Pith. sign in

REVIEW 4 major objections 5 minor 55 references

Replacing human participants with generative models reveals why decoded neurofeedback sometimes fails to induce learning.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 21:34 UTC pith:UKXHPILD

load-bearing objection Genuinely new VAE-as-participant framework for DecNef simulation, internally clean and worth review — but the abstract overclaims 'reproducing empirical phenomena' with no human data to back it. the 4 major comments →

arxiv 2511.14555 v4 pith:UKXHPILD submitted 2025-11-18 q-bio.NC cs.AI

DecNefSimulator: A Modular, Interpretable Framework for Decoded Neurofeedback Simulation Using Generative Models

Alexander Olza , Roberto Santana , David Soto This is my paper
classification q-bio.NC cs.AI
keywords decoded neurofeedbacksimulation frameworkgenerative modelvariational autoencoderlatent spacenon-respondersalternative classprotocol design
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Decoded neurofeedback (DecNef) trains people to alter their own brain states using real-time classifier output as reward, but outcomes vary widely, some participants never learn, and experiments are costly. This paper argues that these problems can be studied, and to a large degree explained, in simulation by replacing the human with a latent-variable generative model—a variational autoencoder—whose hidden space stands in for internal cognitive states. With that stand-in, everything that is hidden in real experiments becomes observable: the paper shows that the same 'participant' learns or fails depending on which alternative class the decoder was trained against, where the trajectory starts, and the cumulative randomness of its attempts at regulation. The authors' claim is that such a modular, task-agnostic simulator can reproduce empirical DecNef phenomena, diagnose maladaptive learning, and let researchers test protocols in silico before running them on humans. A careful reader should care because, if the latent-space mapping holds, the framework turns protocol design into an inspectable machine-learning problem instead of a costly black-box trial.

Core claim

The central claim is that a simulation framework called DecNefLab—which replaces the human participant with a variational autoencoder trained on images as stand-ins for fMRI observations—can reproduce the key behavioral phenomena of decoded neurofeedback and reveal the conditions under which feedback fails to drive learning. In this setup, the VAE's latent space is treated as the participant's set of internal cognitive states; its decoder projects those states into an observable space, and a supervised binary classifier (the 'decoder') converts each observable state into a feedback probability for the target class. A hand-built update rule simulates cognition: the participant trusts the feed

What carries the argument

The load-bearing object is the latent-variable generative model used as artificial participant, concretely a variational autoencoder with a two-dimensional latent space Z; the encoder maps observable data to latent cognitive states and the decoder generates observable proxies x = D_G(z). The other central pieces are the binary classifier D (the 'decoder' in DecNef) that supplies feedback p(y = y* | x), and the update rule L(z, p, λ, γ, δ) that models the trainee's exploration-exploitation behavior and one-step memory. The key work done by this machinery is the separation of the cognitive trajectory (z_t) from the observable trajectory (x_t): because z_t is directly accessible in simulation,

Load-bearing premise

The paper's own load-bearing premise, stated in Section 5, is that the variational autoencoder's latent space corresponds to human cognitive states; if that mapping does not hold for real participants, the protocol-design recommendations drawn from simulation do not transfer.

What would settle it

Run the same simulation with a third alternative class (e.g., Coat) using identical initial states and random seeds; if its learning curves and final latent-state distributions resemble the Dress condition more than the Trouser condition, the claimed alternative-class dependence needs revision. More decisively, in a real DecNef study with a fixed participant pool, if swapping only the alternative class does not change the fraction of learners and the measured neural states, the central claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Protocol designers can vet the choice of alternative class in advance: a poor contrast can turn a learner into an apparent non-responder before any human is scanned.
  • Researchers can distinguish true non-response from experimental artifact by running many trajectories from many initial states; high variance across random seeds signals that the 'non-responder' label is not a stable trait.
  • A positive feedback trajectory is not evidence of target-state induction: the framework shows rewards can rise while the latent state drifts away from the target, so success measures should include distance to the intended state, not just reward.
  • Because the generator, classifier, and update rule are modular, the simulator can serve as a testbed for new protocols, self-supervised discriminators, or other generative architectures before in vivo implementation.
  • The results bear directly on the 'decoder's dictum' debate: decodability of a state by the experimenter does not imply the brain uses that information, and the simulation makes this discrepancy observable.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the latent-to-cognition mapping holds, a natural next experiment is to hold a real human cohort fixed and vary only the alternative class; the framework predicts responder rates should shift measurably, a comparatively cheap test.
  • Editorial inference: the exploration-suppression effect suggests a concrete protocol idea not explored in the paper—start trainees in neutral or low-reward states, or use reward shaping, to prevent premature convergence to a high-feedback-but-wrong region.
  • Editorial inference: a trajectory-dispersion diagnostic—computing the spread of final latent states across random seeds for a fixed starting state—could be exported to real data to estimate how much non-response is noise-driven.
  • Editorial inference: because the autoencoder is an intentionally imperfect, noisy model, the framework's predictions are about qualitative regime shifts rather than exact effect sizes; extending it to synthetic fMRI with realistic noise would make quantitative predictions testable.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes DecNefLab (the arXiv title says DecNefSimulator), a modular simulation framework for decoded neurofeedback (DecNef). A human participant is replaced by a latent-variable generative model—here a VAE trained on Fashion-MNIST—whose latent space is treated as the participant's cognitive-state space. A supervised classifier is trained on target and alternative image classes to provide the feedback signal. The participant's learning strategy is a hand-designed update rule (Eqs. 2–3): exploration variance shrinks when feedback increases, grows when feedback decreases, and the state reverts to the previous state after a sufficiently large reward drop. The paper reports simulations varying the alternative class (Trousers vs. Dresses), initial latent states, and random transients, with a random-feedback control. The main claims are that the alternative class is a critical determinant of learning, that initial states with high feedback suppress exploration, and that initial conditions and randomness can make the same participant appear to be a responder or non-responder.

