pith. sign in

arxiv: 1906.10592 · v2 · pith:KC42LXXXnew · submitted 2019-06-25 · 💻 cs.NE · cs.LG· q-bio.NC

Tactile Hallucinations on Artificial Skin Induced by Homeostasis in a Deep Boltzmann Machine

Michael Deistler , Yagmur Yener , Florian Bergner , Pablo Lanillos , Gordon Cheng This is my paper

Pith reviewed 2026-05-25 15:51 UTC · model grok-4.3

classification 💻 cs.NE cs.LGq-bio.NC
keywords tactile hallucinationsdeep Boltzmann machinehomeostasisartificial skinsensory deprivationgenerative modelslatent representationsperception
0
0 comments X

The pith

Homeostasis in a Deep Boltzmann Machine induces hallucinations of previously learned tactile patterns on artificial skin without sensory input.

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

The paper applies a homeostasis rule that changes neuron excitability during sensory deprivation inside a Deep Boltzmann Machine trained on patterns from artificial skin. This produces hallucinations of the learned patterns when input is removed and also improves how well the network reconstructs its internal latent states. The work treats this as a computational account of tactile hallucinations reported in neurological disorders and amputees. A reader would see it as evidence that perception can be generative and that homeostatic adjustment is one concrete way the generation occurs.

Core claim

In a Deep Boltzmann Machine trained on tactile patterns from artificial skin, introducing homeostasis during periods without sensory input causes the network to generate hallucinations of previously learned patterns, induces the formation of meaningful latent representations, and significantly increases the quality of the reconstruction of these latent states.

What carries the argument

Homeostasis rule that adjusts neuron excitability in the Deep Boltzmann Machine to simulate effects of sensory deprivation.

If this is right

  • Hallucinations of learned patterns appear on the artificial skin when sensory input is absent.
  • Meaningful latent representations form under the homeostasis rule.
  • Reconstruction quality of the latent states increases significantly.
  • The model supplies one possible explanation for the nature of tactile hallucinations.
  • Homeostatic processes are indicated as a candidate mechanism underlying such hallucinations.

Where Pith is reading between the lines

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

  • The same homeostasis mechanism could be inserted into other generative models to test whether it produces hallucinations in visual or auditory domains.
  • Comparing the DBM output directly to spike recordings from deprived skin nerves would provide a concrete test of biological plausibility.
  • Prosthetic skins might need explicit homeostatic compensation to suppress unwanted phantom patterns during periods of low input.

Load-bearing premise

The Deep Boltzmann Machine together with the chosen homeostasis rule is a sufficient model of the generative processes that produce tactile hallucinations in biological systems.

What would settle it

If neural recordings from biological tactile pathways during sensory deprivation show different patterns from those generated by the homeostatic DBM, or if removing homeostasis does not reduce reconstruction quality, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 1906.10592 by Florian Bergner, Gordon Cheng, Michael Deistler, Pablo Lanillos, Yagmur Yener.

Figure 1
Figure 1. Figure 1: Deep Boltzmann Machine (DBM) for modelling tactile hallucinations. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Biologically inspired multi-modal skin [14], [24]. The skin is [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training accuracy curves as measured by the performance measure [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Decoded states when clamping to one of the training patterns for circular connectivity. The measure of performance Q is 0.87. Training Homeostasis Blanking the input [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of the measures [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: Correlation between Qloss and Qgain. On the x-axis, the difference in reconstruction performance between clamping to a training pattern or a blank pattern Qloss is shown. On the y-axis, the gain in performance through hallucinations Qgain is shown. The two traces correspond to the two types of connectivity namely circular (orange), and linear (blue). In all cases, there is a clear correlation (respective c… view at source ↗
Figure 8
Figure 8. Figure 8: Demonstration of the reconstructed patterns given a sample trace dur [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
read the original abstract

Perceptual hallucinations are present in neurological and psychiatric disorders and amputees. While the hallucinations can be drug-induced, it has been described that they can even be provoked in healthy subjects. Understanding their manifestation could thus unveil how the brain processes sensory information and might evidence the generative nature of perception. In this work, we investigate the generation of tactile hallucinations on biologically inspired, artificial skin. To model tactile hallucinations, we apply homeostasis, a change in the excitability of neurons during sensory deprivation, in a Deep Boltzmann Machine (DBM). We find that homeostasis prompts hallucinations of previously learned patterns on the artificial skin in the absence of sensory input. Moreover, we show that homeostasis is capable of inducing the formation of meaningful latent representations in a DBM and that it significantly increases the quality of the reconstruction of these latent states. Through this, our work provides a possible explanation for the nature of tactile hallucinations and highlights homeostatic processes as a potential underlying mechanism.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces homeostasis as an excitability change in a Deep Boltzmann Machine trained on tactile patterns to model hallucinations on artificial skin. It claims that, in the absence of sensory input, the homeostasis rule induces hallucinations of previously learned patterns, promotes formation of meaningful latent representations, and significantly improves reconstruction quality of those states, thereby supplying a possible mechanistic explanation for tactile hallucinations in biological systems.

