A new study of a connectome-constrained model of the fruit fly visual system has uncovered a striking limit to neural scale stability: a frozen visual representation can preserve motion direction across a broad range of image sizes, but its ability to generalize collapses abruptly when stimuli move beyond the scale range encountered during training. The work, published in Neuroinformatics, suggests that biological plausibility and computational flexibility do not automatically produce unlimited scale invariance. Instead, the model appears to maintain a highly reliable internal code within a familiar visual range while failing at an upper boundary, offering a rare and precise demonstration of how generalization can break inside a mechanistic brain model.
The research focused on FlyVis, a computational representation of the Drosophila visual system whose architecture is constrained by anatomical and connectivity data from the fly brain. FlyVis contains pathways corresponding to known visual circuits, including neurons associated with motion processing and the direction-selective T4 and T5 pathways. Rather than training the model from scratch to recognize motion, the researchers used an independently pretrained FlyVis checkpoint and froze its parameters. They then asked whether a simple linear decoder could recover the direction of a moving stimulus from the model’s activity when one apparent retinal scale was deliberately withheld during training. This design is important because it tests the transferability of an existing representation rather than measuring whether a newly optimized classifier can simply memorize six direction classes.
The investigators constructed a scale-varying stimulus set containing 192 unique rendered conditions. The stimuli included moving bars and translating targets presented at eight apparent scales, with motion directions serving as the classification labels. Each condition represented a distinct combination of visual form, scale, and direction, allowing the analysis to be conducted at the level of unique rendered examples rather than treating repeated measurements as independent evidence. The central evaluation used a leave-one-apparent-scale-out protocol. For each test, the decoder was trained on direction-labeled examples from the other seven scales and then evaluated on the entirely held-out scale. Geometrically scale-neutral moving edges served as a control, helping distinguish failures caused by scale changes from failures caused by motion direction itself.
Across all eight held-out scales, the frozen FlyVis representation supported an accuracy of 0.885, with a Wilson 95 percent confidence interval ranging from 0.833 to 0.923. The overall result indicates that direction information remains strongly and linearly accessible even when the decoder has not seen the test scale during supervision. But the average concealed an unusually sharp boundary. Across the six interior scales, transfer was nearly perfect, reaching approximately 0.986 accuracy. At the upper boundary, corresponding to an apparent scale of 24, performance fell to 0.167, essentially the chance level for a six-way direction problem. The model therefore did not exhibit a smooth, gradual decline as stimuli became larger. Instead, it showed a sudden extrapolation failure after maintaining an exceptionally stable code across the range between the training scales.
That pattern has broad implications for how researchers interpret scale invariance in neural systems. In computer vision, a representation is often described as scale-stable when a classifier can recognize an object or feature despite changes in size. Yet a model can appear highly invariant when tested within the distribution of scales it has effectively experienced, while remaining unable to extrapolate outside that distribution. The FlyVis results make this distinction unusually visible. The representation appears to organize motion direction consistently over its interior operating range, but the organization does not extend indefinitely. The finding resembles a visual system with a calibrated working range: within that range, different apparent sizes map onto compatible activity patterns, while beyond it the mapping becomes incompatible with the decoder’s learned decision boundaries.
To determine whether the strong results were simply caused by the stimulus pixels or by an especially favorable sampling pattern, the researchers trained capacity- and sampling-matched temporal models directly on direction labels. These comparison systems included a small temporal residual convolutional network and models using alternative neighbourhood operators. The tested operators included a hexagonal arrangement, a collision-free square-projected neighbourhood, and a self-only configuration. Their performance ranged from 0.983 to 0.986, demonstrating that the raw visual task itself was readily solvable under the experimental conditions. At the same time, the controls showed that success could not be attributed solely to one particular sampling geometry. The FlyVis model’s behavior therefore reflects more than the existence of a learnable motion signal; it reflects the interaction between pretrained optical-flow-related representations, connectome-constrained dynamics, and the specific parameters that shape the model’s internal activity.
