Barely depictive: Predicting imagery vividness relative to perception with EEGNet

This study introduces a probabilistic deep learning framework using EEGNet to quantify visual mental imagery vividness on a shared neural-behavioral scale with visual perception, revealing that imagery is a "barely depictive" process that is substantially weaker than perception yet scales with individual subjective ratings.

The Gist

The Big Question: How Real is Your Mental Movie?

Imagine you are sitting in a dark room. Someone asks you to look at a bright, high-definition picture of a strawberry on a screen. That is Visual Perception (VP).

Then, they ask you to close your eyes and imagine that same strawberry. That is Visual Mental Imagery (VMI).

For decades, scientists have debated: Is your mental image just a "weak" version of seeing the real thing, or is it something totally different? Most people say, "My mental image feels pretty clear!" But scientists suspect that, deep down in the brain, the "mental image" is actually much fainter than we think.

The problem is: How do you measure the "brightness" of a thought? You can't stick a ruler in someone's mind.

The Solution: The "Brain Translator"

The researchers in this study came up with a clever trick. Instead of asking people how bright their thoughts feel, they built a Brain Translator (a computer program called EEGNet) to "read" the brain's electrical signals.

Here is how they did it, step-by-step:

1. Training the Translator (The "Perception" Phase)

First, they showed 34 volunteers pictures of a cat and a strawberry. But they didn't just show them clear pictures. They showed them a whole spectrum of clarity:

• Level 1: A blank grey screen (nothing).
• Level 2: A very blurry, faint ghost of a strawberry.
• Level 3: A medium-blur strawberry.
• Level 4: A mostly clear strawberry.
• Level 5: A crystal-clear, high-definition strawberry.

While the volunteers looked at these, the researchers recorded their brainwaves using an EEG cap (like a swim cap with sensors). They fed this data into the computer.

The Analogy: Think of the computer as a student learning to read a language. The "teacher" (the real pictures) shows the student examples of "faint," "medium," and "bright." The student learns to recognize the specific electrical "accent" the brain uses for each level of brightness.

2. The Test (The "Imagination" Phase)

Next, the volunteers closed their eyes. The computer played a sound, and the volunteers had to imagine the strawberry. They had to rate how clear their mental image felt on a scale of 1 to 5.

The Twist: The computer didn't ask the volunteers what they felt. Instead, it looked at their brainwaves while they were imagining and asked: "Based on what I learned from the real pictures, how bright does this thought look to my sensors?"

The Surprising Results

The computer's "translation" revealed a fascinating truth:

1. The "Barely Depictive" Reality
When the computer looked at the brainwaves of people imagining the strawberry, it didn't see a "Level 5" (crystal clear) image. It didn't even see a "Level 2" (faint ghost).
It saw something barely above the "blank screen" (Level 1).

The Metaphor: Imagine you are looking at a real, bright firework (Perception). Now, imagine you are trying to remember that firework in your mind (Imagery).

• What people say: "I see the firework! It's almost as bright as the real one!"
• What the brain says: "Actually, the electrical signal for that memory is barely a flicker. It's more like a tiny, dying ember compared to the real explosion."

The study concludes that our mental images are "barely depictive." They exist, but they are incredibly weak compared to real seeing.

2. The Gap Between Feeling and Reality
Here is the most interesting part:

• Most people reported their mental images were quite vivid (around a 4 out of 5).
• The computer said their brain activity was only a 2 out of 5 (barely above nothing).

Why the difference?
The researchers suggest that when we imagine things, our brains might be "filling in the blanks" with our knowledge. We know what a strawberry looks like, so our brain tells us, "That's a strawberry!" and we feel like we see it clearly. But the raw visual signal in the back of the brain is actually very dim.

It's like listening to a song on the radio with static. Your brain is so good at recognizing the melody that you think you hear the song perfectly, even though the actual signal is full of noise.

Why Does This Matter?

This study is a game-changer for understanding Aphantasia (the inability to visualize images).

• Previously, people with aphantasia were told they had a "deficit" because they couldn't see images.
• This study suggests that everyone has very weak images in their brain, but most people just think they are seeing clearly because their brain fills in the gaps.
• People with aphantasia might just be the ones who are honest about how faint the signal actually is!

The Takeaway

We often think of our imagination as a "mental movie theater" with a big screen. This study suggests that the theater is actually a tiny, flickering projector in a dark room. We think the picture is clear because our brain is a great storyteller, but the actual "pixels" in our brain are barely there.

The researchers built a new tool (EEGNet) that can measure this faint signal, giving us a way to finally compare "seeing" and "imagining" on the same scale. It turns out, the gap between the two is much wider than we ever imagined.

Technical Summary

1. Problem Statement

Visual Mental Imagery (VMI) is widely considered a "weaker form" of top-down Visual Perception (VP). While neuroimaging studies suggest shared neural representations between VMI and VP, particularly in early visual areas (EVAs), the precise quantitative relationship remains ill-defined. Specifically:

• The Gap: It is unclear how much "weaker" VMI is compared to VP at the neural level.
• The Challenge: Traditional decoding approaches rely on categorical labels (e.g., "cat" vs. "house"). However, vividness is a continuous, subjective property. Obtaining objective, ground-truth labels for VMI vividness is difficult because it relies on introspection, which correlates weakly with neural activity.
• The Question: Can a model trained on the neural signatures of perceived vividness (where ground truth is controllable) be used to predict the neural vividness of imagined stimuli, thereby placing VMI and VP on a shared, objective scale?

