Biological Processing Units: Leveraging an Insect Connectome to Pioneer Biofidelic Neural Architectures

Beiya Xu; Jason Brown; Joshua T. Vogelstein; R. Jacob Vogelstein; Siyu Yu; Tingshan Liu; Zihan Qin

REVIEW 4 major objections 5 minor 22 references

Biological Processing Units: Leveraging an Insect Connectome to Pioneer Biofidelic Neural Architectures

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A fixed recurrent network built from the complete Drosophila larval connectome, with only its input and output projections trained, matches or beats size-matched MLPs on MNIST and CIFAR-10 and outperforms larger transformers on chess…

desk verdict The idea is new and worth a look, but the reported superiority claims rely on unfair baselines and an undescribed chess search. read the letter →

arxiv 2507.10951 v1 pith:SH4RC7RV submitted 2025-07-15 cs.NE cs.AIq-bio.NC

Siyu Yu , Zihan Qin , Tingshan Liu , Beiya Xu , R. Jacob Vogelstein , Jason Brown , Joshua T. Vogelstein This is my paper · ORCID

classification cs.NEcs.AIq-bio.NC

keywords biologicalprocessingunitDrosophilaconnectomelotteryticketreservoircomputingexpansiondegree-correctedstochasticblockmodelchesspuzzlesolvingfrozenrecurrentnetwork

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

The paper claims that a complete biological connectome — the 3,000-neuron, 65,000-synapse wiring diagram of a larval Drosophila brain — can serve as a frozen recurrent core for artificial intelligence. Nothing inside the network is trained: the synaptic weights stay exactly as measured, and only the linear projections that feed stimuli in and read answers out are learned. With that setup, the unmodified Biological Processing Unit reaches 98 percent accuracy on MNIST and 58 percent on CIFAR-10, beating a two-layer MLP matched on trainable parameter count. Expanding the connectome with a signed degree-corrected stochastic block model raises CIFAR-10 accuracy monotonically, and on chess the same frozen core, paired with lightweight encoders, performs favorably against larger transformers. The thesis is that evolutionarily selected circuits are reusable, general-purpose computational substrates — biological lottery tickets.

What carries the argument

The central object is the Biological Processing Unit (BPU): a fixed recurrent core built directly from the larva's axon-to-dendrite adjacency matrix, with each connection signed excitatory or inhibitory using neurotransmitter-based annotations. Its dynamics update sensory, internal, and output populations as $S^{(t+1)} = f(W_{ss}S^{(t)} + W_{rs}I^{(t)} + W_{os}O^{(t)} + E^{(t)})$ with analogous updates for the internal and output pools, where the $W_{xy}$ matrices are the untouched connectome weights and $f$ is a nonlinearity; the external input $E^{(t)}$ and the output decoding are the only trainable routes. Performance at scale comes from a directed, signed degree-corrected stochastic block model (DCSBM), which expands the connectome up to $5\times$ while reproducing its block-level weight densities, sign probabilities, and degree structure, then restores the original sub-matrix so expanded reservoirs stay anchored in real biology. For chess, learnable graph (GINEConv) or convolutional encoders turn board positions into fixed-size embeddings that drive the same frozen core.

What would settle it

A concrete test would replace the connectome's synaptic weights with randomly permuted or rewired weights (preserving the same degree sequence and sign balance) and retrain only input and output projections; if the scrambled reservoir reproduces the 98% MNIST and 58% CIFAR-10 numbers, the specific biological wiring is not doing the work. For chess, evaluate the CNN-BPU without the depth-6 minimax search against the ~2M and 9M transformers; if the gap disappears, the search procedure rather than the connectome carries the 91.7% result.

