From the Drosophila Visual Connectome to General-Purpose Computer Vision

A connectome-informed backbone separates reusable circuit organization from task interfaces

• FlyVision uses parallel learned input branches followed by three recurrent graph-mixing stages, global pooling and a task-specific head.The backbone scales channel widths and embedding size while retaining this shared computation.Source: A connectome-informed backbone… · p. 3
• The shared design supports compact 2D models, 224×224 ImageNet training, checkpoint-based radiography transfer and multi-view MRI analysis.BrainAGE applies the pretrained encoder independently to eight sagittal, eight coronal and eight axial slices per scan.Source: A connectome-informed backbone… · p. 3

Compact recognition on MNIST

• 99.34% top-1 accuracy with 80,608 parameters placed FlyVision within 0.04 percentage points of the highest observed MNIST accuracy, while ResNet18 used about 139 times as many parameters.ResNet18 with ImageNet initialization reached 99.38%; scratch ResNet18 reached 99.22%.Source: Compact recognition on MNIST · p. 3
• LeNet is smaller than FlyVision but less accurate, whereas ResNet18 is substantially larger for only a small absolute accuracy increase.This comparison positions FlyVision as a high-accuracy, compact, end-to-end-trainable model in the low-resolution regime.Source: Compact recognition on MNIST · p. 3

Natural-image recognition reveals architecture-dependent transfer on CIFAR-10

• From scratch on CIFAR-10, ResNet18 reached 83.44% top-1 accuracy and FlyVision reached 78.03%, ahead of LeNet and the MLP.Direct ImageNet initialization raised ResNet18 to 85.84%, the highest observed CIFAR-10 accuracy in the benchmark.Source: Natural-image recognition reveals… · p. 3
• MNIST initialization changed FlyVision CIFAR-10 accuracy from 78.03% to 77.62% (-0.41 percentage points), while transfer effects differed across architectures and source tasks.Direct ImageNet initialization increased ResNet18 from 83.44% to 85.84% (+2.40 points); sequential ImageNet-to-MNIST-to-CIFAR initialization reached 84.04%.Source: Natural-image recognition reveals… · p. 4

ImageNet-1K establishes large-scale visual recognition

• FlyVision Base and Large global-surround models reached 60.79% and 66.25% top-1 accuracy on ImageNet-1K, respectively.Their paired top-5 accuracies were 82.33% and 86.07%.Source: ImageNet-1K establishes large-scale… · p. 4
• 66.53% top-1 accuracy was reached by the 3.744-million-parameter FlyVision Large local-k7 model with a learned low-frequency branch, versus 69.25% for 11.690-million-parameter ResNet18.The profiled FlyVision Large model required 14.803 GFLOPs, compared with 3.628 GFLOPs for ResNet18, separating parameter efficiency from computational cost.Source: ImageNet-1K establishes large-scale… · p. 4

Skin-disease imaging tests biomedical transfer outside radiography

• On SkinDisease, Large local-k7 with a learned low-frequency branch reached 63.78% held-out accuracy, while Large global achieved 95.59% macro-AUROC; ResNet18 reached 66.19% accuracy and 95.96% macro-AUROC.Source: Skin-disease imaging tests biomedical… · p. 4
• The prepared image-level split contained 11,506 training, 2,042 validation and 1,538 test images across 22 labeled classes.The source collection contained 15,444 images; six evaluation exports shared identical test keys, labels and class mappings.Source: Skin-disease imaging tests biomedical… · p. 4
• Evaluation estimates describe held-out image-level generalization because patient and lesion identifiers were unavailable in the metadata.Five runs used resize/rotation jitter, while the Large local-k7/low-frequency run used spatial affine augmentation.Source: Skin-disease imaging tests biomedical… · p. 4

FlyVision occupies a favorable accuracy-parameter regime on four-class chest radiography