2. Methodology

Experimental Design

• Participants: 34 healthy adults (after exclusions).
• Stimuli: Two categories (cat, strawberry) manipulated across 5 levels of "vividness" by parametrically degrading spatial sharpness (Gaussian blur) and luminance contrast.
  • Level 1: Baseline (grey screen, no image).
  • Level 2–4: Intermediate degradation.
  • Level 5: Original, high-fidelity image.
• Tasks:
  1. Perception Task (VP): Participants viewed stimuli at varying vividness levels and rated them.
  2. Imagery Task (VMI): Participants imagined the cued stimuli (based on auditory cues) and rated their mental image vividness.
• Data Acquisition: 64-channel EEG recorded at 512 Hz.

Data Preprocessing & Feature Extraction

• Pipeline: Bad channel removal (RANSAC), artifact rejection (ICA with ICLabel), band-pass filtering (0.1–42 Hz), and baseline correction.
• Input for Model: Single-trial EEG epochs (0–1000 ms post-stimulus) from 8 posterior electrodes (O1, O2, Oz, PO3, PO4, PO7, PO8, POz) targeting Early Visual Areas (EVAs). Data was resampled to 128 Hz.

Machine Learning Model (EEGNet)

• Architecture: A compact Convolutional Neural Network (CNN) designed for Brain-Computer Interfaces (EEGNet).
  • Layers: Temporal convolution (learning frequency filters) $\rightarrow$ Depthwise convolution (learning spatial filters) $\rightarrow$ Separable convolution (temporal summary and feature mixing) $\rightarrow$ Dense classification layer.
• Training Strategy: Leave-One-Subject-Out (LOSO) cross-validation. The model was trained on 33 participants and tested on the held-out participant, repeated 34 times. This ensures the model learns a subject-independent neural signature.
• Output Metrics:
  1. Discrete Prediction: The class with the highest probability.
  2. Expected Vividness (Continuous): A probabilistic weighted average ($\sum p_c \times v_c$) allowing for interpolation between trained classes and quantification of uncertainty.

Validation Approach

• Phase 1 (VP Validation): Train on VP data to decode 5 classes. Test on held-out VP data to verify the model learns a graded neural signature.
• Phase 2 (Interpolation): Train on 3 classes (Levels 1, 2, 5) to test generalization to unseen intermediate levels (3 and 4) in VP.
• Phase 3 (VMI Projection): Apply the trained VP model to unlabeled VMI trials to generate "expected vividness" scores. Compare these neural predictions against:
  • Neural predictions during VP.
  • Participants' self-reported (subjective) vividness.

3. Key Results

Perception Task Performance

• 5-Class Model: Achieved 44.32% accuracy (chance = 20%). Misclassifications were primarily between adjacent levels, indicating the model learned an ordinal relationship.
• 3-Class Model: Achieved 69.31% accuracy (chance = 33%).
• Interpolation: The model successfully generalized to unseen vividness levels (3 and 4) during perception, with expected vividness scores tracking the true ordinal rank of the stimuli. This confirmed the model learned a graded neural signature rather than memorizing discrete categories.

Imagery Task Findings

• Neural Vividness of VMI: When projecting VMI trials onto the perceptual scale:
  • Mean expected VMI vividness was significantly higher than the baseline (Level 1).
  • Mean expected VMI vividness was significantly lower than the lowest perceived vividness (Level 2).
  • Quantification: VMI vividness was approximately 38–46% lower than the vividness of the original high-fidelity stimuli (Level 5) and only slightly above the "no image" baseline.
• Conclusion: The neural data supports the description of VMI as "barely depictive" rather than "quasi-depictive." It engages EVAs but with substantially attenuated signal strength compared to actual perception.

Relationship with Subjective Reports

• Discrepancy: Participants reported high subjective vividness (mean 3.7/5), while the neural model predicted low vividness (1.9–2.1/5).
• Correlation:
  • Across the whole sample, the correlation between neural and subjective vividness was weak/non-significant.
  • Crucial Subgroup Analysis: When excluding 3 participants with extremely low self-reported vividness (suspected aphantasia), a significant positive correlation emerged ($\rho \approx 0.31–0.42$) for the remaining 91% of participants.
  • This suggests that for typical imagers, neural and subjective vividness scale together, but the absolute magnitude of the neural signal remains low.

4. Key Contributions

1. Novel Framework: Introduced a probabilistic deep learning approach to quantify VMI vividness on a shared neural-behavioural scale with VP, bypassing the need for ground-truth labels in imagery tasks.
2. Quantification of "Weakness": Provided empirical evidence that VMI is "barely depictive," quantifying it as a signal that is detectable above baseline but significantly weaker than even the most degraded perceptual input.
3. Generalizable Neural Signature: Demonstrated that a subject-independent CNN can decode a graded neural signature of vividness in EVAs that generalizes across participants and interpolates to unseen stimulus levels.
4. Re-evaluation of Aphantasia: Highlighted that individuals reporting "no image" may actually have neural signals closer to the "true" perceptual baseline, suggesting that high subjective reports in typical imagers might reflect an overestimation or different scaling of the vividness metric.

5. Significance and Implications

• Theoretical: Supports the "top-down VP" hypothesis but refines it, showing that while VMI shares neural mechanisms with VP, the signal strength is fundamentally different. It challenges the notion that VMI is a robust "picture in the mind" at the level of early visual cortex.
• Methodological: Offers a principled tool to study individual differences in imagery (e.g., aphantasia vs. hyperphantasia) without relying solely on subjective questionnaires, which are prone to bias.
• Clinical/Research: The framework can be adapted to other sensory domains (auditory, somatosensory) and could help distinguish between voluntary imagery, spontaneous mental images, and hallucinations by quantifying the "perceptual weight" of neural activity.
• Future Directions: The authors suggest that the discrepancy between neural and subjective reports may stem from how participants interpret rating scales or the temporal dynamics of imagery (e.g., lack of active maintenance). Future work should explore finer-grained perceptual thresholds to pinpoint the exact neural boundary between perception and imagery.