it provides a valuable external sanity check by validating salience maps against known pretectum-to-tail anatomy. The public code and data repositories, as well as the use of off-the-shelf model weights, make the pipeline easy to adopt. However, the headline 'DeepSeek best' rests on small absolute effect sizes (best r ≈ 0.10) and on statistical tests that do not account for the dependence between overlapping sliding windows, while the abstract's pre-training claim is internally inconsistent with the authors' own Table 2. The paper therefore establishes a promising direction rather than a definitive ranking, and the central claims require additional statistical and experimental support.

major comments (3)
  1. [§3.4, Table 1, Fig. 2a] The test sequences are sliding windows of length s with stride one, so adjacent windows overlap by s−1 frames. The per-sequence RMSE values used in the Mann-Whitney U tests and the reported standard errors are therefore not independent, which inflates the effective sample size and makes the p<0.05 stars anti-conservative. Because the RMSE differences are small (e.g., 0.052 vs 0.056 at s=20), the significance asterisks, not the raw means, are doing the evidentiary work for the claim that DeepSeek-c7b is 'significantly improved.' Please provide cluster-level inference (e.g., clustering by fish or using non-overlapping windows) and report effective sample sizes.
  2. [Abstract, §4.2, Table 2] The abstract states that 'LLMs become better at neural decoding when they use pre-trained weights learned from textual natural language data.' This claim is contradicted by Table 2: for GPT-2 at s=20 the untrained model has lower RMSE (0.052 vs 0.058) and much higher Pearson r (0.17 vs 0.04), and for DeepSeek-c7b the untrained model has higher r at both s=5 and s=20 (0.11 vs 0.10 and 0.12 vs 0.09). The text in §4.2 acknowledges the GPT-2 RMSE reversal but does not reconcile the r reversals. Please qualify the pre-training claim (e.g., to RMSE and to specific models) or re-analyze the data.
  3. [§3.4, Table 1] All models are compared under a single fine-tuning protocol (learning rate 1e-6 for Experiment 1, 5–10 epochs, fixed batch sizes), and the RC and LSTM hyperparameters (reservoir size, leaking rate, input scaling, regularization, hidden size, and hidden layers) are not reported in the main text or appendix. Without evidence that the smaller baselines are reasonably tuned, the conclusion that DeepSeek-c7b is 'best' could be an artifact of under-tuned competitors. Please report the full hyperparameter settings for all models and include a small sensitivity analysis or hyperparameter search for the LSTM and RC.
minor comments (5)
  1. [§3.4] The text says 'The sum of RMSE values across all sequences is the metric reported,' but Table 1 reports values around 0.05 that must be means, not sums; please clarify whether the mean or the sum is used and define the aggregation consistently.
  2. [Appendix A.6, Eq. (14)] In Eq. (14), the bias term for the second expert projection is written as b_ei^(1) in both places; the second bias should presumably be b_ei^(2).
  3. [Table 2] The dagger symbol in Table 2 indicates a Wilcoxon signed-rank test, but the caption does not describe the test, the number of fish, or how the paired structure was defined; please describe the test and state n.
  4. [Fig. 2c] The caption should explicitly state that the reservoir computer prediction is shown only in the Δθ0 panel, since the reader may otherwise expect to see it in all panels.
  5. [§4.4, Fig. 4c] The Pt-vs-Hb salience comparison uses a Student's t-test with only three fish; please report effect sizes and consider a non-parametric or paired test, and state whether the comparison is across the three fish or across neurons.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the evaluation is a held-out empirical comparison, and the paper's self-citations are external biological validations rather than inputs to the derivation.