Significance. If its central claim—that the framework can reproduce empirical phenomena of DecNef learning and guide protocol design—were established, the paper would offer a useful, transparent, and modular in-silico tool for the DecNef community. The framework has notable strengths: the cognitive trajectory is fully observable, control experiments with random feedback are included and behave as expected (chance-level trajectories), and random seeds are matched across conditions so that alternative-class effects are not confounded by initialization. The demonstration of maladaptive high-feedback-but-off-target states, enabled by directly comparing latent trajectories with classifier output, is a valuable proof of concept. However, the paper's practical payoff—protocol recommendations for human experiments—rests on an interpretive mapping between VAE latent space and human cognitive states, which the authors explicitly acknowledge is outside the scope of the study. This limits the external validity of the results and makes the current text stronger as a demonstration of a modeling framework than as a reproduction of empirical DecNef phenomena.

major comments (4)
  1. [Abstract and Section 5] The abstract claims the framework can '(i) reproduce empirical phenomena of DecNef learning,' but no comparison with human DecNef data, nor with earlier DecNef simulations (e.g., Oblak et al. 2017, Shibata et al. 2019, Annicchiarico et al. 2025), is presented. Section 5 explicitly concedes that 'The correspondence we assume between the model's latent space and human cognitive states is itself an interpretive assumption' and that full evaluation is out of scope. This concession, placed where it is, directly undercuts the reproduction claim. The central claim should be either validated against at least one published empirical DecNef dataset (e.g., the 2011 Shibata orientation study) at the level of qualitative learning curves, or weakened to 'illustrates candidate mechanisms under stated assumptions.' As written, the claim is load-bearing and unsupported.
  2. [Section 4.1.2, Eqs. (2)–(3), and Section 4.3.4] Several of the paper's 'findings' are built into the update rule rather than discovered. Eq. (2) directly makes the exploration variance σ shrink as p approaches 1 and grow as p falls; Eq. (3) makes low feedback produce random exploration and a large feedback drop trigger reversion to the previous state. The report in Section 4.3.4 that 'initial states associated with low reward signals promote exploration' and 'initial states with high feedback suppress exploration' is therefore a restatement of the model's definition, not an emergent result. The same applies to the non-responder trajectories in Section 4.3.3, which are consequences of the stochastic term in Eq. (3). This circularity does not invalidate the framework, but the manuscript should explicitly frame these outcomes as consequences of the assumed learning strategy and should include sensitivity analyses over λ, γ, δ, and altern
  3. [Section 4.3.2 and Section 4.2.4] The conclusion that alternative-class choice is a 'critical determinant' is based on a comparison of two classifiers that differ not only in the alternative class but also in classification accuracy (0.96 for Trousers, 0.88 for Dresses, reported in Figure 2). The paper does not report calibration, confidence distributions, or any adjustment for classifier quality. It is therefore not possible to separate the effect of the alternative class's semantic relation to the target from the effect of overall classifier performance. The claim that the entire difference in learning dynamics is 'entirely because of the choice of the alternative class' (Section 4.3.3) overstates what the design can establish. The authors should either match classifiers for accuracy/calibration or explicitly analyze the contribution of classifier performance.
  4. [Section 4.3.3 and Section 4.3.4] The evaluation of learning success or failure is qualitative. The text states that trajectories with increasing p_t or final p_t above 0.5 would be labeled 'successful,' but no quantitative thresholds or proportions are reported. The claim that 'the same participant would have been far more likely to be deemed incapable' with the Dress classifier is not supported by statistical evidence—no counts, percentages, or confidence intervals are given. Also, the A–E regions in Figure 4 appear to be chosen post hoc; no independent operational definitions are provided. Adding explicit success criteria and reporting trajectory-level statistics would make the paper's central claims testable and would strengthen the framework's utility as a diagnostic tool.
minor comments (5)
  1. [Title and Abstract] The arXiv title says 'DecNefSimulator' while the full text uses 'DecNefLab.' This mismatch should be corrected for indexing and consistency.
  2. [Figure 1] The figure caption says 'The decoder E_G computes the corresponding x_0,' but E_G was previously defined as the encoder; the pipeline should use D_G for decoding. This is likely a typographical error but is confusing.
  3. [Section 4.2.2] The warm-up procedure is described as providing static p_t = 0.5 for t = 0,...,3 while simultaneously computing the 'real' probabilities. The sentence 'the first true feedback value visible to the participant was computed using the average' should be clarified: does the participant receive 0.5 or the average during the first four steps? The current wording is ambiguous.
  4. [Section 5] The phrase 'The correspondence we assume ... is itself an interpretive assumption' is a major limitation and should appear in a dedicated 'Limitations' subsection, not only in the discussion. Given that this assumption is the basis for protocol-transfer claims, it deserves prominence.
  5. [General] There is no code availability statement. Given the paper's emphasis on modularity and reproducibility, sharing the simulation code (with seeds and hyperparameters) would be a substantial addition and would let other researchers test alternative generators, classifiers, and update rules.