Watch

Extended reading notes

Core claim

The central claim is that the intact Drosophila larval connectome, taken verbatim from electron-microscopy reconstructions and left structurally untouched, is sufficient to support competitive performance on perception and planning tasks. The authors convert the wiring diagram into a fixed recurrent reservoir: neurons are partitioned into sensory (430), internal (2,304), and output (218) pools, synaptic counts are signed by neurotransmitter annotations, and the recurrent dynamics are unrolled over a fixed number of steps. Only the input and output projection matrices are trained by gradient descent. This frozen reservoir yields 98% MNIST and 58% CIFAR-10 accuracy, both above a size-matched MLP baseline; connectome expansion via a directed, signed degree-corrected stochastic block model pushes CIFAR-10 higher. For chess, a GNN-BPU with 232,912 trainable parameters reaches about 60% puzzle accuracy after 10,000 training games, and a CNN-BPU with ~2M parameters beats an equivalent-size Transformer; adding depth-6 minimax search at inference raises accuracy to 91.71%, above a 9M-parameter Transformer. The authors read these results as evidence that biological circuits carry computational inductive priors that survive without internal weight training.

Load-bearing premise

The load-bearing premise is that the 'size-matched' baselines are fair comparisons: the MLP matches only the BPU's trainable projection count while the BPU also has a fixed 65,000-weight structured recurrent core, and the chess comparison grants the CNN-BPU a depth-6 minimax search that the transformer baselines do not use.

Editorial extensions

If this is right

If the connectome is a biological lottery ticket, architecture search over the recurrent core becomes unnecessary: a fixed, measured wiring graph with trained read-in and read-out projections already reaches competitive accuracy on image classification.
Synthetic connectome expansion preserves the biological prior at larger scales, so CIFAR-10 accuracy grows monotonically with expansion factor without any additional training of the recurrent weights.
The same frozen core transfers across vision and chess, suggesting the substrate is general-purpose rather than specialized for one sensory modality.
The chess results indicate strong data efficiency: a 232,912-parameter GNN-BPU reaches about 60% puzzle accuracy after only 10,000 training games, far outpacing transformers in the low-data regime.
Search at inference stacks on the connectome's representation: depth-6 minimax with alpha-beta pruning lifts the CNN-BPU to 91.71% accuracy, surpassing a 9M-parameter transformer that did not use search.

Reading between the lines

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

Editorial inference: if the biological lottery ticket idea holds, the same frozen-reservoir recipe should be testable on larger connectomes (adult Drosophila, mouse) once available; the paper motivates this trajectory but does not perform it, and its DCSBM expansion already provides a scaling tool for such tests.
Editorial inference: the reported margins may shrink under stricter matching — a fairer MLP baseline would receive a comparably sized fixed random reservoir (65,000 frozen weights of its own), isolating what the biological wiring specifically contributes over random wiring.
Editorial inference: the chess comparison conflates representation and search; adding the same depth-6 minimax to the transformer baselines would separate how much of the 91.7% comes from the connectome-derived features versus classical search.
Editorial inference: the modality ablation's odd ordering (a 26-neuron respiratory group beating a 29-neuron sight group at small training sizes) hints that wiring substructure, not neuron count, carries inductive bias; correlating task performance with graph motifs or path statistics would turn this hint into a testable hypothesis.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

First, what's real: the paper is the first to use the complete Drosophila larval connectome as a fixed recurrent reservoir for image classification and chess. That's a new and sensible experiment. The DCSBM scaling method is a modest addition but useful for studying scale. The reported numbers—98% MNIST, 58% CIFAR-10, and the GNN-BPU chess results—are plausible in principle.

The central problem is the baseline. Section 2.3's MLP uses a fixed random hidden-to-hidden projection, so it's not a trained MLP at all. Matching trainable parameters doesn't match computational capacity: the BPU unrolls a 3,000-neuron graph for T steps, so it has far more forward computation. The gap between 98% vs 97% and 58% vs 52% can be explained by that extra computation. The chess comparison is worse: the abstract claims 'nearly 10x better' but the text says competitive or superior with no numbers. The CNN-BPU with depth-6 minimax at inference lacks any description of the value function or how it integrates with the network, and the Transformer doesn't get search. That's not a fair test.

These are load-bearing, because the headline is that biofidelic networks beat standard architectures. As presented, that claim is not supported. The stress-test note is right on both counts.

That said, the underlying hypothesis—that a complete biological connectome is a useful computational substrate—remains interesting and worth testing with proper controls. The paper shows clear thinking; the flaws are in the experimental design and reporting, not in the logic.