• The benchmark used a deterministic class-stratified 70:15:15 image-level split of 21,165 image-mask pairs.The paired lung mask was applied in image space before ImageNet normalization.Source: FlyVision occupies a favorable… · pp. 4–5
• 92.76% accuracy was reached by ImageNet-pretrained FlyVision Large on the held-out test set, compared with 91.56% for ImageNet-pretrained ResNet18.FlyVision Base reached 92.60% accuracy; its random-initialized version reached 92.19%.Source: FlyVision occupies a favorable… · p. 5
• FlyVision Base used 1.33 million parameters and approximately 1.28 GFLOPs per image, versus 11.18 million parameters and approximately 3.63 GFLOPs for ResNet18.Within this benchmark, Base's point estimates were 1.04 percentage points higher for accuracy and 0.77 points higher for macro-F1.Source: FlyVision occupies a favorable… · p. 5
• Across the stronger models, the largest residual ambiguity was between lung opacity and normal radiographs, while viral pneumonia was comparatively well separated.COVID-19 sensitivity was 0.895 for FlyVision Base and 0.923 for FlyVision Large, compared with 0.862 for ResNet18.Source: FlyVision occupies a favorable… · p. 5

Hierarchical decomposition localizes the principal radiography errors

• The hierarchy separates normal-versus-disease detection from disease-subtype classification, or shares a backbone between binary and subtype heads.The two-stage system reconstructs final four-class predictions from the binary and subtype outputs.Source: Hierarchical decomposition localizes… · p. 5
• 92.60% accuracy and 92.75% macro-F1 for the flat model exceeded the two-stage and multi-head hierarchical results.The two-stage model reached 91.53% accuracy and 91.68% macro-F1; the multi-head model reached 91.75% and 92.17%.Source: Hierarchical decomposition localizes… · p. 5
• The subtype models achieved 95.02% accuracy and 95.58% macro-F1 for disease-only classification, while routing errors concentrated at normal-versus-disease decisions.True normal images were most often routed toward lung opacity.Source: Hierarchical decomposition localizes… · p. 5

Multi-view FlyVision extends the pretrained visual representation to volumetric brain-age regression

• BrainAGE encodes 24 slices with an ImageNet-pretrained FlyVision Large backbone, then combines confidence-modulated, Gaussian-position-weighted votes within each axis and averages the three axis estimates.The test split contained 433 scans, with participant-grouped partitioning.Source: Multi-view FlyVision extends the… · p. 5
• Three-axis BrainAGE fusion achieved 5.98 years MAE and R2 = 0.868, reducing MAE by 1.45–1.50 years versus individual axes.Axis-specific MAEs were 7.47 years sagittal, 7.44 coronal and 7.49 axial on the held-out set.Source: Multi-view FlyVision extends the… · p. 6

• One circuit principle remains trainable from 28×28 grayscale digits through 32×32 natural images and 224×224 large-scale or biomedical images.ImageNet-pretrained FlyVision Large is also reused across anatomical slices for scan-level brain-age regression from three-dimensional T1w MRI.Source: Cross-task synthesis · p. 6
• The sequence tests compact recognition, transfer, 1,000-class recognition, two biomedical image regimes, hierarchical routing and multi-view continuous prediction.The benchmarks include MNIST, CIFAR-10, ImageNet-1K, skin-disease classification, chest radiography and BrainAGE.Source: Cross-task synthesis · p. 6
• Parameter count, computation, predictive quality and error structure provide complementary views of how the connectome-informed design scales.Source: Cross-task synthesis · p. 6

• One computational motif persists across digit, natural-image, ImageNet, biomedical classification and volumetric structural-MRI tasks as model capacity scales.The shared design preserves parallel input pathways, recurrent processing and population-level interaction.Source: Discussion · p. 6
• 99.34% MNIST accuracy used 80,608 parameters, while CIFAR-10 FlyVision reached 78.03% and MNIST initialization changed its accuracy by -0.41 percentage points.Transfer effects differed by architecture: the same source task improved LeNet and the MLP.Source: Discussion · p. 6
• 66.53% ImageNet-1K top-1 accuracy for Large local-k7/low-frequency remained below ResNet18's 69.25%, with approximately 3.74 million versus 11.690 million parameters.FlyVision required more Conv/Linear operations than ResNet18, distinguishing parameter efficiency from computational efficiency.Source: Discussion · p. 6
• On the 22-class skin benchmark, FlyVision Large local-k7 with a learned low-frequency branch scored 63.78% accuracy and 95.28% macro-AUROC, versus ResNet18’s 66.19% and 95.96%; chest FlyVision Base and Large reached 92.60% and 92.76% accuracy.The chest models were ImageNet-pretrained, with 1.33 and 2.97 million parameters, respectively.Source: Discussion · pp. 6–7
• Three-axis BrainAGE fusion achieved 5.98 years MAE and R2 = 0.868 on held-out T1w scans, using confidence-modulated Gaussian voting across anatomical slices.Axis-specific MAEs were 7.47, 7.44 and 7.49 years.Source: Discussion · p. 7
• The study frames sparse topology, recurrence, ON/OFF segregation, population interaction and connection sign as independently perturbable design components.The benchmark suite supplies reference conditions spanning performance, transfer, errors, routing, calibration and multi-view fusion.Source: Discussion · p. 7
• Skin and radiography results measure held-out image-level performance; patient-, site- and scanner-separated evaluation would provide a complementary deployment-generalization test.The skin collection lacks patient or lesion identifiers, while radiography uses deterministic image-level stratification.Source: Discussion · p. 7
• The paper treats biological circuit structure as experimentally grounded architectural hypotheses tested through cross-domain scaling, transfer and perturbation.Source: Discussion · p. 7