Circularity Check

0 steps flagged

No material circularity: the framework's outputs are transparent consequences of its stated update rules, and the main caveat is external validity, not a hidden equivalence.

full rationale

The claimed derivation chain is an explicit simulation, not a fitted model disguised as prediction. The update rules (Eqs. 2-3) are stated as modeling assumptions (Section 4.1), and the 'findings' in Sections 4.3 and 5 — e.g., high-feedback initial states suppress exploration, randomness can produce apparent non-responders — are the behavior of those rules. That is model entailment, not circularity: the paper does not claim to infer the rules from human data and then 'predict' them back. No parameter is fitted to a subset and then reported as a prediction. The self-citations ([35], and [1] in the reference list) are background/supporting citations and are not load-bearing for the central framework. The genuine weakness is external validity: Section 5 explicitly concedes that 'The correspondence we assume between the model's latent space and human cognitive states is itself an interpretive assumption' and that 'a full evaluation of this mapping — including comparisons with alternative architectures such as LLMs or CLIP-based models — lies outside the scope of the present study.' This means the abstract's 'reproduce empirical phenomena of DecNef learning' is under-supported without a human-data benchmark, but that is missing evidence, not a circular derivation. Therefore no circular step meets the quote-and-reduction bar; score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claims about learning dynamics rest on six hand-set parameters and four modeling assumptions, none of which are fit to or validated against empirical DecNef data. The framework's contribution is as a tool, but its simulated 'findings' are largely logical consequences of the chosen equations and classifier contrasts, not independent discoveries.

free parameters (6)
  • λ (trust in feedback) = 0.2
    Hand-set in Section 4.2.2; weights current state vs. random exploration in Eq. (3). Determines how strongly feedback pushes the latent state.
  • γ (reactivity/impulsivity) = 0.2
    Hand-set in Section 4.2.2; controls exploration scaling in Eq. (2); high γ makes σ respond fast to feedback.
  • δ (revert threshold) = 0.75
    Hand-set in Section 4.2.2; if p_t < δ p_{t-1}, subject reverts to previous state (Eq. 3); directly implements 'cautious participant' behavior.
  • w (feedback averaging window / warm-up) = 4
    Section 4.2.2; feedback is averaged over last w probabilities; affects smoothness of the reward signal.
  • T (simulation horizon) = 500
    Section 4.2.2; trajectory length; affects whether a participant is labeled as learner or non-learner at session end.
  • m (VAE latent dimension) = 2
    Section 4.1.1; chosen for visualization; shapes the geometry of the cognitive-state space and the feedback landscape.
axioms (4)
  • domain assumption The VAE latent space Z corresponds to the subject's cognitive states, and decoder outputs X are observable proxies (analogous to fMRI).
    Section 4.1.1 and Section 5: 'The correspondence we assume between the model's latent space and human cognitive states is itself an interpretive assumption.' Load-bearing for any transfer of conclusions to real participants.
  • ad hoc to paper The learning strategy given by Eqs. (2)-(3) captures exploration/exploitation and reversal behavior of DecNef participants.
    Section 4.1.2: defined from three stated assumptions, not calibrated against human time series. All dynamic 'findings' are consequences of this rule.
  • domain assumption The binary classifier's out-of-distribution probabilities describe how DecNef feedback is computed in practice.
    Assumed in Sections 4.3.1-4.3.2; the classifier's overconfidence is a known ML phenomenon, but its exact landscape in the simulation is model-specific.
  • domain assumption Participants maximize instantaneous target-class probability p_t without knowledge of the target.
    Standard DecNef assumption noted in Sections 2 and 4.1; used to define learning dynamics.