I'd send this to review, but major revision required. The authors need to rerun baselines with a properly trained MLP of comparable actual forward pass, either give the Transformer a search or remove search from BPU, describe the minimax details, and fix the abstract. The novelty is sufficient; the execution is not.

Referee Report

4 major / 5 minor

Summary. The paper introduces the Biological Processing Unit (BPU), a fixed-weight recurrent network derived from the complete Drosophila larval connectome, and reports that the unmodified BPU achieves 98% accuracy on MNIST and 58% on CIFAR-10, surpassing a size-matched MLP baseline. It also proposes a directed, signed degree-corrected stochastic block model to expand the connectome up to 5x, reports monotonic CIFAR-10 improvements with expansion, and evaluates GNN-BPU and CNN-BPU variants on chess puzzles, claiming that a lightweight GNN-BPU reaches 60% move accuracy and that a CNN-BPU with a depth-6 minimax search reaches 91.7% puzzle accuracy, exceeding Transformer baselines.

Significance. If the comparative claims were sound, this would be a notable demonstration of the 'biological lottery ticket' idea: a fixed, evolutionarily selected circuit serving as a reusable neural substrate for perception and planning. The architectural concept is interesting, and the DCSBM expansion is a concrete and potentially useful generative model for scaling connectome-derived reservoirs. The paper's central claims, however, are not currently established: the image-classification baseline is not a trained MLP, the abstract's 'nearly 10x better' chess claim is unsupported by the body, and the chess success with minimax adds inference-time computation not applied to the Transformer baselines. These issues are load-bearing, so the significance of the reported results is substantially diminished as presented.

major comments (4)

[2.3, 3.1, Eq. (1)] The 'size-matched MLP' baseline is not a trained MLP: its hidden-to-hidden transform is a fixed random projection, and only the input and output projections are trained. Matching the trainable projection parameter count does not match the capacity or the amount of computation performed, because the BPU unrolls a 3,000-neuron recurrent core with 65,000 fixed weights for T steps (Eq. 1). The gaps 98% versus 97% on MNIST and 58% versus 52% on CIFAR-10 therefore cannot be attributed to the biological origin of the weights; a fully trained MLP with matched parameters or matched FLOPs is needed to support the 'surpassing size-matched MLPs' claim.
[Abstract, 3.2, Fig. 4] The abstract's claim that the GNN-BPU achieves "nearly 10x better than any size transformer" is not supported anywhere in the paper. Section 3.2 reports 59%, 61%, and 63% accuracy and says the model is "competitive with substantially larger models," with no 10x figure and no definition of whether the comparison is move accuracy, puzzle accuracy, error rate, or parameter efficiency; either a precise comparison or a retraction is required.
[3.2, Fig. 5] The CNN-BPU result of 91.7% includes a depth-6 minimax search with alpha-beta pruning at inference, while the Transformer baselines are not given any equivalent search. The paper never specifies the value function used by the search, how it is integrated with the BPU policy, or the relative compute budgets, so the "outperforms parameter-matched Transformers" claim conflates the architecture with additional inference-time computation; a search-augmented Transformer or an equal-compute comparison is needed.
[2.2, 3.1, Fig. 2B] The DCSBM expansion is fit to the connectome's degree and block statistics (Eqs. 2-3), so Fig. 2B's monotonic CIFAR-10 improvement could come from generic graph size or degree structure rather than preserved biological wiring. The paper should include controls such as degree-preserving edge rewiring or shuffled block labels to show that the specific connectome-derived structure, not the DCSBM prior, is responsible for the gains.

minor comments (5)

[2.4] Section 2.4 says "10 × 106 board positions," which is ambiguous; it should be written as 10^6 or 10 × 10^6 to avoid confusion.
[2.4] The phrase "The precision of the puzzle is measured under the training budgets of the games" is grammatically unclear; rephrase as "Puzzle accuracy is reported for training budgets of 10^4, 10^5, and 10^6 games."
[2.3, Eq. (1)] Section 2.3 says the external input E(t) is supplied at t=0, but Eq. (1) includes E(t) at every time step; please reconcile the notation.
[4] Section 4 contains a typo: "efficent" should be "efficient."
[References] Reference [20] (Geroldinger and Khadam, "On the arithmetic of monoids of ideals") appears unrelated to the sentence about constrained plasticity mechanisms; please verify and replace the citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the frozen connectome is external input; tasks train only readout projections.