FlyVision architecture

• The two-dimensional FlyVision implementation uses a dual ON/OFF stem, three recurrent graph-mixing stages, adaptive pooling, an embedding and a task-specific head.The stem uses global per-channel means or local-k7 means; “+ LF” adds a learnable low-frequency convolutional pathway.Source: FlyVision architecture · p. 7

MNIST and CIFAR-10 datasets

• MNIST and CIFAR-10 use 16/32/64-channel stages, 1/2/2 recurrent steps and deterministic class-stratified validation splits.The datasets contain 60,000/10,000 MNIST train/test images and 50,000/10,000 CIFAR-10 images; CIFAR-10 training uses padded random crops and horizontal flips.Source: FlyVision architecture · p. 8, MNIST and CIFAR-10 datasets · p. 8

MNIST/CIFAR-10 optimization, transfer and evaluation

• Classification runs use 100 epochs, batch size 128 and AdamW; CIFAR-10 transfer initializes compatible states from MNIST and resets the target classifier.Scratch-versus-transfer comparisons use paired predictions and exact McNemar tests with Holm adjustment.Source: MNIST/CIFAR-10 optimization, transfer… · p. 8

ImageNet-1K pretraining

• ImageNet-1K checkpoints were trained at 224×224 for 90 scheduled epochs with augmentation, warm-up and cosine decay.The reported global-surround models were trained from scratch; ResNet18 was also trained under the common launcher.Source: ImageNet-1K pretraining · p. 8

Skin-disease dataset and evaluation

• The skin benchmark uses 15,444 images across 22 source classes and reports image-level generalization because patient or lesion identifiers are unavailable.Duplicate-content screening and label-conflict handling produced a prepared 11,506/2,042/1,538 train/validation/test split.Source: Skin-disease dataset and evaluation · p. 9

Chest-radiography preprocessing and optimization

• Radiography uses a deterministic 70/15/15 image split; images and paired masks are resized, jointly augmented and lung-masked before normalization.Flat four-class models use class-weighted cross-entropy and validation macro-F1 early stopping.Source: Chest-radiography dataset and split · p. 9, Chest-radiography preprocessing and… · p. 9

Hierarchical radiography models

• Radiography FlyVision models load ImageNet checkpoint tensors into downstream reconstructions before replacing the classifier; hierarchical variants combine binary detection and subtype predictions.Operation counts reflect the downstream forward schedule, which differs from ImageNet; the hierarchy also includes a shared-backbone multi-head formulation.Source: Radiography model initialization and… · pp. 9–10, Hierarchical radiography models · p. 10

BrainAGE cohort, preprocessing and multi-view voting

• BrainAGE keeps repeated scans within participant-grouped 70:15:15 splits (seed 2026), samples 24 T1-weighted slices per scan, and uses a shared FlyVision Large encoder with confidence- and position-weighted axis voting.Training combines scan-level Huber loss with 0.25-times Gaussian-weighted slice-level Huber loss. Confidence intervals for MAE, RMSE, R2, and Pearson correlation use 1,000 bootstrap resamples.Source: BrainAGE cohort, preprocessing and… · pp. 10–11