pith-pipeline@v1.3.0-alltime-deepseek · 17972 in / 11766 out tokens · 122972 ms · 2026-08-03T21:34:48.568717+00:00 · methodology

0 comments
read the original abstract

Decoded Neurofeedback (DecNef) is a promising non-invasive approach to brain modulation with wide-ranging applications in neuromedicine and cognitive neuroscience. However, progress in DecNef research remains constrained by subject-dependent learning variability, reliance on indirect measures to quantify progress, and the high cost and time demands of experimentation. We present DecNefSimulator, a modular and interpretable simulation framework that formalizes DecNef as a machine learning problem. Beyond providing a virtual laboratory, DecNefSimulator enables researchers to model, analyze and understand neurofeedback dynamics. Using latent variable generative models as simulated participants, DecNefSimulator allows direct observation of internal cognitive states and systematic evaluation of how different protocol designs and subject characteristics influence learning. We demonstrate how this approach can (i) reproduce empirical phenomena of DecNef learning, (ii) identify conditions under which DecNef feedback fails to induce learning, and (iii) guide the design of more robust and reliable DecNef protocols in silico before human implementation. In summary, DecNefSimulator bridges computational modeling and cognitive neuroscience, offering a principled foundation for methodological innovation, robust protocol design, and ultimately, a deeper understanding of DecNef-based brain modulation.

Figures

Figures reproduced from arXiv: 2511.14555 by Alexander Olza, David Soto, Roberto Santana.

Figure 1
Figure 1. Figure 1: Pipeline of a generic DecNef simulation. [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Probabilities given by each binary discriminator to the images generated by the VAE via [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Evolution of pt during the 1000 simulations for each experiment, grouped and averaged by initial point z0 (thin lines). The color indicates the category of the Gaussian latent prototype from which each z0 was sampled. The bold lines represent the average pt across all trajectories for which the initial point was sampled around the same latent prototype. happened for most trajectories in this specific exper… view at source ↗
Figure 4
Figure 4. Figure 4: Trajectories in the latent space for the DecNef simulation with target-class [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 24 canonical work pages · 2 internal anchors

  1. [1]

    Enhancing fMRI decoded neu- rofeedback with co-adaptive training: simulation and proof-of-principle evidence

    Najemeddine Abdennour, Pedro Margolles, and David Soto. “Enhancing fMRI decoded neu- rofeedback with co-adaptive training: simulation and proof-of-principle evidence”. In:bioRxiv (2025).doi:10.1101/2025.02.21.639408

    work page doi:10.1101/2025.02.21.639408 2025

  2. [2]

    Can We Predict Who Will Respond to Neurofeedback? A Review of the Inefficacy Problem and Existing Predictors for Successful EEG Neurofeedback Learning

    Oren Alkoby et al. “Can We Predict Who Will Respond to Neurofeedback? A Review of the Inefficacy Problem and Existing Predictors for Successful EEG Neurofeedback Learning”. In: Neuroscience378 (2018), pp. 155–164.doi:10.1016/j.neuroscience.2016.12.050

    doi:10.1016/j.neuroscience.2016.12.050 2018

  3. [3]

    Learning to Associate Orientation with Color in Early Visual Areas by Associative Decoded fMRI Neurofeedback

    Kaoru Amano et al. “Learning to Associate Orientation with Color in Early Visual Areas by Associative Decoded fMRI Neurofeedback”. In:Current Biology26.14 (2016), pp. 1861–1866. doi:10.1016/j.cub.2016.05.014

    work page doi:10.1016/j.cub.2016.05.014 2016

  4. [4]

    An Active Inference perspective on Neurofeedback Training

    Cˆ ome Annicchiarico, Fabien Lotte, and J´ er´ emie Mattout. “An Active Inference perspective on Neurofeedback Training”. In:arXiv(2025).doi:https://doi.org/10.48550/arXiv. 2505.03308

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.03308 2025

  5. [5]

    Whole-brain causal connectivity during decoded neurofeedback: a meta study

    Fahimeh Arab et al. “Whole-brain causal connectivity during decoded neurofeedback: a meta study”. In:bioRxiv(2024).doi:10.1101/2024.11.16.623939. 23

    work page doi:10.1101/2024.11.16.623939 2024

  6. [6]

    Generative Feedback Explains Distinct Brain Activity Codes for Seen and Mental Images