Significance. If the reported effects prove robust under quantitative controls and the DBM dynamics can be shown to be at least a plausible proxy for somatosensory generative processes, the work would link homeostatic regulation to perceptual inference in a generative model, adding a computational perspective on how sensory deprivation can produce structured hallucinations. The absence of metrics, ablations, and biological grounding currently limits the strength of that contribution.

major comments (2)
  1. [Abstract] Abstract: the assertion that homeostasis 'significantly increases the quality of the reconstruction of these latent states' is presented without quantitative metrics, error bars, statistical tests, or ablation controls, which is load-bearing for the central performance claim.
  2. [Abstract] Abstract: the further claim that the model 'provides a possible explanation for the nature of tactile hallucinations' rests on the untested assumption that the chosen homeostasis rule and resulting sampling statistics constitute a plausible proxy for biological generative mechanisms; no comparison to electrophysiological recordings (e.g., firing-rate distributions or correlations under sensory deprivation) or to alternative generative architectures is supplied.
minor comments (1)
  1. The abstract supplies only qualitative outcomes; the methods section should explicitly state the homeostasis update rule, all free parameters, training protocol, and the precise definition of 'meaningful latent representations' to permit independent verification.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed review and the opportunity to address the concerns raised. Below we respond point-by-point to the major comments.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion that homeostasis 'significantly increases the quality of the reconstruction of these latent states' is presented without quantitative metrics, error bars, statistical tests, or ablation controls, which is load-bearing for the central performance claim.

    Authors: The abstract is a concise summary; the manuscript contains the supporting quantitative results, including reconstruction-error metrics, comparisons with and without homeostasis, and visualizations of the effects. To strengthen the abstract we will add a short clause referencing the observed quantitative improvement in reconstruction quality. revision: yes

  2. Referee: [Abstract] Abstract: the further claim that the model 'provides a possible explanation for the nature of tactile hallucinations' rests on the untested assumption that the chosen homeostasis rule and resulting sampling statistics constitute a plausible proxy for biological generative mechanisms; no comparison to electrophysiological recordings (e.g., firing-rate distributions or correlations under sensory deprivation) or to alternative generative architectures is supplied.

    Authors: The work demonstrates, within a generative model (DBM) trained on tactile data, that a biologically motivated homeostasis rule produces structured hallucinations of learned patterns when external input is removed. This supplies a concrete computational mechanism that can be viewed as one possible explanation for tactile hallucinations under sensory deprivation. The discussion section already notes the biological motivation of the homeostasis rule and its relation to perceptual inference; we do not claim the model is a direct replica of cortical circuitry, only that it illustrates a plausible generative-process account. revision: no

Circularity Check

0 steps flagged

No significant circularity; simulation results are direct consequences of added mechanism

full rationale

The paper introduces homeostasis as an external rule applied to a pre-trained DBM and reports simulation outcomes (hallucinations and improved reconstruction) under zero-input conditions. These are not reductions by construction to the inputs, as the rule is an independent addition whose effects are measured rather than presupposed. No self-definitional equations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the derivation chain. The biological explanatory claim is framed as a possible model rather than a forced mathematical identity, leaving the work self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the untested premise that a DBM plus a simple excitability shift is an adequate generative model of biological tactile perception. No external benchmarks or independent data are supplied in the abstract.

axioms (1)
  • domain assumption A DBM trained on tactile patterns can serve as a proxy for the generative processes underlying biological tactile perception.
    Invoked when the authors equate model output to 'tactile hallucinations' without additional validation.

pith-pipeline@v0.9.0 · 5711 in / 1232 out tokens · 16561 ms · 2026-05-25T15:51:51.081607+00:00 · methodology

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

34 extracted references · 34 canonical work pages

  1. [1]

    The genesis of phanton (deenervation) hallucinations: An hypothesis,

    P. V . Rabins, “The genesis of phanton (deenervation) hallucinations: An hypothesis,” International journal of geriatric psychiatry , vol. 9, no. 10, pp. 775–777, 1994

    work page 1994

  2. [2]

    Hallucinations experienced by visually impaired: Charles bonnet syndrome,

    L. Pang, “Hallucinations experienced by visually impaired: Charles bonnet syndrome,” Optometry and Vision Science , vol. 93, no. 12, p. 1466, 2016

    work page 2016

  3. [3]

