Predictive coding and stochastic resonance as fundamental principles of auditory phantom perception

Brain, Journal Year: 2023, Volume and Issue: 146(12), P. 4809 - 4825

Published: July 27, 2023

Mechanistic insight is achieved only when experiments are employed to test formal or computational models. Furthermore, in analogy lesion studies, phantom perception may serve as a vehicle understand the fundamental processing principles underlying healthy auditory perception. With special focus on tinnitus-as prime example of perception-we review recent work at intersection artificial intelligence, psychology and neuroscience. In particular, we discuss why everyone with tinnitus suffers from (at least hidden) hearing loss, but not loss tinnitus. We argue that intrinsic neural noise generated amplified along pathway compensatory mechanism restore normal based adaptive stochastic resonance. The increase can then be misinterpreted input perceived This formalized Bayesian brain framework, where percept (posterior) assimilates prior prediction (brain's expectations) likelihood (bottom-up signal). A higher mean lower variance (i.e. enhanced precision) shifts posterior, evincing misinterpretation sensory evidence, which further confounded by plastic changes underwrite predictions. Hence, two provide most explanatory power for emergence perceptions: predictive coding top-down resonance complementary bottom-up mechanism. conclude both also play crucial role Finally, context neuroscience-inspired improve contemporary machine learning techniques.

Language: Английский

---

Analysis and visualization of sleep stages based on deep neural networks

Neurobiology of Sleep and Circadian Rhythms, Journal Year: 2021, Volume and Issue: 10, P. 100064 - 100064

Published: March 14, 2021

Automatic sleep stage scoring based on deep neural networks has come into focus of researchers and physicians, as a reliable method able to objectively classify stages would save human resources simplify clinical routines. Due novel open-source software libraries for machine learning, in combination with enormous recent progress hardware development, paradigm shift the field research towards automatic diagnostics might be imminent. We argue that modern learning techniques are not just tool perform classification, but also creative approach find hidden properties physiology. have already developed established algorithms visualize cluster EEG data, facilitating first assessments health terms sleep-apnea consequently reduced daytime vigilance. In following study, we further analyze cortical activity during by determining probabilities momentary stages, represented hypnodensity graphs then computing vectorial cross-correlations different channels. can show this measure serves estimate period length cycles thus help disturbances due pathological conditions.

Language: Английский

---

Multi-modal cognitive maps for language and vision based on neural successor representations

Neurocomputing, Journal Year: 2025, Volume and Issue: unknown, P. 129662 - 129662

Published: Feb. 1, 2025

---

Neural network based successor representations to form cognitive maps of space and language

Scientific Reports, Journal Year: 2022, Volume and Issue: 12(1)

Published: July 4, 2022

How does the mind organize thoughts? The hippocampal-entorhinal complex is thought to support domain-general representation and processing of structural knowledge arbitrary state, feature concept spaces. In particular, it enables formation cognitive maps, navigation on these thereby broadly contributing cognition. It has been proposed that multi-scale successor representations provides an explanation underlying computations performed by place grid cells. Here, we present a neural network based approach learn such representations, its application different scenarios: spatial exploration task supervised learning, reinforcement non-spatial where linguistic constructions have be inferred observing sample sentences. all scenarios, correctly learns approximates structure building representations. Furthermore, resulting firing patterns are strikingly similar experimentally observed cell patterns. We conclude maps network-based structured provide promising way overcome some short comings deep learning towards artificial general intelligence.

Language: Английский

---

Tinnitus is associated with improved cognitive performance and speech perception–Can stochastic resonance explain?

Frontiers in Aging Neuroscience, Journal Year: 2022, Volume and Issue: 14

Published: Dec. 16, 2022

OPINION article Front. Aging Neurosci., 16 December 2022Sec. Neuroinflammation and Neuropathy Volume 14 - 2022 | https://doi.org/10.3389/fnagi.2022.1073149

Language: Английский

---

Neural network based formation of cognitive maps of semantic spaces and the putative emergence of abstract concepts

Scientific Reports, Journal Year: 2023, Volume and Issue: 13(1)

Published: March 4, 2023

Abstract How do we make sense of the input from our sensory organs, and put perceived information into context past experiences? The hippocampal-entorhinal complex plays a major role in organization memory thought. formation navigation cognitive maps arbitrary mental spaces via place grid cells can serve as representation memories experiences their relations to each other. multi-scale successor is proposed be mathematical principle underlying cell computations. Here, present neural network, which learns map semantic space based on 32 different animal species encoded feature vectors. network successfully similarities between species, constructs ‘animal space’ representations with an accuracy around 30% near theoretical maximum regarding fact that all have more than one possible successor, i.e. nearest neighbor space. Furthermore, hierarchical structure, scales maps, modeled representations. We find that, fine-grained vectors are evenly distributed In contrast, coarse-grained highly clustered according biological class, amphibians, mammals insects. This could putative mechanism enabling emergence new, abstract concepts. Finally, even completely new or incomplete represented by interpolation remarkable high up 95%. conclude weighted pointer experiences, may therefore crucial building block include prior knowledge, derive knowledge novel input. Thus, model provides tool complement contemporary deep learning approaches road towards artificial general intelligence.