full rationale

The derivation chain is not circular. The BPU's recurrent weights are taken directly from an external biological dataset (the Drosophila larval connectome, Winding et al. 2023) and remain frozen; only input/output readout projections are trained on task labels. This is standard supervised readout training, not a construction that encodes the test labels into the fixed core. The DCSBM expansion is fitted only to connectome wiring statistics (Eqs. 2-3), then used to generate new frozen reservoirs whose performance is measured; the reported accuracies are empirical outputs, not restatements of the fitted block densities or sign probabilities. Chess results likewise compare a fixed connectome core plus trained encoder/readout against external ChessBench Transformer baselines from Ruoss et al.; no uniqueness theorem or self-citation is invoked to rule out alternatives. The only notable weaknesses are comparability of the MLP baseline (Section 2.3 fixes the hidden-to-hidden transform as an untrained random projection) and the depth-6 minimax added to CNN-BPU at inference in Figure 5; these are experimental fairness concerns, not circular derivations, because the BPU's accuracy is not defined in terms of those baselines or the minimax. No equation reduces to its own input by construction.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The central claim depends on the accuracy of the connectome, the polarity assignments, and the reservoir computing paradigm. The DCSBM expansion adds fitted parameters that are not validated against biology. The chess results additionally assume the ChessBench labels and the unspecified minimax integration are reliable.

free parameters (4)

T (recurrent time steps) = not stated
Number of unrolled recurrent steps in the reservoir, chosen to match characteristic synaptic propagation path length (Section 2.1), affects dynamics and accuracy.
DCSBM block densities omega_gh = estimated from connectome (Eq. 2)
Block-pair edge densities fitted to the empirical connectome; used to generate expanded reservoirs.
DCSBM sign probabilities p_gh = estimated from connectome (Eq. 3)
Probability of excitatory versus inhibitory sign per block pair, fitted to the empirical connectome.
DCSBM degree correction parameters theta_out and theta_in = bootstrapped from core neurons per block and rescaled
Node degree parameters used in the Poisson sampling for expanded graphs.

assumptions (6)

domain assumption The Drosophila larval connectome reconstruction is accurate and complete (Winding et al., 2021)
The entire BPU is built from this wiring diagram; if the reconstruction contains errors, the reservoir is not the true fly brain.
domain assumption Neurotransmitter-based annotations correctly determine synaptic polarity
Synaptic counts are multiplied by neurotransmitter annotations to create signed weights (Section 2.1, ref [9]); incorrect polarity would change the reservoir dynamics.
standard math A fixed recurrent network with trained readout can solve the tasks (reservoir computing principle)
The entire BPU approach relies on the reservoir computing paradigm that fixed nonlinear recurrent networks with trained linear readouts are sufficient.
ad hoc to paper The DCSBM expansion preserves task-relevant structure of the connectome
The paper assumes that stochastically generated graphs from a fitted block model retain the computational properties of the real connectome (Section 2.2).
domain assumption ChessBench labels and Stockfish annotations are reliable ground truth
All chess accuracy numbers depend on the correctness of the ChessBench dataset [8] annotations.
domain assumption Minimax search with alpha-beta pruning can be driven by the BPU's outputs without additional training
The paper claims search refines move selection but does not specify how the BPU generates values or policies for search nodes (Section 2.4).

invented entities (1)

Hub node in GNN encoder
purpose: A 65th node connecting all board squares to provide global context for chess position encoding
A graph construction choice for encoding chess positions; not empirically validated as necessary or biologically motivated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Biological Processing Units: Leveraging an Insect Connectome to Pioneer Biofidelic Neural Architectures." pith.science (2026). https://pith.science/paper/SH4RC7RV