full rationale

The central claim that DeepSeek-c7b best decodes tail kinematics is supported by held-out test-set RMSE and Pearson r (Table 1; Eq. 15), with a 70:15:15 train/validate/test split (Section 3.4). No parameter is fitted to the test metric and then reported as a prediction, so the fitted-input-called-prediction pattern does not apply. The pre-trained versus untrained comparison (Experiment 2) is a controlled ablation: architecture is held fixed and only weights differ, so the conclusion that pre-training helps is not built into the comparison. The salience maps (Eq. 1, Eq. 21) are post-hoc input-gradient scores of a trained model, and Section 6 explicitly disclaims causality ('this does not imply causality'); validating them against the pretectum-to-tail relationship uses published anatomy ([3], and [34] for the known optomotor response circuit). Reference [34] overlaps with two present authors, but it is an independent published study, not an unverified uniqueness theorem or ansatz imported to force the conclusion, and it is corroborated by the non-self citation [3]. There is no self-definitional equation pair, no renaming of a known result as organization, and no assertion that the authors' prior work uniquely forces the model choice. A possible concern about Mann-Whitney tests applied to overlapping sliding windows would be a statistical-independence issue affecting significance claims, not a definitional circularity in which an output equals an input by construction. The paper is therefore self-contained as an empirical benchmark comparison.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper is an empirical benchmark; the central claims rest on data processing choices (Suite2P sources as neurons, temporal averaging of tail kinematics), statistical test assumptions (independence of overlapping windows), and a uniform fine-tuning schedule across models. No invented entities are introduced.

free parameters (3)
  • Fine-tuning learning rate = 1e-6 (Exp 1,4), 1e-4 (Exp 2,3)
    A single LR is applied to all models in each experiment without per-model tuning; the ranking of models could change with different LRs.
  • Number of fine-tuning epochs = 5-10
    Fixed range without early stopping per model; affects relative performance.
  • Top-k salient neuron count = k=10, k=20, top 50
    Chosen for display, not derived; the anatomical validation uses these hand-picked thresholds.
assumptions (3)
  • domain assumption Calcium fluorescence traces from Suite2P are assumed to represent single-neuron activity
    Section 3.1: 'we refer to these extracted sources as neurons.'
  • domain assumption Averaging tail kinematics at 200 Hz down to the 1.1 Hz frame rate preserves the behaviorally relevant signal
    Section 3.4: 'we compressed this sample set... by simply averaging across all output time steps in a given frame.' High-frequency tail dynamics are lost, which may reduce the apparent decoding accuracy.
  • ad hoc to paper The Mann-Whitney U test and Student's t-test are valid when applied to overlapping sliding-window sequences
    Section 3.4 treats each sequence as an independent sample even though windows overlap, inflating the effective N.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NLP4Neuro: Sequence-to-sequence learning for neural population decoding." pith.science (2026). https://pith.science/paper/I7UXFB4O

@misc{pith2026250702264,
  author       = {Pith},
  title        = {Pith review of: NLP4Neuro: Sequence-to-sequence learning for neural population decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I7UXFB4O}},
  note         = {Machine review of arXiv:2507.02264}
}
read the original abstract

Delineating how animal behavior arises from neural activity is a foundational goal of neuroscience. However, as the computations underlying behavior unfold in networks of thousands of individual neurons across the entire brain, this presents challenges for investigating neural roles and computational mechanisms in large, densely wired mammalian brains during behavior. Transformers, the backbones of modern large language models (LLMs), have become powerful tools for neural decoding from smaller neural populations. These modern LLMs have benefited from extensive pre-training, and their sequence-to-sequence learning has been shown to generalize to novel tasks and data modalities, which may also confer advantages for neural decoding from larger, brain-wide activity recordings. Here, we present a systematic evaluation of off-the-shelf LLMs to decode behavior from brain-wide populations, termed NLP4Neuro, which we used to test LLMs on simultaneous calcium imaging and behavior recordings in larval zebrafish exposed to visual motion stimuli. Through NLP4Neuro, we found that LLMs become better at neural decoding when they use pre-trained weights learned from textual natural language data. Moreover, we found that a recent mixture-of-experts LLM, DeepSeek Coder-7b, significantly improved behavioral decoding accuracy, predicted tail movements over long timescales, and provided anatomically consistent highly interpretable readouts of neuron salience. NLP4Neuro demonstrates that LLMs are highly capable of informing brain-wide neural circuit dissection.

Figures

Figures reproduced from arXiv: 2507.02264 by the authors.