Language: Английский

---

Deep learning based decoding of single local field potential events

NeuroImage, Journal Year: 2024, Volume and Issue: 297, P. 120696 - 120696

Published: June 21, 2024

How is information processed in the cerebral cortex? In most cases, recorded brain activity averaged over many (stimulus) repetitions, which erases fine-structure of neural signal. However, obviously a single-trial processor. Thus, we here demonstrate that an unsupervised machine learning approach can be used to extract meaningful from electro-physiological recordings on basis. We use auto-encoder network reduce dimensions single local field potential (LFP) events create interpretable clusters different patterns. Strikingly, certain LFP shapes correspond latency differences recording channels. Hence, determine direction flux cortex. Furthermore, after clustering, decoded cluster centroids reverse-engineer underlying prototypical event shapes. To evaluate our approach, applied it both extra-cellular rodents, and intra-cranial EEG humans. Finally, find channel during spontaneous sample realm possible stimulus evoked A finding so far has only been demonstrated for multi-channel population coding.

Language: Английский

---

Simulated transient hearing loss improves auditory sensitivity

Scientific Reports, Journal Year: 2021, Volume and Issue: 11(1)

Published: July 20, 2021

Recently, it was proposed that a processing principle called adaptive stochastic resonance plays major role in the auditory system, and serves to maintain optimal sensitivity even highly variable sound pressure levels. As side effect, case of reduced input, such as permanent hearing loss or frequency specific deprivation, this mechanism may eventually lead perception phantom sounds like tinnitus Zwicker tone illusion. Using computational modeling, biological plausibility already demonstrated. Here, we provide experimental results further support model perception. In particular, Mongolian gerbils were exposed moderate intensity, non-damaging long-term notched noise, which mimics for frequencies within notch. Remarkably, animals developed significantly increased sensitivity, i.e. improved thresholds, centered notch, but not outside addition, most treated with new paradigm showed identical behavioral signs (tinnitus) acoustic trauma induced tinnitus. contrast, broadband noise control condition did show any significant threshold change, nor

Language: Английский

---

Dynamics and Information Import in Recurrent Neural Networks

Frontiers in Computational Neuroscience, Journal Year: 2022, Volume and Issue: 16

Published: April 27, 2022

Recurrent neural networks (RNNs) are complex dynamical systems, capable of ongoing activity without any driving input. The long-term behavior free-running RNNs, described by periodic, chaotic and fixed point attractors, is controlled the statistics connection weights, such as density d non-zero connections, or balance b between excitatory inhibitory connections. However, for information processing purposes, RNNs need to receive external input signals, it not clear which regimes optimal this import. We use both average correlations C mutual I momentary vector next system state quantitative measures import analyze their dependence on network. Remarkably, resulting phase diagrams ( b, ) highly consistent, pointing a link systems information-processing approach systems. Information maximal at “edge chaos,” optimally suited computation, but surprisingly in low-density regime border regime. Moreover, we find completely new type resonance phenomenon, call “Import Resonance” (IR), where shows maximum, i.e., peak-like coupling strength RNN its IR complements previously found Recurrence Resonance (RR), correlation successive states peak certain amplitude noise added system. Both RR can be exploited optimize artificial might also play crucial role biological

Language: Английский

---

Word class representations spontaneously emerge in a deep neural network trained on next word prediction

Published: Dec. 15, 2023

How do humans learn language, and can the first language be learned at all? These fundamental questions are still hotly debated. In contemporary linguistics, there two major schools of thought that give completely opposite answers. According to Chomsky's theory universal grammar, cannot because children not exposed sufficient data in their linguistic environment. contrast, usage-based models assume a profound relationship between structure use. particular, contextual mental processing representations assumed have cognitive capacity capture complexity actual use all levels. The prime example is syntax, i.e., rules by which words assembled into larger units such as sentences. Typically, syntactic expressed sequences word classes. However, it remains unclear whether classes innate, implied or they emerge during acquisition, suggested approaches. Here, we address this issue from machine learning natural perspective. trained an artificial deep neural network on predicting next word, provided consecutive input. Subsequently, analyzed emerging activation patterns hidden layers network. Strikingly, find internal nine-word input cluster according class tenth predicted output, even though did receive any explicit information about training. This surprising result suggests, also human brain, abstract representational categories may naturally consequence predictive coding acquisition.

Language: Английский

---