@misc{pith2026250710951,
  author       = {Pith},
  title        = {Pith review of: Biological Processing Units: Leveraging an Insect Connectome to Pioneer Biofidelic Neural Architectures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SH4RC7RV}},
  note         = {Machine review of arXiv:2507.10951}
}

read the original abstract

The complete connectome of the Drosophila larva brain offers a unique opportunity to investigate whether biologically evolved circuits can support artificial intelligence. We convert this wiring diagram into a Biological Processing Unit (BPU), a fixed recurrent network derived directly from synaptic connectivity. Despite its modest size 3,000 neurons and 65,000 weights between them), the unmodified BPU achieves 98% accuracy on MNIST and 58% on CIFAR-10, surpassing size-matched MLPs. Scaling the BPU via structured connectome expansions further improves CIFAR-10 performance, while modality-specific ablations reveal the uneven contributions of different sensory subsystems. On the ChessBench dataset, a lightweight GNN-BPU model trained on only 10,000 games achieves 60% move accuracy, nearly 10x better than any size transformer. Moreover, CNN-BPU models with ~2M parameters outperform parameter-matched Transformers, and with a depth-6 minimax search at inference, reach 91.7% accuracy, exceeding even a 9M-parameter Transformer baseline. These results demonstrate the potential of biofidelic neural architectures to support complex cognitive tasks and motivate scaling to larger and more intelligent connectomes in future work.

Figures

Figures reproduced from arXiv: 2507.10951 by the authors.

**Figure 2.** (A) Test accuracy on MNIST and CIFAR-10 for the original connectomederived BPU. (B) CIFAR-10 test accuracy as a function of expansion factor for expanded BPUs via DCSBM. Shaded bands indicate average over five runs and are compared to a size-matched 2-layer MLP baseline. monotonically with expansion factor: a 2× graph already surpasses the original BPU, and performance continues to climb, remaining consistently a… view at source ↗

**Figure 3.** shows an ablation study that evaluates the contribution of different sensory modalities to image classification. Performance does not scale directly with neuron count, e.g., the respiratory group (26 neurons) outperforms the larger sight-related group (29 neurons) when trained with a small training sample size, highlighting the role of functional specificity. This may reflect evolved relevance of certain modalities… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Puzzle-solving accuracy (%) with GNN–BPU model and ChessBench reference models of multiple sizes. Despite having only 232,912 trainable parameters, the GNN–BPU converges even with small dataset size and achieves competitive or superior accuracy to the baselines. 3.2 C…

**Figure 5.** Figure 5: Bars show the percentage of puzzles solved correctly within each Elo bin. The legend indicates model type, parameter count, and overall accuracy. At equal scale (∼2M), CNN–BPU outperforms the Transformer baseline. With search, CNN–BPU surpasses even a 9M-parameter Tran…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 20 canonical work pages

[1]

The connectome of an insect brain.Science, 379(6636):eadd9330, 2023

Michael Winding, Benjamin D Pedigo, Christopher L Barnes, Heather G Patsolic, Youngser Park, Tom Kazimiers, Akira Fushiki, Ingrid V Andrade, Avinash Khan- delwal, Javier Valdes-Aleman, et al. The connectome of an insect brain.Science, 379(6636):eadd9330, 2023

work page 2023
[2]

The lottery ticket hypothesis: Finding sparse, trainable neural networks, 2019

Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks, 2019

2019
[3]

A natural lottery ticket winner: Reinforcement learning with ordinary neural cir- cuits

Ramin Hasani, Mathias Lechner, Alexander Amini, Daniela Rus, and Radu Grosu. A natural lottery ticket winner: Reinforcement learning with ordinary neural cir- cuits. In Hal Daumé III and Aarti Singh, editors,Proceedings of the 37th Inter- national Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pages 4082–4093. PMLR,...

work page 2020
[4]

Connectome- constrained networks predict neural activity across the fly visual system.Nature, 634:1132–1140, 2024

Juho K Lappalainen, Fabian D Tschopp, Sai Prakhya, et al. Connectome- constrained networks predict neural activity across the fly visual system.Nature, 634:1132–1140, 2024

work page 2024
[5]

Incorporating neuro-inspired adaptability for continual learning in artificial intelligence.Nature Machine Intelligence, 5:1356– 1368, 2023