Figure 1
Figure 1. Use of NLP4Neuro to predict behavioral output from simultaneous neural activity in larval zebrafish (a) Volumetric two-photon calcium imaging of head-fixed, tail-freed larval zebrafish during visual stimulation while recording tail movements at 200 Hz. Pt: pretectum; nMLF: nucleus of the medial longitudinal fasciculus; Hb: hindbrain. (b) The NLP4Neuro pipeline selects an LLM to learn a mapping of GCaMP6s fluorescenc… view at source ↗
Figure 2
Figure 2. Pre-trained DeepSeek-c7b captures tail kinematic events, simple RNNs do not. (a) Grouped model test set performance across fish, with increasing sequence lengths, s. (b) Pre￾trained versus un-trained performance for transformer models. For (a-b), we use data from N = 5 fish. A * denotes significant differences (p < 0.05) using the Mann-Whitney U test. (c) Sampled behavioral output and test set predictions with 95% c… view at source ↗
Figure 3
Figure 3. Relative position improves performance across transformers. DeepSeek-c7b is not sensitive to embedding strategies. Performance is shown for a representative fish, with s = 20. Differences (*) across model groups, and within groups (**-**** for GPT-2, DeepSeek, and BERT-bu respectively) are shown for each embedding strategy, using a Mann-Whitney U test (p < 0.05). For a sequence length s = 20 and representative fish,… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Globally salient neurons are found in the pretectum (Pt). (a) Neuron salience for DeepSeek-c7b, projected onto the GCaMP6s image for a representative fish. Dot size reflects salience score for predicting tail output. (b) Regional clustering of all neurons. (c) Mean sco…
Figure 5
Figure 5. Figure 5: DeepSeek predicts rightward and leftward tail movements, and is interpretable. Salient neurons show bilateral symmetry. (a) Predicted and ground truth sequences (s = 20) for right (red) and left (blue) tail movements. (b) Left (blue) and right (red) motion salience ove…
Figure 6
Figure 6. Figure 6: Moment-to-moment salience mapping identifies known functional regions and lateral￾isation. (a) An exemplar sequence with predicted right tail movements from DeepSeek-c7b (s = 20). (b) Spatial distribution of the top 50 most salient pretectal (Pt), hindbrain (Hb), and o…
Figure 7
Figure 7. Figure 7: PCA applied to the full neural data token, with each point in PC-space corresponding to a [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: PCA applied to all neurons, with each point indicating a single frame of data for fish 1. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 63 canonical work pages

  1. [1]

    https://www

    Neural data transformer 3: A foundation model for motor cortical decoding. https://www. krellinst.org/csgf/conf/2024/abstracts/ye2022. Accessed: 2025-5-15

    work page 2024

  2. [2]

    Neuronal dynamics regulating brain and behavioral state transitions

    Aaron S Andalman, Vanessa M Burns, Matthew Lovett-Barron, Michael Broxton, Ben Poole, Samuel J Yang, Logan Grosenick, Talia N Lerner, Ritchie Chen, Tyler Benster, Philippe Mourrain, Marc Levoy, Kanaka Rajan, and Karl Deisseroth. Neuronal dynamics regulating brain and behavioral state transitions. Cell, 177(4):970–985.e20, May 2019

    work page 2019

  3. [3]

    Pretectal neurons control hunting behaviour

    Paride Antinucci, Mónica Folgueira, and Isaac H Bianco. Pretectal neurons control hunting behaviour. Elife, 8, October 2019

    work page 2019

  4. [4]

    Multi-session, multi-task neural decoding from distinct cell-types and brain regions

    Mehdi Azabou, Krystal Xuejing Pan, Vinam Arora, Ian Jarratt Knight, Eva L Dyer, and Blake Aaron Richards. Multi-session, multi-task neural decoding from distinct cell-types and brain regions. In The Thirteenth International Conference on Learning Representations, October 2024

    work page 2024

  5. [5]

    Population transformer

    Andrei Barbu. Population transformer. https://glchau.github.io/ population-transformer/. Accessed: 2025-5-8

    work page 2025

  6. [6]

    Elegans-AI: How the connectome of a living organism could model artificial neural networks

    Francesco Bardozzo, Andrea Terlizzi, Claudio Simoncini, Pietro Lió, and Roberto Tagliaferri. Elegans-AI: How the connectome of a living organism could model artificial neural networks. Neurocomputing, 584(127598):127598, June 2024

    work page 2024

  7. [7]