    Jesse L. Breedlove et al. “Generative Feedback Explains Distinct Brain Activity Codes for Seen and Mental Images”. In:Current Biology30.12 (2020), 2211–2224.e6.doi:https:// doi.org/10.1016/j.cub.2020.04.014

    doi:10.1016/j.cub.2020.04.014 2020

  7. [7]

    Unconscious reinforcement learning of hidden brain states supported by confidence

    Aurelio Cortese, Hakwan Lau, and Mitsuo Kawato. “Unconscious reinforcement learning of hidden brain states supported by confidence”. In:Nature Communications11.1 (2020), p. 4429.doi:10.1038/s41467-020-17828-8

    work page doi:10.1038/s41467-020-17828-8 2020

  8. [8]

    Decoded fMRI neurofeedback can induce bidirectional confidence changes within single participants

    Aurelio Cortese et al. “Decoded fMRI neurofeedback can induce bidirectional confidence changes within single participants”. In:NeuroImage149 (2017), pp. 323–337.doi:https: //doi.org/10.1016/j.neuroimage.2017.01.069

    work page doi:10.1016/j.neuroimage.2017.01.069 2017

  9. [9]

    Multivoxel neurofeedback selectively modulates confidence without changing perceptual performance

    Aurelio Cortese et al. “Multivoxel neurofeedback selectively modulates confidence without changing perceptual performance”. In:Nature Communications7.1 (2016), p. 13669.doi: 10.1038/ncomms13669

    work page doi:10.1038/ncomms13669 2016

  10. [10]

    Toward brain-inspired foundation model for EEG signal processing: our opinion

    Suhan Cui, Dongwon Lee, and Dong Wen. “Toward brain-inspired foundation model for EEG signal processing: our opinion”. In:Frontiers in NeuroscienceVolume 18 - 2024 (2024).doi: 10.3389/fnins.2024.1507654

    arXiv 2024

  11. [11]

    Chapter 3 - The multi-stage theory of neurofeedback learning: a framework for understanding mechanisms

    Eddy J. Davelaar. “Chapter 3 - The multi-stage theory of neurofeedback learning: a framework for understanding mechanisms”. In:Current Research in Neuroadaptive Technology. Ed. by Stephen H. Fairclough and Thorsten O. Zander. Academic Press, 2022, pp. 43–57.doi:https: //doi.org/10.1016/B978-0-12-821413-8.00008-7

    work page doi:10.1016/b978-0-12-821413-8.00008-7 2022

  12. [12]

    Mechanisms of Neurofeedback: A Computation-theoretic Approach

    Eddy J. Davelaar. “Mechanisms of Neurofeedback: A Computation-theoretic Approach”. In: Neuroscience378 (2018), pp. 175–188.doi:https://doi.org/10.1016/j.neuroscience. 2017.05.052

    work page doi:10.1016/j.neuroscience 2018

  13. [13]

    Adversarially Learned Inference

    Vincent Dumoulin et al. “Adversarially Learned Inference”. In:Proceedings of the Interna- tional Conference on Learning Representations. 2017.doi:10.48550/arXiv.1606.00704

    Pith/arXiv arXiv doi:10.48550/arxiv.1606.00704 2017

  14. [14]

    Predictors of real time fMRI neurofeedback performance and improve- ment: A machine learning mega-analysis

    Amelie Haugg et al. “Predictors of real time fMRI neurofeedback performance and improve- ment: A machine learning mega-analysis”. In:NeuroImage237 (2021), p. 118207.doi:https: //doi.org/10.1016/j.neuroimage.2021.118207

    arXiv 2021

  15. [15]

    Hierarchical VAEs Know What They Don't Know

    Jakob D. Havtorn et al. “Hierarchical V AEs Know What They Don’t Know”. In:Proceedings of the 38th International Conference on Machine Learning. Ed. by Marina Meila and Tong Zhang. Vol. 139. PMLR, 2021, pp. 4117–4128.doi:https://doi.org/10.48550/arXiv. 2102.08248

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2102.08248 2021

  16. [16]

    Quality and denoising in real-time functional magnetic resonance imaging neurofeedback: A methods review

    Stephan Heunis et al. “Quality and denoising in real-time functional magnetic resonance imaging neurofeedback: A methods review”. In:Hum. Brain Mapp.41.12 (2020), pp. 3439– 3467.doi:10.1002/hbm.25010

    work page doi:10.1002/hbm.25010 2020

  17. [17]

    Visual Image Reconstruction from Brain Activity via Latent Representation

    Yukiyasu Kamitani, Misato Tanaka, and Ken Shirakawa. “Visual Image Reconstruction from Brain Activity via Latent Representation”. In:Annual Review of Vision Science11.Volume 11, 2025 (2025), pp. 611–634.doi:https://doi.org/10.1146/annurev-vision-110423- 023616

    work page doi:10.1146/annurev-vision-110423- 2025

  18. [18]

