A neuronal basis for mental imagery

www.nature.com/cr www.cell-research.com Cell Research RESEARCH HIGHLIGHT A neuronal basis for mental imagery Stephen M. Fleming1,2,3 ✉ and Nadine Dijkstra4 © The Author(s) under exclusive licence to Center for Excellence in Molecular Cell Science, Chinese Academy of Sciences 2026 1234567890();,: Cell Research (2026) 0:1–2; https://doi.org/10.1038/s41422-026-01260-6 Wadia et al. provide rare single-neuron evidence that mental imagery reactivates a perceptual code in the human ventral temporal cortex, with a substantial subset of object-tuned neurons showing shared tuning across seeing and imagining. By characterizing neuronal activity in a feature space derived from a deep neural network, the study demonstrates that imagery recapitulates the structure of perceptual representations. When sitting at our desks in the depths of the British winter, it is tempting to call to mind an image of a Mediterranean holiday: the warm sun on our faces, a cool drink in hand. This ability to internally activate mental representations — to create and recombine a range of previous experiences — is thought to undergird artistic and creative endeavours and is a central feature of planning and autobiographical memory. Conversely, overactive mental imagery may contribute to psychological disturbances including hallucinations and intrusive memories. The neural basis of mental imagery has been at the core of the imagery debate — is the format of imagery more perception-like (iconic) or language-like (propositional)?1,2 Neuroimaging research has made progress on this question, showing that mental imagery co-opts or reactivates neuronal resources used for perception.3 One consequence of this architecture is that damage to these brain areas can lead to problems not only with visual perception, but also with imagination.4 Some have taken these findings to suggest that the imagery debate is now resolved.5 However, overlap at the level of non-invasive recordings such as fMRI or EEG does not reveal overlap at a neuronal or circuit level. It is possible that imagery is subserved by activation of separate or parallel circuitry located in similar cortical regions. Addressing this question is critical for our understanding of mental imagery as a psychological phenomenon — is it really recruiting the same resources as when we perceive the outside world? Or does it only seem that way? This is where the study by Wadia and colleagues breaks new ground in mental imagery research.6 The authors recorded single neurons in human epilepsy patients (implanted with electrodes for clinical purposes) while they viewed and subsequently imagined objects. Focusing on neurons recorded from the ventral temporal cortex (VTC), the authors built a computational model in which neuronal activations were hypothesized to track one of 50 principal components of a low-dimensional feature space (obtained by passing the images through a deep convolutional neural network, AlexNet). In keeping with previous research,7 this feature space provided a good account of the VTC neuronal activity — so good, in fact, that a linear combination of neuronal activations was sufficient to reconstruct the object that participants were looking at. They then asked what happened in the brain when people were asked to imagine, rather than perceive, those same objects. Strikingly, a similar perceptual neuronal code was reactivated when imagining the objects — with ~40% of the object-tuned VTC neurons recorded during perception being reactivated by imagery. Notably, a computational account of the sensory code was important for understanding the type of reactivation induced by mental imagery. By tracking activation in a latent feature space, Wadia and colleagues were able to provide more secure inference that the same computation is recapitulated during imagination, rather than detecting reactivation at the level of neuronal activity alone. Furthermore, they were able to show that a visual code was a better explanation of the activation patterns than a semantic code. These results help imagery research move beyond simply asking whether a brain area is active, to an understanding of how its activity supports perception and imagery. It also showcases the importance of convergent evidence from both non-invasive brain imaging and studies of single-unit activity. Several questions emerge for future investigation. First of all, what are the mechanisms that allow the visual code to be harnessed in such a precise way? During perception, activation in the visual cortex is ultimately caused by bottom-up signals from the retina. In contrast, during imagery, these bottom-up signals are absent by definition. It is remarkable that such a nuanced and complex code can be recruited at will, by an instruction to imagine. How exactly can information from memory be combined to generate these rich visual representations in a purely top-down manner? Second, what supports the vividness or subjective force of mental imagery? This question could be asked by supplementing instructions to imagine simple objects with subjective ratings of imagery vividness. Other studies have noted that people with aphantasia — a lack of reported mental imagery experience — can nevertheless activate similar brain areas, and share a capacity to perform tasks that rely on perceptual reactivation, such as visual working memory.8,9 One possibility is that the same neuronal code is reactivated in both aphantasia and imagery, but that downstream regions (e.g., in the prefrontal cortex) fail to metacognitively access that information in aphantasia.10 Alternatively, a recent study showed that there are subtle differences in the way that visual representations are instantiated in people with aphantasia, suggesting that the exact format of the neuronal code during imagery might play a role in whether it reaches awareness.11 1 Department of Experimental Psychology and Institute of Cognitive Neuroscience, University College London, London, UK. 2Max Planck UCL Centre for Computational Psychiatry and Ageing Research, University College London, London, UK. 3CIFAR Program in Brain, Mind and Consciousness, Toronto, ON, Canada. 4Department of Imaging Neuroscience, University College London, London, UK. ✉email: stephen.fl[email protected] S.M. Fleming and N. Dijkstra 2 Finally, it remains unknown how the brain solves the “reality monitoring” problem — the problem of distinguishing between perceived and imagined sensory activities. Wadia and colleagues showed that a decoder could be trained to distinguish perceived and imagined objects from only a small number of VTC neurons — implying that there is sufficient information for reality monitoring within the VTC. However, it is unclear what exactly drives this classification. The 40% shared neurons between imagery and perception most likely provides a lower bound on the amount of overlap due to noise in the neuronal measurement, the selection of the computational mode (...truncated)