    Laplacian eigenmaps for dimensionality reduction and data representation

    Mikhail Belkin and Partha Niyogi. Laplacian eigenmaps for dimensionality reduction and data representation. Neural Comput., 15(6):1373–1396, June 2003

    work page 2003

  8. [8]

    MeLM, a generative pretrained language modeling framework that solves for- ward and inverse mechanics problems

    Markus J Buehler. MeLM, a generative pretrained language modeling framework that solves for- ward and inverse mechanics problems. J. Mech. Phys. Solids, 181(105454):105454, December 2023

    work page 2023

Show all 63 references
  1. [9]

    Population transformer: Learning population- level representations of neural activity

    Geeling Chau, Christopher Wang, Sabera Talukder, Vighnesh Subramaniam, Saraswati Soedar- madji, Yisong Yue, Boris Katz, and Andrei Barbu. Population transformer: Learning population- level representations of neural activity. arXiv [cs.LG], June 2024

    2024

  2. [10]

    PepMLM: Target sequence-conditioned generation of therapeutic peptide binders via span masked language modeling

    Tianlai Chen, Madeleine Dumas, Rio Watson, Sophia Vincoff, Christina Peng, Lin Zhao, Lauren Hong, Sarah Pertsemlidis, Mayumi Shaepers-Cheu, Tian Zi Wang, Divya Srijay, Connor Monticello, Pranay Vure, Rishab Pulugurta, Kseniia Kholina, Shrey Goel, Matthew P DeLisa, Ray Truant, ...

    arXiv

  3. [11]

    Visual control of walking speed in drosophila

    Matthew S Creamer, Omer Mano, and Damon A Clark. Visual control of walking speed in drosophila. Neuron, 100(6):1460–1473.e6, December 2018

    2018

  4. [12]

    Bridging the gap between the connectome and whole-brain activity in C

    Matthew S Creamer, Andrew M Leifer, and Jonathan W Pillow. Bridging the gap between the connectome and whole-brain activity in C. elegans. bioRxiv, page 2024.09.22.614271, September 2024

    2024

  5. [13]

    Sparse autoencoders find highly interpretable features in language models

    Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. arXiv [cs.LG], September 2023

    2023

  6. [14]

    Extraction of salient sentences from labelled documents

    Misha Denil, Alban Demiraj, and Nando de Freitas. Extraction of salient sentences from labelled documents. arXiv [cs.CL], December 2014

    2014

  7. [15]

    BERT: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv [cs.CL], October 2018

    2018

  8. [16]

    Exploring deep learning models for EEG neural decoding

    Laurits Dixen, Stefan Heinrich, and Paolo Burelli. Exploring deep learning models for EEG neural decoding. arXiv [cs.LG], March 2025. 10

    2025

  9. [17]

    Effects of connectivity on narrative temporal processing in structured reservoir computing

    Peter Ford Dominey, Timothy M Ellmore, and Jocelyne Ventre-Dominey. Effects of connectivity on narrative temporal processing in structured reservoir computing. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1–8, July 2022

    2022

  10. [18]

    Neural circuits underlying divergent visuomotor strategies of zebrafish and danionella cerebrum

    Kaitlyn E Fouke, Zichen He, Matthew D Loring, and Eva A Naumann. Neural circuits underlying divergent visuomotor strategies of zebrafish and danionella cerebrum. Curr. Biol., May 2025

    2025

  11. [19]

    Next generation reservoir computing

    Daniel J Gauthier, Erik Bollt, Aaron Griffith, and Wendson A S Barbosa. Next generation reservoir computing. Nat. Commun., 12(1):5564, September 2021

    2021

  12. [20]

    CaImAn an open source tool for scalable calcium imaging data analysis

    Andrea Giovannucci, Johannes Friedrich, Pat Gunn, Jérémie Kalfon, Brandon L Brown, Sue Ann Koay, Jiannis Taxidis, Farzaneh Najafi, Jeffrey L Gauthier, Pengcheng Zhou, Baljit S Khakh, David W Tank, Dmitri B Chklovskii, and Eftychios A Pnevmatikakis. CaImAn an open source tool f...