    Variational Diffusion Models

    Diederik Kingma et al. “Variational Diffusion Models”. In:Advances in Neural Information Processing Systems. Ed. by M. Ranzato et al. Vol. 34. Curran Associates, Inc., 2021, pp. 21696– 21707.doi:10.48550/arXiv.2107.00630. 24

    Pith/arXiv arXiv doi:10.48550/arxiv.2107.00630 2021

  19. [19]

    Adam: A Method for Stochastic Optimization

    Diederik P. Kingma and Jimmy Ba. “Adam: A Method for Stochastic Optimization”. In: Proceedings of the 3rd International Conference for Learning Representations. 2015.doi:10. 48550/arXiv.1412.6980

    Pith/arXiv arXiv doi:10.48550/arxiv.1412.6980 2015

  20. [20]

    An Introduction to Variational Autoencoders

    Diederik P. Kingma and Max Welling. “An Introduction to Variational Autoencoders”. In: Found. Trends Mach. Learn.12.4 (2019), pp. 307–392.doi:10.48550/arXiv.1906.02691

    Pith/arXiv arXiv doi:10.48550/arxiv.1906.02691 2019

  21. [21]

    Essentials of the self-organizing map

    Teuvo Kohonen. “Essentials of the self-organizing map”. In:Neural Networks37 (2013), pp. 52–65.doi:https://doi.org/10.1016/j.neunet.2012.09.018

    work page doi:10.1016/j.neunet.2012.09.018 2013

  22. [22]

    Fear reduction without fear through reinforcement of neural activity that bypasses conscious exposure

    Ai Koizumi et al. “Fear reduction without fear through reinforcement of neural activity that bypasses conscious exposure”. In:Nature Human Behaviour1.1 (2016), p. 0006.doi:10 . 1038/s41562-016-0006

    2016

  23. [23]

    Signal quality and Bayesian signal processing in neurofeedback based on real-time fMRI

    Yury Koush et al. “Signal quality and Bayesian signal processing in neurofeedback based on real-time fMRI”. In:NeuroImage59.1 (2012), pp. 478–489.doi:https://doi.org/10.1016/ j.neuroimage.2011.07.076

    2012

  24. [24]

    Representation, Pattern Information, and Brain Signatures: From Neurons to Neuroimaging

    Philip A. Kragel et al. “Representation, Pattern Information, and Brain Signatures: From Neurons to Neuroimaging”. In:Neuron99.2 (2018), pp. 257–273.doi:10.1016/j.neuron. 2018.06.009

    doi:10.1016/j.neuron 2018

  25. [25]

    Decoding fMRI brain states in real-time

    Stephen M. LaConte. “Decoding fMRI brain states in real-time”. In:NeuroImage56.2 (2011), pp. 440–454.doi:https://doi.org/10.1016/j.neuroimage.2010.06.052

    work page doi:10.1016/j.neuroimage.2010.06.052 2011

  26. [26]

    Self-regulation of the anterior insula: Reinforcement learning using real-time fMRI neurofeedback

    Emma J. Lawrence et al. “Self-regulation of the anterior insula: Reinforcement learning using real-time fMRI neurofeedback”. In:NeuroImage88 (2014), pp. 113–124.doi:https://doi. org/10.1016/j.neuroimage.2013.10.069

    work page doi:10.1016/j.neuroimage.2013.10.069 2014

  27. [27]

    Handwritten digit recognition with a back-propagation network

    Y. Le Cun et al. “Handwritten digit recognition with a back-propagation network”. In:Pro- ceedings of the 3rd International Conference on Neural Information Processing Systems. Cam- bridge, MA, USA: MIT Press, 1989, pp. 396–404

    1989

  28. [28]

    What we can do and what we cannot do with fMRI

    Nikos K. Logothetis. “What we can do and what we cannot do with fMRI”. In:Nature 453.7197 (2008), pp. 869–878.doi:10.1038/nature06976

    doi:10.1038/nature06976 2008

  29. [29]

    Neurofeedback through the lens of reinforcement learning

    Nitzan Lubianiker et al. “Neurofeedback through the lens of reinforcement learning”. In: Trends in Neurosciences45.8 (2022), pp. 579–593.doi:10.1016/j.tins.2022.03.008

    doi:10.1016/j.tins.2022.03.008 2022

  30. [30]

    Unconscious Manipulation of Conceptual Representations with De- coded Neurofeedback Impacts Search Behavior

    Pedro Margolles et al. “Unconscious Manipulation of Conceptual Representations with De- coded Neurofeedback Impacts Search Behavior”. In:Journal of Neuroscience44.2 (2024). doi:10.1523/JNEUROSCI.1235-23.2023

    work page doi:10.1523/jneurosci.1235-23.2023 2024

  31. [31]