• The study used secondary analyses of existing image datasets, with no prospective recruitment or new human-subject data collection.Source: Ethics and data governance · p. 11
• Skin-disease files lacked patient or lesion identifiers, so the analysis did not perform patient-level grouping, linkage, or re-identification.Source: Ethics and data governance · p. 11
• Neuroimaging records are not redistributed by the authors, and source data use remains governed by the originating repositories’ access and data-use terms.Source: Ethics and data governance · p. 11

• MNIST, CIFAR-10, ImageNet-1K, radiography, skin-disease, and neuroimaging datasets are available from their respective repositories subject to source-specific access, licensing, and data-use conditions.Source: Data availability · p. 11
• Restricted BrainAGE participant-level data must be obtained from originating repositories under applicable data-use agreements; the skin-disease images are available through Kaggle.Source: Data availability · p. 11
• ConnectomeX source code and author-generated reproduction resources are proprietary, and findings-supporting data may be requested from the corresponding author where third-party terms permit.Source: Data availability · p. 11

Figures 1–2

• Figure 1 maps connectome populations and feed-forward, recurrent, and feedback connections to FlyVision’s ON/OFF backbone and cross-domain tests, while Figure 2 summarizes the architecture family and task progression.Source: Figure 1 · p. 15, Figure 2 · p. 16

Figures 3–4

• 99.34% top-1 accuracy with 80,608 parameters is reported for scratch FlyVision on MNIST; the figures also plot accuracy against macro-F1 and parameter count.Source: Figure 4 · p. 17, Figure 3 · p. 17

Figures 5–6

• Figure 5 plots held-out CIFAR-10 accuracy against macro-F1 across scratch and transfer conditions, while Figure 6 shows top-1 accuracy and transfer-induced changes, including +2.40 points for ResNet18 and -0.41 points for FlyVision.Source: Figure 6 · p. 18, Figure 5 · p. 18

Figure 7

• Figure 7 compares ImageNet-1K validation top-1 with paired top-5 accuracy and uses bubble size for parameter count, including the Large local-k7/low-frequency checkpoint evaluated on all 50,000 validation images.Source: Figure 7 · p. 19

Figures 8–9

• Figure 8 compares skin-disease test accuracy and macro-F1 across six models, with accuracy intervals and parameter-sized bubbles; Figure 9 shows the 22-class confusion matrix and reports 63.78% overall accuracy.Source: Figure 8 · p. 19, Figure 9 · p. 20

Figures 10–11 and 13

• Chest-radiography figures show held-out accuracy, macro-F1, AUROC, AUPRC, and efficiency against parameter count, compute, latency, and throughput; intervals summarize bootstrap uncertainty.Source: Figure 10 · p. 21, Figure 11 · p. 21, Figure 13 · p. 22

Figures 12 and 14–15

• Further radiography figures compare classwise F1, recall, specificity, confusion matrices, hierarchical systems, and subtype-routing errors, including false-positive destinations for true-normal images.Source: Figure 12 · p. 22, Figure 14 · p. 23, Figure 15 · p. 24

Figures 16–17

• Three-axis BrainAGE fusion on 433 held-out scans achieved MAE 5.98 years and R^2 = 0.868; the figures also show predicted-versus-chronological age, agreement, axis-specific errors, and eight-slice axis estimates.Source: Figure 16 · p. 24, Figures · p. 24, Figure 17 · p. 25

Table 1

• Table 1 reports MNIST test performance.Source: Table 1 · p. 26

Table 2

• Table 2 reports CIFAR-10 test performance alongside transfer conditions.Source: Table 2 · p. 26

Table 3

• Table 3 reports ImageNet-1K validation performance at best-top-1 checkpoints, pairing top-5 with the selected top-1 checkpoint and defining the forward-operation counting convention.Source: Table 3 · p. 26

Table 4

• Table 4 covers the identical 1,538-image skin-disease test set and defines macro-AUROC as the unweighted mean of 22 one-versus-rest class AUROCs.Source: Table 4 · p. 26

Table 5

• Table 5 reports chest-radiography performance and efficiency on a shared 3,175-image test split, with task-specific graph conventions for GFLOPs and one-versus-rest macro-AUROC.Source: Table 5 · p. 27

Table 6

• Table 6 reports hierarchical chest-radiography performance.Source: Table 6 · p. 27

Table 7

• Table 7 reports BrainAGE held-out regression values, including bootstrap confidence intervals for primary three-axis fusion metrics and fixed held-out errors for individual axes.Source: Table 7 · p. 27