    Tactile hallucinations in parkinson’s disease,

    G. F ´enelon, S. Thobois, A.-M. Bonnet, E. Broussolle, and F. Tison, “Tactile hallucinations in parkinson’s disease,” Journal of neurology , vol. 249, no. 12, pp. 1699–1703, 2002

    work page 2002

  4. [4]

    Hallucinations in schizophrenia,

    K. Mueser, A. Bellack, and E. Brady, “Hallucinations in schizophrenia,” Acta Psychiatrica Scandinavica , vol. 82, no. 1, pp. 26–29, 1990

    work page 1990

  5. [5]

    A review on neural network models of schizophrenia and autism spectrum disorder,

    P. Lanillos, D. Oliva, A. Philippsen, Y . Yamashita, Y . Nagai, and G. Cheng, “A review on neural network models of schizophrenia and autism spectrum disorder,” arXiv preprint arXiv:1906.10015 , 2019

    work page arXiv 1906

  6. [6]

    Pavlovian conditioning– induced hallucinations result from overweighting of perceptual priors,

    A. R. Powers, C. Mathys, and P. Corlett, “Pavlovian conditioning– induced hallucinations result from overweighting of perceptual priors,” Science, vol. 357, no. 6351, pp. 596–600, 2017

    work page 2017

  7. [7]

    Book review: Neural network models of schizophrenia,

    R. E. Hoffman and T. H. McGlashan, “Book review: Neural network models of schizophrenia,” The Neuroscientist, vol. 7, no. 5, pp. 441–454, 2001

    work page 2001

  8. [8]

    Hallucinations in charles bonnet syndrome induced by homeostasis: a deep boltzmann machine model,

    P. Series, D. P. Reichert, and A. J. Storkey, “Hallucinations in charles bonnet syndrome induced by homeostasis: a deep boltzmann machine model,” in Advances in Neural Information Processing Systems , 2010, pp. 2020–2028

    work page 2010

  9. [9]

    A neural network model of phantom limbs,

    M. Spitzer, P. B ¨ohler, M. Weisbrod, and U. Kischka, “A neural network model of phantom limbs,” Biological cybernetics , vol. 72, no. 3, pp. 197–206, 1995

    work page 1995

  10. [10]

    Ac- tive inference, sensory attenuation and illusions,

    H. Brown, R. A. Adams, I. Parees, M. Edwards, and K. Friston, “Ac- tive inference, sensory attenuation and illusions,” Cognitive processing, vol. 14, no. 4, pp. 411–427, 2013

    work page 2013

  11. [11]

    A computational model unifies apparently contradictory findings concerning phantom pain,

    K. J. Bostr ¨om, M. H. De Lussanet, T. Weiss, C. Puta, and H. Wag- ner, “A computational model unifies apparently contradictory findings concerning phantom pain,” Scientific reports, vol. 4, p. 5298, 2014

    work page 2014

  12. [12]

    Complex visual hallucinations in the visually impaired: the charles bonnet syn- drome,

    G. J. Menon, I. Rahman, S. J. Menon, and G. N. Dutton, “Complex visual hallucinations in the visually impaired: the charles bonnet syn- drome,” Survey of ophthalmology , vol. 48, no. 1, pp. 58–72, 2003

    work page 2003

  13. [13]

    The charles bonnet syndrome:phantom visual images,

    G. Schultz and R. Melzack, “The charles bonnet syndrome:phantom visual images,” Perception, vol. 20, no. 6, pp. 809–825, 1991

    work page 1991

  14. [14]

    Humanoid multimodal tactile-sensing modules,

    P. Mittendorfer and G. Cheng, “Humanoid multimodal tactile-sensing modules,” IEEE Transactions on Robotics , vol. 27, no. 3, pp. 401–410, 2011

    work page 2011

  15. [15]

    Computational models of hallucinations,

    R. Jardri and S. Den `eve, “Computational models of hallucinations,” in The neuroscience of hallucinations . Springer, 2013, pp. 289–313

    work page 2013

  16. [16]

    A neural model of delusions and hallucinations in schizophrenia,

    E. Ruppin, J. A. Reggia, and D. Horn, “A neural model of delusions and hallucinations in schizophrenia,” in Advances in Neural Information Processing Systems, 1995, pp. 149–156

    work page 1995

  17. [17]

    The computational anatomy of psychosis,

    R. A. Adams, K. E. Stephan, H. R. Brown, C. D. Frith, and K. J. Friston, “The computational anatomy of psychosis,” Frontiers in psychiatry , vol. 4, p. 47, 2013

    work page 2013

  18. [18]

    Hierarchical somatosensory processing,

    Y . Iwamura, “Hierarchical somatosensory processing,” Current opinion in neurobiology, vol. 8, no. 4, pp. 522–528, 1998

    work page 1998

  19. [19]

