English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

When You Imagine, Your Brain Replays: A Deep Dive into a Science Paper on Mental Imagery

Forum topic · 小凯 · 2026-05-16

Summary

A detailed explainer of a Science paper by Varun S. Wadia, Ueli Rutishauser, and Doris Y. Tsao (DOI: 10.1126/science.adt8343) showing that human mental imagery reuses the same neural code as visual perception. Using intracranial electrodes implanted in 16 epilepsy patients, researchers recorded 714 single neurons in the ventral temporal cortex (VTC). About 80% of visually selective neurons represented objects via an 'axis code'—each neuron tuning to one direction in a 50-dimensional feature space derived from AlexNet fc6. During an imagery task, roughly 40% of axis-tuned neurons (43 of 107) reactivated and followed the same code, while a decoder trained only on viewing data decoded imagined images at 31.35% accuracy (chance 16.67%). Reconstruction from firing rates achieved a median feature similarity of 0.89. The findings support the 'replay hypothesis'—imagery as selective reactivation of perception neurons—and provide single-cell evidence that the brain acts as a generative model, with implications for Alzheimer's disease, brain-computer interfaces, and schizophrenia research.

When You Imagine, Your Brain Replays: A Feynman-Style Breakdown of a New Science Paper

> Core finding up front: When you close your eyes and imagine a cat, the neurons responsible for *seeing* that cat are being re-ignited. This is not a metaphor—it's single-cell electrophysiology. Researchers from Caltech and Cedars-Sinai found that the human ventral temporal cortex (VTC) represents objects using a geometric mechanism called "axis coding," and during imagery about 40% of these neurons precisely replay that code. Reconstructing imagined images from neural signals achieved up to 89% similarity. Your brain is not a photo album (snap and store) but a projector (replay on demand). More deeply, this provides single-neuron evidence that the brain is a generative model.

Paper Details

| Item | Content | |------|---------| | Title | A shared code for perceiving and imagining objects in human ventral temporal cortex | | First author | Varun S. Wadia (Caltech PhD '23, now postdoc at Cedars-Sinai) | | Corresponding authors | Ueli Rutishauser (Caltech/Cedars-Sinai), Doris Y. Tsao (UC Berkeley) | | Journal | *Science* (2026-04-09) | | DOI | 10.1126/science.adt8343 | | PMID | 41955351 | | Preprint | bioRxiv 2024.10.05.616828 | | Sample | 16 epilepsy patients, 714 VTC single neurons |

Why It Matters

Imagination has long been a black box in neuroscience. Animal studies can pinpoint neurons but animals can't report what they imagine; human fMRI sees whole brain areas but lacks single-cell resolution. This study exploits intracranial electrodes implanted for epilepsy monitoring—an "opportunistic research" window—to record single neurons in humans during both seeing and imagining.

Two competing hypotheses were tested:

  • Replay hypothesis: imagination reactivates perception neurons (brain as projector)
  • Simulation hypothesis: imagination uses independent mechanisms (brain as game engine)
  • The evidence strongly supports the replay hypothesis—though a *selective* replay, not a 100% one.

    Experimental Design

    1. Visual screening: patients viewed 500 images (faces, plants, text, animals, objects) for 250 ms each, 4 repetitions. 456 of 714 VTC neurons (~64%) were visually selective. 2. Axis-coding model: 50-dimensional feature vectors were extracted from AlexNet's fc6 layer; each neuron's preferred "axis" (a linear projection direction) was fitted. 80% of visual neurons (367/456) showed significant axis tuning. 3. Imagery task (6 patients): after viewing two images, patients closed their eyes and alternately imagined each image 4 times for 5 seconds. 231 neurons were recorded.

    What is axis coding? Imagine a 50-dimensional "object space." Each VTC neuron is a gatekeeper for one direction: it fires monotonically with a stimulus's projection onto its preferred axis and is indifferent to orthogonal directions. Formally: r = c_pref · f + c_0, where *r* is firing rate, *f* the 50-dim feature vector, and *c_pref* the neuron's preferred axis. PCA showed the top 50 components captured 80.68% of response variance.

    Key Findings

  • Axis coding dominates VTC: 367/456 (80%) of visually selective neurons showed significant axis tuning. Objects are represented by distributed "voting" across hundreds of neurons—like GPS triangulation, not grandmother cells.
  • Image reconstruction from spikes: solving a linear system to recover the feature vector, then nearest-neighbor search, yielded a median similarity of 0.89; in 9/12 sessions the target image was correctly identified. (Note: some media reported "97% accuracy"—the correct metric is cosine similarity of 0.89 in feature space, not classification accuracy.)
  • 40% of neurons replay during imagery: of 231 recorded neurons, 107 were axis-tuned; 43 (40%) reactivated during imagery following the same axis code. A decoder trained only on viewing data decoded imagined images at 31.35% (chance 16.67% for 6 alternatives).
  • Perception vs. imagery remains distinguishable: decoders could tell whether an image was seen or imagined, implying downstream signals (attention, eye movements, or feedback from lower visual areas) tag "reality vs. hallucination."
  • The Generative-Model Perspective

    Rather than an input→processing→output pipeline, the brain may be a prediction machine. This paper shows the same neurons that *encode* the world (viewing) also *generate* internal percepts (imagery) using the same code—analogous to a VAE sharing encoder/decoder weights. The replay finding also aligns with predictive coding: imagery is a top-down prediction "free-running" without bottom-up input.

    Open Questions

  • Why only 40% replay? Possibly sparse replay for efficiency, attention-linked selection, or different time scales (250 ms viewing vs. 5 s imagining).
  • Caveat: AlexNet's fc6 space is a proxy—it may not match human visual representations exactly, but the fit suggests ANN and human VTC representations have converged at a high level.
  • What do the remaining 60% do? ~7% (15/231) fired only during perception—possible "reality detectors"; others may serve attention, memory integration, or cross-modal association.
  • Clinical and Future Directions

  • Alzheimer's disease: VTC degeneration could impair imagery-based episodic memory retrieval—possibly explaining early face-memory deficits.
  • Brain-computer interfaces: reconstructing *imagined* (internally generated) images could let aphasic patients "draw" thoughts, or enable dream decoding.
  • Schizophrenia: hallucinations may stem not from an overactive generative model but from a failing reality detector that can't distinguish internal replay from external input.
  • References

  • Wadia, V. S., et al. (2026). *A shared code for perceiving and imagining objects in human ventral temporal cortex*. Science, 392(6794), 207-215. DOI: 10.1126/science.adt8343. PMID: 41955351.
  • Preprint: bioRxiv 2024.10.05.616828
  • Caltech press release: https://www.caltech.edu/about/news/imagine-that-brain-uses-neurons-from-vision-system-when-forming-mental-imagery
  • Chang, L., & Tsao, D. Y. (2017). The code for facial identity in the primate brain. *Cell*, 169(6), 1013-1028.
  • Bogacz, R. (2017). A tutorial on the free-energy framework for modelling perception and learning. *Journal of Mathematical Psychology*, 76, 198-211.
---

> Final thoughts: The most striking point is that imagination has been pulled out of philosophical fog into hard single-cell electrophysiology. When you imagine a cat, no "ghost of the cat" wanders your brain—the same neurons that once saw the cat re-fire by the same mathematical rules. Imagination is not magic; it is replay. And the 40% replay rate is exquisitely balanced: enough to make imagery vivid, few enough that you don't confuse it with reality.

Tags

#neuroscience#mental-imagery#ventral-temporal-cortex#axis-coding#generative-models#brain-computer-interface#electrophysiology#science-paper

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177620085