Lei Wang, Xiaohui Zhang, Qian Li, et al. Incorporating neuro-inspired adaptability for continual learning in artificial intelligence.Nature Machine Intelligence, 5:1356– 1368, 2023

work page 2023
[6]

Ryali, Benjamin Hoover, Leopold Grinberg, Saket Navlakha, Mohammed J

Yuchen Liang, Chaitanya K. Ryali, Benjamin Hoover, Leopold Grinberg, Saket Navlakha, Mohammed J. Zaki, and Dmitry Krotov. Can a fruit fly learn word embeddings?, 2021

work page 2021
[7]

A completeelectronmicroscopyvolumeofthebrainofadultdrosophilamelanogaster

Zhihao Zheng, Jason S Lauritzen, Eric Perlman, Craig G Robinson, Matthew Nichols, Daniel Milkie, Oriol Torrens, Jackson Price, Conrad B Fisher, Nick Shar- ifi, Sarah A Calle-Schuler, Lenka Kmecova, Iman J Ali, Benjamin Karsh, Emily T Trautman, John A Bogovic, Philipp Hanslovsky, Gregory S X E Jefferis, Michael Kazhdan, Khaled Khairy, Stephan Saalfeld, Ric...

work page 2018
[8]

Lewis, Joel Veness, and Tim Ge- newein

Anian Ruoss, Grégoire Delétang, Sourabh Medapati, Jordi Grau-Moya, Li Kevin Wenliang, Elliot Catt, John Reid, Cannada A. Lewis, Joel Veness, and Tim Ge- newein. Amortized planning with large-scale transformers: A case study on chess. arXiv preprint arXiv:2402.04494, 2024

arXiv 2024

Show all 22 references

[9]

Why do we have so many excitatory neurons?bioRxiv, pages 2024–09, 2024

Qingyang Wang, Albert Cardona, Marta Zlatic, Joshua T Vogelstein, and Carey E Priebe. Why do we have so many excitatory neurons?bioRxiv, pages 2024–09, 2024

2024
[10]

Stochastic blockmodels and community structure in networks.Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 83(1):016107, 2011

Brian Karrer and Mark EJ Newman. Stochastic blockmodels and community structure in networks.Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 83(1):016107, 2011

2011
[11]

Community detection in degree-corrected block models.The Annals of Statistics, 2018

Chao Gao, Zongming Ma, Anderson Y Zhang, and Harrison H Zhou. Community detection in degree-corrected block models.The Annals of Statistics, 2018

2018
[12]

Playing chess with large language models, 2023

Nicholas Carlini. Playing chess with large language models, 2023

2023
[13]

Standard portable game notation specification and implementation guide

SJ Edwards, SD Forsyth, J Stanback, and A Saremba. Standard portable game notation specification and implementation guide. 1994.URL https://ia902908. us. archive. org/26/items/pgn-standard-1994-03-12/PGN standard, pages 03–12, 1994

1994
[14]

torch_geometric.nn.conv.GINEConv — pytorch geometric 2.5.1 documentation, 2024

PyTorch Geometric Development Team. torch_geometric.nn.conv.GINEConv — pytorch geometric 2.5.1 documentation, 2024

2024
[15]

Claude E. Shannon. Programming a computer for playing chess. Philosophical Magazine, 41:256–275, 1950. Introduced the Minimax algorithm in chess. 10 Yu et al

1950
[16]

Superior computer chess with model predictive control, reinforcement learning, and rollout, 2024

Atharva Gundawar, Yuchao Li, and Dimitri Bertsekas. Superior computer chess with model predictive control, reinforcement learning, and rollout, 2024

2024
[17]

Knuth and Ronald W

Donald E. Knuth and Ronald W. Moore. An analysis of alpha-beta pruning. Artificial Intelligence, 6(4):293–326, 1975

1975
[18]

Department of Computer Science, Carnegie-Mellon University, 1973

Samuel H Fuller, John G Gaschnig, JJ Gillogly, et al.Analysis of the alpha-beta pruning algorithm. Department of Computer Science, Carnegie-Mellon University, 1973

1973
[19]

Generativenetworkmodelingrevealsquantitative definitions of bilateral symmetry exhibited by a whole insect brain connectome

