Search

The human sleep-cycle has been divided into discrete sleep stages that can be recognized in electroencephalographic (EEG) and other bio-signals by trained specialists or machine learning systems. It is however unclear whether these human-defined stages can be re-discovered with unsupervised methods of data analysis, using only a minimal amount of generic pre-processing. Based on EEG data, recorded overnight from sleeping human subjects, we investigate the degree of clustering of the sleep stages using the General Discrimination Value as a quantitative measure of class separability. Virtually no clustering is found in the raw data, even after transforming the EEG signals of each thirty-second epoch from the time domain into the more informative frequency domain. However, a Principal Component Analysis (PCA) of these epoch-wise frequency spectra reveals that the sleep stages separate significantly better in the low-dimensional sub-space of certain PCA components. In particular the component $C_1(t)$ can serve as a robust, continuous 'master variable' that encodes the depth of sleep and therefore correlates strongly with the 'hypnogram', a common plot of the discrete sleep stages over time. Moreover, $C_1(t)$ shows persistent trends during extended time periods where the sleep stage is constant, suggesting that sleep may be better understood as a continuum. These intriguing properties of $C_1(t)$ are not only relevant for understanding brain dynamics during sleep, but might also be exploited in low-cost single-channel sleep tracking devices for private and clinical use.

Data classification, the process of analyzing data and organizing it into categories, is a fundamental computing problem of natural and artificial information processing systems. Ideally, the performance of classifier models would be evaluated using unambiguous data sets, where the 'correct' assignment of category labels to the input data vectors is unequivocal. In real-world problems, however, a significant fraction of actually occurring data vectors will be located in a boundary zone between or outside of all categories, so that perfect classification cannot even in principle be achieved. We derive the theoretical limit for classification accuracy that arises from the overlap of data categories. By using a surrogate data generation model with adjustable statistical properties, we show that sufficiently powerful classifiers based on completely different principles, such as perceptrons and Bayesian models, all perform at this universal accuracy limit. Remarkably, the accuracy limit is not affected by applying non-linear transformations to the data, even if these transformations are non-reversible and drastically reduce the information content of the input data. We compare emerging data embeddings produced by supervised and unsupervised training, using MNIST and human EEG recordings during sleep. We find that categories are not only well separated in the final layers of classifiers trained with back-propagation, but to a smaller degree also after unsupervised dimensionality reduction. This suggests that human-defined categories, such as hand-written digits or sleep stages, can indeed be considered as 'natural kinds'.

How is information processed in the brain during perception? Mechanistic insight is achieved only when experiments are employed to test formal or computational models. In analogy to lesion studies, phantom perception may serve as a vehicle to understand the fundamental processing principles underlying auditory perception. With a special focus on tinnitus -- as the prime example of auditory phantom perception -- we review recent work at the intersection of artificial intelligence, psychology, and neuroscience. In particular, we discuss why everyone with tinnitus suffers from hearing loss, but not everyone with hearing loss suffers from tinnitus. We argue that the increase of sensory precision due to Bayesian inference could be caused by intrinsic neural noise and lead to a prediction error in the cerebral cortex. Hence, two fundamental processing principles - being ubiquitous in the brain - provide the most explanatory power for the emergence of tinnitus: predictive coding as a top-down, and stochastic resonance as a complementary bottom-up mechanism. We conclude that both principles play a crucial role in healthy auditory perception., Comment: arXiv admin note: substantial text overlap with arXiv:2010.01914

In neural networks with identical neurons, the matrix of connection weights completely describes the network structure and thereby determines how it is processing information. However, due to the non-linearity of these systems, it is not clear if similar microscopic connection structures also imply similar functional properties, or if a network is impacted more by macroscopic structural quantities, such as the ratio of excitatory and inhibitory connections (balance), or the ratio of non-zero connections (density). To clarify these questions, we focus on motifs of three binary neurons with discrete ternary connection strengths, an important class of network building blocks that can be analyzed exhaustively. We develop new, permutation-invariant metrics to quantify the structural and functional distance between two given network motifs. We then use multidimensional scaling to identify and visualize clusters of motifs with similar structural and functional properties. Our comprehensive analysis reveals that the function of a neural network is only weakly correlated with its microscopic structure, but depends strongly on the balance of the connections.