    2019

  13. [21]

    Behavioral assessment of visual function via optomotor response and cognitive function via Y-maze in diabetic rats

    Kaavya Gudapati, Anayesha Singh, Danielle Clarkson-Townsend, Stephen Q Phillips, Amber Douglass, Andrew J Feola, and Rachael S Allen. Behavioral assessment of visual function via optomotor response and cognitive function via Y-maze in diabetic rats. J. Vis. Exp., (164): 10.379...

    doi

  14. [22]

    DeepSeek-coder: When the large language model meets programming – the rise of code intelligence

    Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y Wu, Y K Li, Fuli Luo, Yingfei Xiong, and Wenfeng Liang. DeepSeek-coder: When the large language model meets programming – the rise of code intelligence. arXiv [cs.SE], January 2024

    2024

  15. [23]

    A brain-wide circuit model of heat-evoked swimming behavior in larval zebrafish

    Martin Haesemeyer, Drew N Robson, Jennifer M Li, Alexander F Schier, and Florian Engert. A brain-wide circuit model of heat-evoked swimming behavior in larval zebrafish. Neuron, 98(4): 817–831.e6, May 2018

    2018

  16. [24]

    Decoding layer saliency in language transformers

    Elizabeth M Hou and Gregory Castanon. Decoding layer saliency in language transformers. arXiv [cs.CL], August 2023

    2023

  17. [25]

    Long short-term memory-based neural decoding of object categories evoked by natural images

    Wei Huang, Hongmei Yan, Chong Wang, Jiyi Li, Xiaoqing Yang, Liang Li, Zhentao Zuo, Jiang Zhang, and Huafu Chen. Long short-term memory-based neural decoding of object categories evoked by natural images. Hum. Brain Mapp., 41(15):4442–4453, October 2020

    2020

  18. [26]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv [cs.LG], December 2014

    2014

  19. [27]

    Connectome-constrained networks predict neural activity across the fly visual system

    Janne K Lappalainen, Fabian D Tschopp, Sridhama Prakhya, Mason McGill, Aljoscha Nern, Kazunori Shinomiya, Shin-Ya Takemura, Eyal Gruntman, Jakob H Macke, and Srinivas C Turaga. Connectome-constrained networks predict neural activity across the fly visual system. Nature, 634(80...

    2024

  20. [28]

    Do emergent abilities exist in quantized large language models: An empirical study

    Peiyu Liu, Zikang Liu, Ze-Feng Gao, Dawei Gao, Wayne Xin Zhao, Yaliang Li, Bolin Ding, and Ji-Rong Wen. Do emergent abilities exist in quantized large language models: An empirical study. arXiv [cs.CL], July 2023

    2023

  21. [29]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv [cs.LG], November 2017

    2017

  22. [30]

    ZAPBench: A benchmark for whole-brain activity prediction in zebrafish

    Jan-Matthis Lueckmann, Alexander Immer, Alex Bo-Yuan Chen, Peter H Li, Mariela D Petkova, Nirmala A Iyer, Luuk Willem Hesselink, Aparna Dev, Gudrun Ihrke, Woohyun Park, Alyson Petruncio, Aubrey Weigel, Wyatt Korff, Florian Engert, Jeff W Lichtman, Misha B Ahrens, Michał Janusz...

    2025

  23. [31]

    Decoding the brain: From neural representations to mechanistic models

    Mackenzie Weygandt Mathis, Adriana Perez Rotondo, Edward F Chang, Andreas S Tolias, and Alexander Mathis. Decoding the brain: From neural representations to mechanistic models. Cell, 187(21):5814–5832, October 2024

    2024

  24. [32]

    Connectomes inform function: from time-varying dynamics to animal behaviour

    Jacob Morra, Kaitlyn Fouke, Eva A Naumann, and Mark Daley. Connectomes inform function: from time-varying dynamics to animal behaviour. Nat. Comput., pages 1–18, June 2025. 11

    2025

  25. [33]