Benjamin D Pedigo, Mike Powell, Eric W Bridgeford, Michael Winding, Carey E Priebe,andJoshuaTVogelstein. Generativenetworkmodelingrevealsquantitative definitions of bilateral symmetry exhibited by a whole insect brain connectome. Elife, 12:e83739, 2023

2023
[20]

Azeem Khadam

Alfred Geroldinger and M. Azeem Khadam. On the arithmetic of monoids of ideals, 2021

2021
[21]

Vishwanathan, A

A. Vishwanathan, A. Sood, J. Wu, et al. Predicting modular functions and neural coding of behavior from a synaptic wiring diagram.Nature Neuroscience, 27:2443– 2454, 2024

2024
[22]

P. K. Shiu, G. R. Sterne, N. Spiller, et al. A drosophila computational brain model reveals sensorimotor processing.Nature, 634:210–219, 2024

2024

Pith tools

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

[1] [1]

The connectome of an insect brain.Science, 379(6636):eadd9330, 2023

Michael Winding, Benjamin D Pedigo, Christopher L Barnes, Heather G Patsolic, Youngser Park, Tom Kazimiers, Akira Fushiki, Ingrid V Andrade, Avinash Khan- delwal, Javier Valdes-Aleman, et al. The connectome of an insect brain.Science, 379(6636):eadd9330, 2023

work page 2023

[2] [2]

The lottery ticket hypothesis: Finding sparse, trainable neural networks, 2019

Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks, 2019

2019

[3] [3]

A natural lottery ticket winner: Reinforcement learning with ordinary neural cir- cuits

Ramin Hasani, Mathias Lechner, Alexander Amini, Daniela Rus, and Radu Grosu. A natural lottery ticket winner: Reinforcement learning with ordinary neural cir- cuits. In Hal Daumé III and Aarti Singh, editors,Proceedings of the 37th Inter- national Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pages 4082–4093. PMLR,...

work page 2020

[4] [4]

Connectome- constrained networks predict neural activity across the fly visual system.Nature, 634:1132–1140, 2024

Juho K Lappalainen, Fabian D Tschopp, Sai Prakhya, et al. Connectome- constrained networks predict neural activity across the fly visual system.Nature, 634:1132–1140, 2024

work page 2024

[5] [5]

Incorporating neuro-inspired adaptability for continual learning in artificial intelligence.Nature Machine Intelligence, 5:1356– 1368, 2023

Lei Wang, Xiaohui Zhang, Qian Li, et al. Incorporating neuro-inspired adaptability for continual learning in artificial intelligence.Nature Machine Intelligence, 5:1356– 1368, 2023

work page 2023

[6] [6]

Ryali, Benjamin Hoover, Leopold Grinberg, Saket Navlakha, Mohammed J

Yuchen Liang, Chaitanya K. Ryali, Benjamin Hoover, Leopold Grinberg, Saket Navlakha, Mohammed J. Zaki, and Dmitry Krotov. Can a fruit fly learn word embeddings?, 2021

work page 2021

[7] [7]

A completeelectronmicroscopyvolumeofthebrainofadultdrosophilamelanogaster

Zhihao Zheng, Jason S Lauritzen, Eric Perlman, Craig G Robinson, Matthew Nichols, Daniel Milkie, Oriol Torrens, Jackson Price, Conrad B Fisher, Nick Shar- ifi, Sarah A Calle-Schuler, Lenka Kmecova, Iman J Ali, Benjamin Karsh, Emily T Trautman, John A Bogovic, Philipp Hanslovsky, Gregory S X E Jefferis, Michael Kazhdan, Khaled Khairy, Stephan Saalfeld, Ric...

work page 2018

[8] [8]

Lewis, Joel Veness, and Tim Ge- newein

Anian Ruoss, Grégoire Delétang, Sourabh Medapati, Jordi Grau-Moya, Li Kevin Wenliang, Elliot Catt, John Reid, Cannada A. Lewis, Joel Veness, and Tim Ge- newein. Amortized planning with large-scale transformers: A case study on chess. arXiv preprint arXiv:2402.04494, 2024

arXiv 2024