Introduction
Hallucinations—sensory perceptions in the absence of external stimuli—are typically associated with neuropsychiatric conditions such as schizophrenia or Parkinson’s disease. However, they also manifest in healthy individuals under sensory deprivation and in amputees as phantom limb sensations. While computational models for visual and auditory hallucinations exist, the somatosensory (tactile) system has remained under-explored. This research demonstrates that tactile hallucinations can be computationally modeled by applying a homeostatic mechanism to a Deep Boltzmann Machine (DBM) interfaced with artificial skin. The study suggests that hallucinations are not merely random noise but are driven by a biological drive to maintain stable neuronal activity levels.
System Architecture and Homeostasis
The researchers developed a framework that integrates neuro-inspired machine learning with tactile hardware. The system utilizes a biologically inspired artificial skin composed of 18 hexagonal cells, each equipped with multi-modal sensors (force, proximity, and temperature).
At the core of the model is a Deep Boltzmann Machine (DBM), a generative neural network capable of learning data distributions. To simulate sensory deprivation, the network is presented with a "blank" input (all zeros). The authors implemented a homeostatic rule: if a neuron's activity falls below a learned baseline, its internal bias is increased to boost excitability. This hypersensitivity causes the network to mistake internal generative activity for external reality, effectively "pulling" previously learned patterns out of the vacuum of sensory deprivation.
Experimental Dataset
The model was evaluated using a custom-generated tactile dataset recorded from the 18-cell hexagonal grid. The dataset consisted of three distinct triangle patterns, each comprising three active cells. These sparse, geometric patterns provided a clear baseline for measuring how effectively the network could reconstruct or "hallucinate" specific shapes from corrupted or null inputs.
Results and Discussion
The study found that homeostasis acts as a powerful generative force. Rather than producing random activity, the homeostatic process significantly improved the quality of latent representations, helping the network converge toward meaningful, learned states even when receiving zero external stimulus.
Key quantitative findings include:
- Performance Gains: Using the Dice-coefficient ($Q$) to measure similarity, homeostasis increased accuracy from a baseline of $Q \approx 0.50$ (blank input) to $Q \approx 0.72$ for circular connectivity.
- Convergence: The homeostatic adjustment typically saturated after 1,200 time steps, at which point the network consistently "felt" one of the trained triangle patterns.
- Clinical Alignment: The timescale of these adjustments in the model aligns with clinical observations of phantom limb syndrome, which often emerges within 24 hours to a week following amputation.
Conclusion
This work provides a computational bridge between biological homeostasis and the emergence of tactile hallucinations. By demonstrating that "ghost" sensations can arise from a network's attempt to maintain activity during deprivation, the research offers a potential neurological explanation for phantom limb syndrome. Future work will aim to scale the system to thousands of sensors and incorporate continuous, non-binary signals to better mimic real biological skin. Ultimately, these models may inform clinical treatments, such as targeted electrical stimulation to prevent the homeostatic triggers that lead to distressing hallucinations.
Summary updated