    Harnessing behavioral diversity to understand neural computations for cognition

    Simon Musall, Anne E Urai, David Sussillo, and Anne K Churchland. Harnessing behavioral diversity to understand neural computations for cognition. Curr. Opin. Neurobiol., 58:229–238, October 2019

    2019

  26. [34]

    From whole-brain data to functional circuit models: The zebrafish optomotor response

    Eva A Naumann, James E Fitzgerald, Timothy W Dunn, Jason Rihel, Haim Sompolinsky, and Florian Engert. From whole-brain data to functional circuit models: The zebrafish optomotor response. Cell, 167(4):947–960.e20, November 2016

    2016

  27. [35]

    Andrew Y . Ng. Sparse autoencoders. CS294A Lecture Notes 72, Stanford University, 2011. URL https://web.stanford.edu/class/cs294a/sparseAutoencoder.pdf

    2011

  28. [36]

    Suite2p: beyond 10,000 neurons with standard two-photon microscopy

    Marius Pachitariu, Carsen Stringer, Mario Dipoppa, Sylvia Schröder, L Federico Rossi, Henry Dalgleish, Matteo Carandini, and Kenneth D Harris. Suite2p: beyond 10,000 neurons with standard two-photon microscopy. bioRxiv, page 061507, June 2016

    2016

  29. [37]

    PyTorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fan...

    2019

  30. [38]

    Language models are unsupervised multitask learners

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. February 2019

    2019

  31. [39]

    Scaling language models: Methods, analysis & insights from training gopher

    Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan, Jacob Menick, Albin Cassirer, Richard Powell, George van den Driessche, Lisa Anne Hendricks, ...

    2021

  32. [40]

    Is attention interpretable? arXiv [cs.CL], June 2019

    Sofia Serrano and Noah A Smith. Is attention interpretable? arXiv [cs.CL], June 2019

    2019

  33. [41]

    Self-attention with relative position repre- sentations

    Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position repre- sentations. arXiv [cs.CL], March 2018

    2018

  34. [42]

    Comparative analysis of neural decoding algorithms for brain-machine interfaces

    Olena Shevchenko, Sofiia Yeremeieva, and Brokoslaw Laschowski. Comparative analysis of neural decoding algorithms for brain-machine interfaces. bioRxiv, page 2024.12.05.627080, December 2024

    2024

  35. [43]

    Optimization of optomotor response-based visual function assessment in mice

    Cong Shi, Xuedong Yuan, Karen Chang, Kin-Sang Cho, Xinmin Simon Xie, Dong Feng Chen, and Gang Luo. Optimization of optomotor response-based visual function assessment in mice. Sci. Rep., 8(1):9708, June 2018

    2018

  36. [44]

    Deep inside convolutional networks: Visualising image classification models and saliency maps

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv [cs.CV], December 2013

    2013

  37. [45]

    Applications of brain-computer interfaces in neurodegenerative diseases

    Hossein Tayebi, Sina Azadnajafabad, Seyed Farzad Maroufi, Ahmad Pour-Rashidi, Mirhojjat Khorasanizadeh, Sina Faramarzi, and Konstantin V Slavin. Applications of brain-computer interfaces in neurodegenerative diseases. Neurosurg. Rev., 46(1):131, May 2023

    2023

  38. [46]

    A connectome based hexagonal lattice convolutional network model of the drosophila visual system

    Fabian David Tschopp, Michael B Reiser, and Srinivas C Turaga. A connectome based hexagonal lattice convolutional network model of the drosophila visual system. arXiv [q- bio.NC], June 2018. 12

    2018

  39. [47]

    Large-scale neural recordings call for new insights to link brain and behavior

    Anne E Urai, Brent Doiron, Andrew M Leifer, and Anne K Churchland. Large-scale neural recordings call for new insights to link brain and behavior. Nat. Neurosci., 25(1):11–19, January 2022

    2022

  40. [48]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv [cs.CL], June 2017

    2017

  41. [49]

    Computation through neural population dynamics

    Saurabh Vyas, Matthew D Golub, David Sussillo, and Krishna V Shenoy. Computation through neural population dynamics. Annu. Rev. Neurosci., 43(1):249–275, July 2020

    2020

  42. [50]