    Distributed hierarchical processing in the primate cerebral cortex

    D. J. Felleman and D. E. Van, “Distributed hierarchical processing in the primate cerebral cortex.” Cerebral cortex (New York, NY: 1991) , vol. 1, no. 1, pp. 1–47, 1991

    work page 1991

  20. [20]

    The self-tuning neuron: synaptic scaling of excitatory synapses,

    G. G. Turrigiano, “The self-tuning neuron: synaptic scaling of excitatory synapses,” Cell, vol. 135, no. 3, pp. 422–435, 2008

    work page 2008

  21. [21]

    The dialectic of hebb and homeostasis,

    ——, “The dialectic of hebb and homeostasis,” Philosophical Transac- tions of the Royal Society B: Biological Sciences , vol. 372, no. 1715, p. 20160258, 2017

    work page 2017

  22. [22]

    Homeostatic plasticity and stdp: keeping a neuron’s cool in a fluctuating world,

    A. J. Watt and N. S. Desai, “Homeostatic plasticity and stdp: keeping a neuron’s cool in a fluctuating world,” Frontiers in synaptic neuroscience, vol. 2, p. 5, 2010

    work page 2010

  23. [23]

    Synaptic plasticity in neural networks needs homeostasis with a fast rate detector,

    F. Zenke, G. Hennequin, and W. Gerstner, “Synaptic plasticity in neural networks needs homeostasis with a fast rate detector,” PLoS computational biology, vol. 9, no. 11, p. e1003330, 2013

    work page 2013

  24. [24]

    Event-based signaling for large-scale artificial robotic skin - realization and performance evalua- tion,

    F. Bergner, E. Dean-Leon, and G. Cheng, “Event-based signaling for large-scale artificial robotic skin - realization and performance evalua- tion,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2016, pp. 4918–4924

    work page 2016

  25. [25]

    Phantom limb pain: theories and therapies,

    S. R. Weeks, V . C. Anderson-Barnes, and J. W. Tsao, “Phantom limb pain: theories and therapies,” The neurologist, vol. 16, no. 5, pp. 277– 286, 2010

    work page 2010

  26. [26]

    A comparison of neural network training methods for text classification

    A. de Andrade, “A comparison of neural network training methods for text classification.”

    work page

  27. [27]

    Adaptive robot body learning and estimation through predictive coding,

    P. Lanillos and G. Cheng, “Adaptive robot body learning and estimation through predictive coding,” Intelligent Robots and Systems (IROS), 2018 IEEE/RSJ International Conference on , 2018

    work page 2018

  28. [28]

    Vector reconstruction from firing rates,

    E. Salinas and L. Abbott, “Vector reconstruction from firing rates,” Journal of computational neuroscience , vol. 1, no. 1-2, pp. 89–107, 1994

    work page 1994

  29. [29]

    Neural population encoding of touch intensity,

    O. Franz ´en, D. Kenshalo, and G. Essick, “Neural population encoding of touch intensity,” in Information processing in the somatosensory system . Springer, 1991, pp. 71–80

    work page 1991

  30. [30]

    Robotic homunculus: Learning of artificial skin representation in a humanoid robot motivated by primary somatosensory cortex,

    M. Hoffmann, Z. Straka, I. Farka ˇs, M. Vavreˇcka, and G. Metta, “Robotic homunculus: Learning of artificial skin representation in a humanoid robot motivated by primary somatosensory cortex,” IEEE Transactions on Cognitive and Developmental Systems , vol. 10, no. 2, pp. 163–176, 2017

    work page 2017

  31. [31]

    Information processing in dynamical systems: Founda- tions of harmony theory,

    P. Smolensky, “Information processing in dynamical systems: Founda- tions of harmony theory,” COLORADO UNIV AT BOULDER DEPT OF COMPUTER SCIENCE, Tech. Rep., 1986

    work page 1986

  32. [32]

    Training products of experts by minimizing contrastive divergence,

    G. E. Hinton, “Training products of experts by minimizing contrastive divergence,” Neural computation, vol. 14, no. 8, pp. 1771–1800, 2002

    work page 2002

  33. [33]

    Spike timing–dependent plasticity: a hebbian learning rule,

    N. Caporale and Y . Dan, “Spike timing–dependent plasticity: a hebbian learning rule,” Annu. Rev. Neurosci., vol. 31, pp. 25–46, 2008

    work page 2008

  34. [34]

    Forming sparse representations by local anti-hebbian learn- ing,

    P. F ¨oldiak, “Forming sparse representations by local anti-hebbian learn- ing,” Biological cybernetics, vol. 64, no. 2, pp. 165–170, 1990. APPENDIX A. Boltzmann Machines In general, all of the described neural network architectures are generative models. This means that, after a training phase, in which input patterns are being presented to the network, ...

    work page 1990