The study also examined what happens when the T4 and T5 motion pathways are silenced. This intervention produced a revealing separation between information availability and coordinate stability. When the researchers refitted the decoder after silencing T4 and T5, accuracy remained high at 0.891. However, when they applied the original decoder trained on intact activity, performance dropped to 0.193. In other words, the information needed to determine direction was not eliminated completely, but it was redistributed into a new activity configuration that the old readout could no longer interpret. This distinction is central to modern neural decoding. A drop in performance from a fixed decoder does not necessarily mean that a representation has lost the underlying variable; it may instead mean that the variable survives in reorganized coordinates. The perturbation results therefore point to a distributed direction code that depends on T4/T5 activity for preserving its original linear geometry, even though other network activity can compensate when the decoder is allowed to adapt.
Additional learned-parameter shuffles revealed that the model’s behavior was also sensitive to the particular realization of its parameters. Randomly rearranging learned values disrupted the direction code in some cases and destabilized the resulting dynamics, indicating that performance was not determined by connectivity alone. The anatomical wiring diagram provides the network’s structural constraints, but the learned parameters determine how signals propagate, interact, and settle into task-relevant patterns. This result reinforces a growing theme in connectome-based neuroscience: a wiring diagram is not a complete explanation of computation. Two networks can share a topology while producing different representations because their synaptic strengths, temporal constants, nonlinearities, and training histories differ. In FlyVis, these factors appear to work together to create a direction representation that is robust across interior scales but fragile under certain perturbations and extrapolations.
The authors present the experiment as a reusable benchmark for studying cross-scale representation in mechanistic visual models. Its value lies not only in the headline accuracy, but in the separation of several questions that are often conflated. The frozen-checkpoint experiment tests whether an already formed representation transfers to an unseen scale. The directly trained baselines test whether the stimulus and sampling conditions make the task solvable. The T4/T5 intervention tests whether particular motion pathways preserve the original readout geometry or merely contribute information that can be reconstructed elsewhere. The parameter-shuffle experiments test whether the result depends on the learned realization of the network. Together, these comparisons provide a more detailed picture of generalization than a single classification score could offer.
The findings may ultimately help connect artificial vision, systems neuroscience, and the study of biological robustness. They suggest that scale stability should be reported as a function of the tested range rather than summarized by one aggregate number. A model can achieve nearly perfect transfer across most conditions and still fail catastrophically at one boundary, a pattern that could remain hidden in conventional train-and-test evaluations. For FlyVis, optical-flow pretraining, anatomical constraints, and learned dynamics emerge as complementary ingredients whose individual contributions can be isolated in future factorial studies. The broader lesson is both encouraging and cautionary: a connectome-constrained model can preserve a remarkably clean motion code across unseen visual scales, but that stability is range-limited. Neural representations may be flexible not because they are universally invariant, but because they are exquisitely organized for the environments and stimulus ranges in which they operate.
Subject of Research: Scale stability and motion-direction representations in a connectome-constrained computational model of the fruit fly visual system
Article Title: Range-Limited Scale Stability of Dynamic Direction Representations in a Connectome-Constrained Fly Visual Model
Article References: Dhiman, N., & Panwar, S. “Range-Limited Scale Stability of Dynamic Direction Representations in a Connectome-Constrained Fly Visual Model.” Neuroinformatics 24, Article 54 (2026). Key related references include Lappalainen et al., “Connectome-constrained networks predict neural activity across the fly visual system,” Nature 634, 1132–1140 (2024); Dorkenwald et al., “Neuronal wiring diagram of an adult brain,” Nature 634, 124–138 (2024); and TuragaLab, “FlyVis: A connectome-constrained deep mechanistic network model of the fruit fly visual system in PyTorch” (2026).
Image Credits: AI Generated
DOI: 10.1007/s12021-026-09811-3
Keywords: FlyVis, Drosophila visual system, connectome-constrained networks, motion direction, scale invariance, neural representations, T4/T5 neurons, optical-flow pretraining, neural decoding, computational neuroscience, visual motion, representation transfer, leave-one-scale-out evaluation