    BrainBERT: Self-supervised representation learning for intracranial recordings

    Christopher Wang, Vighnesh Subramaniam, Adam Uri Yaari, Gabriel Kreiman, Boris Katz, Ignacio Cases, and Andrei Barbu. BrainBERT: Self-supervised representation learning for intracranial recordings. arXiv [cs.LG], February 2023

    2023

  43. [51]

    Pre-trained language models and their applications

    Haifeng Wang, Jiwei Li, Hua Wu, Eduard Hovy, and Yu Sun. Pre-trained language models and their applications. Engineering (Beijing), 25:51–65, September 2022

    2022

  44. [52]

    Learning natural language inference with LSTM.arXiv [cs.CL], December 2015

    Shuohang Wang and Jing Jiang. Learning natural language inference with LSTM.arXiv [cs.CL], December 2015

    2015

  45. [53]

    Emergent abilities of large language models

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. Emergent abilities of large language models. ar...

    2022

  46. [54]

    Nystromformer: A nystrom-based algorithm for approximating self-attention

    Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. arXiv [cs.CL], February 2021

    2021

  47. [55]

    Representation learning for neural population activity with neural data transformers

    Joel Ye and Chethan Pandarinath. Representation learning for neural population activity with neural data transformers. bioRxiv, January 2021

    2021

  48. [56]

    Neural data transformer 2: Multi-context pretraining for neural spiking activity

    Joel Ye, Jennifer L Collinger, Leila Wehbe, and Robert Gaunt. Neural data transformer 2: Multi-context pretraining for neural spiking activity. bioRxivorg, page 2023.09.18.558113, September 2023

    2023

  49. [57]

    Generalizable LLM learning of graph synthetic data with reinforcement learning

    Yizhuo Zhang, Heng Wang, Shangbin Feng, Zhaoxuan Tan, Xinyun Liu, and Yulia Tsvetkov. Generalizable LLM learning of graph synthetic data with reinforcement learning. arXiv [cs.LG], June 2025

    2025

  50. [58]

    universal translator

    Yizi Zhang, Yanchen Wang, Donato Jimenez-Beneto, Zixuan Wang, Mehdi Azabou, Blake Richards, Olivier Winter, International Brain Laboratory, Eva Dyer, Liam Paninski, and Cole Hurwitz. Towards a “universal translator” for neural dynamics at single-cell, single-spike resolution. ...

    2024

  51. [59]

    MLPST: MLP is all you need for spatio-temporal prediction

    Zijian Zhang, Ze Huang, Zhiwei Hu, Xiangyu Zhao, Wanyu Wang, Zitao Liu, Junbo Zhang, S Joe Qin, and Hongwei Zhao. MLPST: MLP is all you need for spatio-temporal prediction. arXiv [cs.LG], September 2023

    2023

  52. [60]

    Stytra: An open-source, integrated system for stimulation, tracking and closed-loop behavioral experiments

    Vilim Štih, Luigi Petrucco, Andreas M Kist, and Ruben Portugues. Stytra: An open-source, integrated system for stimulation, tracking and closed-loop behavioral experiments. PLoS Comput. Biol., 15(4):e1006699, April 2019. 13 A Technical Appendices and Supplementary Material A.1...

    2019

  53. [61]

    Multi-Headed Self-Attention: 14 For all tokens u(t) ∈ Rn,m, we construct a key matrix K, query matrix Q, and value matrix V , where Q(t) = u(t)W Q, K (t) = u(t)W K, V (t) = u(t)W V . (8) To determine which tokens in the sequence should be attended to, based on the context of a...

  54. [62]

    u′(t) = LayerNorm u(t) + z(t)

    Layer Normalization After applying multi-headed attention, and generating z(t), we simply update the existing token u(t) with z(t) and apply layer normalization. u′(t) = LayerNorm u(t) + z(t) . (11)

  55. [63]

    In BERT, the gaussian error linear unit, or GELU activation function is used, as described in [15]

    Feed-Forward Network Finally, we run the transformed token u′(t) through a multi-layer perceptron (MLP) which reduces to the dimensionality of the input token. In BERT, the gaussian error linear unit, or GELU activation function is used, as described in [15]. ˜u(t) = GELU u′(t...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.