In a groundbreaking advance at the crossroads of neuroscience and artificial intelligence, researchers have devised a novel metric that promises to revolutionize how we compare artificial neural networks (ANNs) with the biological brains they aim to emulate. Traditionally, the assessment of these computational models has centered on forward predictivity—how well features within an ANN can predict neural responses recorded from primate brains. However, this approach captures only one side of a complex relationship, neglecting whether neural activity can predict the computational units within ANNs. This oversight may conceal fundamental representational discrepancies between artificial and biological networks.
The innovative study, conducted by Muzellec and Kar and published in Nature Machine Intelligence, introduces what they describe as “reverse predictivity.” This diagnostic metric quantifies the extent to which activity in the macaque inferior temporal (IT) cortex can predict the activations of ANN units. By measuring this bidirectional predictability, the team reveals a striking asymmetry that challenges prevailing assumptions. While numerous ANNs achieve impressive forward predictivity levels, explaining roughly 50% of the variance in neural responses, many of their units are inexplicably unpredictable when the causal direction is flipped. This finding suggests that these ANNs contain dimensions inaccessible or irrelevant to biological neural processing.
This asymmetry in predictability lays bare a gap in our modeling frameworks. While ANNs are often evaluated on how well they mirror brain signals in a forward direction—from artificial to biological—the reverse comparison exposes ‘hidden’ features of ANNs that do not correspond to any known brain representation. Importantly, the study shows that monkey-to-monkey comparisons yield symmetric predictivity, reaffirming that the observed asymmetry with ANNs is not a methodological artifact but a genuine disjunction between species and systems. By anchoring the comparison to empirical neural data, this work establishes a new standard for evaluating computational brain models.
The implications of reverse predictivity extend far beyond metrics. By identifying ANN units that are ‘common’ to the biological system—units whose activation patterns are well predicted from neural activity—the researchers delineate shared representational subspaces. These common units are behaviorally relevant and generalize across species boundaries, suggesting they embody functionally meaningful features in visual recognition. Conversely, unique units that cannot be predicted from biological signals may represent spurious or artificial constructs optimized for computational tasks but lacking biological plausibility.
At the heart of these phenomena lies the influence of feature dimensionality within models. The research reveals that models embedding higher-dimensional feature spaces tend to harbor more unpredictable units from the neural perspective. This owes to the fact that increasing dimensionality can introduce abstract representations that do not map cleanly onto the brain’s internal coding. Training objectives, too, play a salient role—models trained under standard supervised learning regimes showed distinct patterns of predictivity compared with those trained with adversarial robustness or alternative loss functions. This invites a reexamination of how we sculpt training criteria to achieve not only performance but neurobiological congruence.
The authors emphasize that reverse predictivity is a conservative and diagnostic tool, one that can serve as a compass for the next generation of ANNs. By harnessing this metric, AI researchers and neuroscientists alike can finely tune architectures and learning protocols to align artificial features more closely with the brain’s representational geometry. Ultimately, such alignment is envisioned to elevate both the interpretability and generalizability of models, paving the way for systems that robustly mimic biological vision in both function and mechanistic detail.
Their findings also resonate deeply with ongoing debates about the nature of neural representations. Are the features extracted by deep networks merely statistical constructs, or do they reflect computational principles shared by evolutionarily honed brains? The asymmetry in predictivity highlighted here suggests that many conventional ANNs harbor representational elements void of direct biological analog. This insight fuels the imperative to prioritize architectures and learning paradigms that yield reciprocal predictability, ensuring artificial systems capture the bidirectional logic of brain processing.
The reverse predictivity approach further offers a framework to probe how adversarial robustness—a model’s ability to resist deceptive perturbations—affects representational alignment. Models trained to withstand adversarial attacks often deviate from naturalistic neural codes, reflecting an intriguing trade-off between robustness and biological fidelity. Quantifying this trade-off with reverse predictivity equips the field with an empirical lens to systematically evaluate competing design choices.
From a methodological standpoint, the study harnesses cutting-edge electrophysiological recordings from macaque IT cortex, a brain region critical for high-level visual object recognition. By coupling these high-resolution neural signals with state-of-the-art deep convolutional networks, the team conducts a rigorous examination of cross-system representational geometry. This integrative approach—balancing computational modeling with in vivo brain data—embodies the contemporary ethos of systems neuroscience, wherein interdisciplinary tools unlock new vistas on cognition.
Moreover, the study underscores the value of bidirectional comparisons in computational neuroscience. Whereas previous paradigms treated network-to-brain predictions as a unidirectional task, incorporating reverse mappings compels the field to confront the full complexity of model-brain relationships. This conceptual expansion is poised to recalibrate how AI-driven neuroscience research unfolds, emphasizing mutual predictability over unilateral fits.
Beyond primates, the observation that certain ANN units generalize across species foreshadows exciting translational possibilities. As common representational motifs bridge humans, monkeys, and machines, emerging models may serve as universal platforms for dissecting perceptual computations. Such models can catalyze cross-species investigations into vision, guiding research that spans animals, humans, and synthetic agents alike.
In parallel, reverse predictivity invites a redefinition of what constitutes success in brain modeling. Rather than narrow goals centered exclusively on task accuracy or forward neural correlations, future efforts might prioritize models exhibiting bidirectional representational reciprocity. These models promise deeper explanatory power—capturing not only how brains give rise to behavior but how artificial systems can faithfully embody brain function.
As the field progresses, this paradigm shift encourages closer integration of developmental and learning theories with architecture design. Understanding how brain-like representations emerge through experience and evolution will inform the crafting of ANN training regimes optimized for reverse predictivity. This iterative cycle of hypothesis, modeling, and empirical testing stands to accelerate breakthroughs in both AI and neuroscience.
Ultimately, this study showcases that achieving biological plausibility requires transcending standard metrics and embracing nuanced comparisons. By illuminating the hidden representational gaps in current ANNs, Muzellec and Kar sow the seeds for refined models attuned to the brain’s intrinsic logic. Their bidirectional metric charts a path toward systems that no longer merely mimic, but authentically emulate the multifaceted computations of sensory cortex.
As artificial intelligence continues to permeate science and society, grounding these systems in the substrate of biological reality gains urgency. Aligning models not only with behavioral output but with the latent neural substrates generating cognition promises a new era of transparent, interpretable AI. Through reverse predictivity, the authors provide the field with a vital tool to realize this vision—a transformative stride in decoding the language shared by brains and machines.
Subject of Research: Computational models of neural representations in primate inferior temporal cortex and their comparison with artificial neural networks.
Article Title: Reverse predictivity for bidirectional comparison of neural networks and biological brains.
Article References:
Muzellec, S., Kar, K. Reverse predictivity for bidirectional comparison of neural networks and biological brains. Nat Mach Intell 8, 474–488 (2026). https://doi.org/10.1038/s42256-026-01204-0
Image Credits: AI Generated
DOI: March 2026