    A framework for disentangling the hyperbolic truth of neurofeedback: Comment on Thibault and Raz (2017)

    T. Micoulaud-Franchi J.-A. & Fovet. “A framework for disentangling the hyperbolic truth of neurofeedback: Comment on Thibault and Raz (2017)”. In:American Psychologist73(7) (2018), pp. 933–935.doi:https://doi.org/10.1037/amp0000340

    work page doi:10.1037/amp0000340 2017

  32. [32]

    Universal adversarial perturbations

    Seyed-Mohsen Moosavi-Dezfooli et al. “Universal adversarial perturbations”. In:Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017, pp. 1765–1773. doi:10.1109/CVPR.2017.17

    doi:10.1109/cvpr.2017.17 2017

  33. [33]

    Reinforcement learning in multidimensional environments relies on atten- tion mechanisms

    Yael Niv et al. “Reinforcement learning in multidimensional environments relies on atten- tion mechanisms”. In:Journal of Neuroscience35.21 (2015), pp. 8145–8157.doi:10.1523/ JNEUROSCI.2978-14.2015. 25

    2015

  34. [34]

    Self-regulation strategy, feedback timing and hemodynamic properties modulate learning in a simulated fMRI neu- rofeedback environment

    Ethan F. Oblak, Jarrod A. Lewis-Peacock, and James S. Sulzer. “Self-regulation strategy, feedback timing and hemodynamic properties modulate learning in a simulated fMRI neu- rofeedback environment”. In:PLOS Computational Biology13.7 (2017), pp. 1–25.doi:10. 1371/journal.pcbi.1005681

    2017

  35. [35]

    Domain Adaptation-enhanced search- light: enabling classification of brain states from visual perception to mental imagery

    Alexander Olza, David Soto, and Roberto Santana. “Domain Adaptation-enhanced search- light: enabling classification of brain states from visual perception to mental imagery”. In: Brain Informatics12.1 (2025), p. 17.doi:10.1186/s40708-025-00263-0

    work page doi:10.1186/s40708-025-00263-0 2025

  36. [36]

    Natural scene reconstruction from fMRI signals using generative latent diffusion

    Furkan Ozcelik and Rufin VanRullen. “Natural scene reconstruction from fMRI signals using generative latent diffusion”. In:Scientific Reports13.1 (2023), p. 15666.doi:10 . 1038 / s41598-023-42891-8

    2023

  37. [37]

    Generative Embeddings of Brain Collective Dynamics Using Varia- tional Autoencoders

    Yonatan Sanz Perl et al. “Generative Embeddings of Brain Collective Dynamics Using Varia- tional Autoencoders”. In:Phys. Rev. Lett.125 (23 2020), p. 238101.doi:10.1103/PhysRevLett. 125.238101

    doi:10.1103/physrevlett 2020

  38. [38]

    Low-dimensional organization of global brain states of reduced consciousness

    Yonatan Sanz Perl et al. “Low-dimensional organization of global brain states of reduced consciousness”. In:Cell Reports42.5 (2023), p. 112491.doi:https://doi.org/10.1016/j. celrep.2023.112491

    arXiv 2023

  39. [39]

    Decoding the Brain: Neural Representation and the Limits of Multivariate Pattern Analysis in Cognitive Neuroscience

    J. Brendan Ritchie, David Michael Kaplan, and Colin Klein. “Decoding the Brain: Neural Representation and the Limits of Multivariate Pattern Analysis in Cognitive Neuroscience”. In:British Journal for the Philosophy of Science70.2 (2019), pp. 581–607.doi:10.1093/ bjps/axx023

    2019

  40. [40]

    Consensus on the reporting and experimental design of clinical and cognitive- behavioural neurofeedback studies (CRED-nf checklist)

    Tomas Ros et al. “Consensus on the reporting and experimental design of clinical and cognitive- behavioural neurofeedback studies (CRED-nf checklist)”. In:Brain143.6 (2020), pp. 1674– 1685.doi:10.1093/brain/awaa009

    work page doi:10.1093/brain/awaa009 2020

  41. [41]

    Chapter 13 - Mechanisms of fMRI neurofeedback

    Kazuhisa Shibata. “Chapter 13 - Mechanisms of fMRI neurofeedback”. In:fMRI Neurofeed- back. Ed. by Michelle Hampson. Academic Press, 2021, pp. 287–313.doi:https://doi.org/ 10.1016/B978-0-12-822421-2.00004-1

    work page doi:10.1016/b978-0-12-822421-2.00004-1 2021

  42. [42]

    Perceptual learning incepted by decoded fMRI neurofeedback with- out stimulus presentation

    Kazuhisa Shibata et al. “Perceptual learning incepted by decoded fMRI neurofeedback with- out stimulus presentation”. In:Science334.6061 (2011), pp. 1413–1415.doi:10 . 1126 / science.1212003

    2011

  43. [43]

    Toward a comprehensive understanding of the neural mechanisms of decoded neurofeedback

    Kazuhisa Shibata et al. “Toward a comprehensive understanding of the neural mechanisms of decoded neurofeedback”. In:NeuroImage188 (2019), pp. 539–556.doi:https://doi.org/ 10.1016/j.neuroimage.2018.12.022

    work page doi:10.1016/j.neuroimage.2018.12.022 2019

  44. [44]

    Closed-loop brain training: the science of neurofeedback

    Ranganatha Sitaram et al. “Closed-loop brain training: the science of neurofeedback”. In: Nature Reviews Neuroscience18.2 (2017), pp. 86–100.doi:10.1038/nrn.2016.164

    doi:10.1038/nrn.2016.164 2017

  45. [45]

    Partially Observable Markov Decision Processes

    Matthijs T. J. Spaan. “Partially Observable Markov Decision Processes”. In:Reinforcement Learning: State-of-the-Art. Ed. by Marco Wiering and Martijn van Otterlo. Berlin, Heidelberg: Springer Berlin Heidelberg, 2012, pp. 387–414.doi:10.1007/978-3-642-27645-3_12

    doi:10.1007/978-3-642-27645-3_12 2012

  46. [46]

    Neural mechanisms of neurofeedback: A review

    Carl Stevens and Silvia E. Kober. “Neural mechanisms of neurofeedback: A review”. In: Frontiers in Human Neuroscience15 (2021), p. 619546.doi:10.3389/fnhum.2021.619546

    arXiv 2021

  47. [47]

    High-resolution image reconstruction with latent diffusion models from human brain activity

    Yu Takagi and Shinji Nishimoto. “High-resolution image reconstruction with latent diffusion models from human brain activity”. In:2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2023, pp. 14453–14463.doi:10.1109/CVPR52729.2023.01389. 26

    arXiv 2023

  48. [48]

    Conducting decoded neurofeedback studies

    Vincent Taschereau-Dumouchel et al. “Conducting decoded neurofeedback studies”. In:Social Cognitive and Affective Neuroscience16.8 (2020), pp. 838–848.doi:10.1093/scan/nsaa063

    work page doi:10.1093/scan/nsaa063 2020

  49. [49]

    The psychology of neurofeedback: Clinical intervention even if applied placebo

    Robert T. Thibault and Amir Raz. “The psychology of neurofeedback: Clinical intervention even if applied placebo”. In:American Psychologist72(7) (2017), pp. 679–688.doi:https: //doi.org/10.1037/amp0000118

    work page doi:10.1037/amp0000118 2017

  50. [50]

    A systematic review of fMRI neurofeedback reporting and effects in clinical populations

    Anita Tursic et al. “A systematic review of fMRI neurofeedback reporting and effects in clinical populations”. In:NeuroImage: Clinical28 (2020), p. 102496.doi:https://doi.org/ 10.1016/j.nicl.2020.102496

    arXiv 2020

  51. [51]

    Analysis of dominant classes in universal adversarial perturbations

    Jon Vadillo, Roberto Santana, and Jose A Lozano. “Analysis of dominant classes in universal adversarial perturbations”. In:Knowledge-Based Systems236 (2022), p. 107719.doi:https: //doi.org/10.1016/j.knosys.2021.107719

    arXiv 2022

  52. [52]

    Predictors of neurofeedback training outcome: A systematic review

    Lydia Anna Weber, Thomas Ethofer, and Ann-Christine Ehlis. “Predictors of neurofeedback training outcome: A systematic review”. In:NeuroImage: Clinical27 (2020), p. 102301.doi: https://doi.org/10.1016/j.nicl.2020.102301

    arXiv 2020

  53. [53]

    Is Neuroimaging Measuring Information in the Brain?

    Lee de-Wit et al. “Is Neuroimaging Measuring Information in the Brain?” In:Psychonomic Bulletin & Review23.5 (2016), pp. 1415–1428.doi:10.3758/s13423-016-1002-0

    work page doi:10.3758/s13423-016-1002-0 2016

  54. [54]

    28, 2017

    Han Xiao, Kashif Rasul, and Roland Vollgraf.Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms. 28, 2017

    2017

  55. [55]

    Sparse estimation automatically selects voxels relevant for the decod- ing of fMRI activity patterns

    Okito Yamashita et al. “Sparse estimation automatically selects voxels relevant for the decod- ing of fMRI activity patterns”. In:NeuroImage42.4 (2008), pp. 1414–1429.doi:10.1016/j. neuroimage.2008.05.050. 27

    doi:10.1016/j 2008