Reliable determination of sensory thresholds is the holy grail of signal detection theory. However, there exists no gold standard for the estimation of thresholds based on neurophysiological parameters, although a reliable estimation method is crucial for both scientific investigations and clinical diagnosis. Whenever it is impossible to communicate with the subjects, as in studies with animals or neonatales, thresholds have to be derived from neural recordings. In such cases when the threshold is estimated based on neuronal measures, the standard approach is still the subjective setting of the threshold to the value where at least a "clear" neuronal signal is detectable. These measures are highly subjective, strongly depend on the noise, and fluctuate due to the low signal-to-noise ratio near the threshold. Here we show a novel method to reliably estimate physiological thresholds based on neurophysiological parameters. Using surrogate data, we demonstrate that fitting the responses to different stimulus intensities with a hard sigmoid function, in combination with subsampling, provides a robust threshold value as well as an accurate uncertainty estimate. This method has no systematic dependence on the noise and does not even require samples in the full dynamic range of the sensory system. It is universally applicable to all types of sensory systems, ranging from somatosensory stimulus processing in the cortex to auditory processing in the brain stem.

Deep neural networks typically outperform more traditional machine learning models in their ability to classify complex data, and yet is not clear how the individual hidden layers of a deep network contribute to the overall classification performance. We thus introduce a Generalized Discrimination Value (GDV) that measures, in a non-invasive manner, how well different data classes separate in each given network layer. The GDV can be used for the automatic tuning of hyper-parameters, such as the width profile and the total depth of a network. Moreover, the layer-dependent GDV(L) provides new insights into the data transformations that self-organize during training: In the case of multi-layer perceptrons trained with error backpropagation, we find that classification of highly complex data sets requires a temporal {\em reduction} of class separability, marked by a characteristic 'energy barrier' in the initial part of the GDV(L) curve. Even more surprisingly, for a given data set, the GDV(L) is running through a fixed 'master curve', independently from the total number of network layers. Furthermore, applying the GDV to Deep Belief Networks reveals that also unsupervised training with the Contrastive Divergence method can systematically increase class separability over tens of layers, even though the system does not 'know' the desired class labels. These results indicate that the GDV may become a useful tool to open the black box of deep learning.

The acoustic startle reflex (ASR) that can be induced by a loud sound stimulus can be used as a versatile tool to, e.g., estimate hearing thresholds or identify subjective tinnitus percepts in rodents. These techniques are based on the fact that the ASR amplitude can be suppressed by a pre-stimulus of lower, non-startling intensity, an effect named pre-pulse inhibition (PPI). For hearing threshold estimation, pure tone pre-stimuli of varying amplitudes are presented before an intense noise burst serving as startle stimulus. The amount of suppression of the ASR amplitude as a function of the pre-stimulus intensity can be used as a behavioral correlate to determine the hearing ability. For tinnitus assessment, the pure-tone pre-stimulus is replaced by a gap of silence in a narrowband noise background, a paradigm termed GPIAS (gap-pre-pulse inhibition of the acoustic startle response). A proper application of these paradigms depend on a reliable measurement of the ASR amplitudes, an exact stimulus presentation in terms of frequency and intensity. Here we introduce a novel open source solution for the construction of a low-cost ASR setup for the above mentioned purpose. The complete software for data acquisition and stimulus presentation is written in Python 3.6 and is provided as an anaconda package. Furthermore, we provide a construction plan for the sensory system based on low-cost hardware components. Exemplary data show that the ratios (1-PPI) of the pre and post trauma ASR amplitudes can be well described by a lognormal distribution being in good accordance to previous studies with already established setups. Hence, the open access solution described here will help to further establish the ASR method in many laboratories and thus facilitate and standardize research in animal models of tinnitus or hearing loss.

We present a new statistical method to analyze multichannel steady-state local field potentials (LFP) recorded within different sensory cortices of different rodent species. Our spatiotemporal multi-dimensional cluster statistics (MCS) method enables statistical analyzing and comparing clusters of data points in n-dimensional space. We demonstrate that using this approach stimulus-specific attractor-like spatiotemporal activity patterns can be detected and be significantly different from each other during stimulation with long-lasting stimuli. Our method may be applied to other types of multichannel neuronal data, like EEG, MEG or spiking responses and used for the development of new read-out algorithms of brain activity and by that opens new perspectives for the development of brain-computer interfaces.

Subjective tinnitus (ST) is generally assumed to be a consequence of hearing loss (HL). In animal studies acoustic trauma can lead to behavioral signs of ST, in human studies ST patients without increased hearing thresholds were found to suffer from so called hidden HL. Additionally, ST is correlated with pathologically increased spontaneous firing rates and neuronal hyperactivity (NH) along the auditory pathway. Homeostatic plasticity (HP) has been proposed as a compensation mechanism leading to the development of NH, arguing that after HL initially decreased mean firing rates of neurons are subsequently restored by increased spontaneous rates. However all HP models fundamentally lack explanatory power since the function of keeping mean firing rate constant remains elusive as does the benefit this might have in terms of information processing. Furthermore the neural circuitry being able to perform the comparison of preferred with actual mean firing rate remains unclear. Here we propose an entirely new interpretation of ST related development of NH in terms of information theory. We suggest that stochastic resonance (SR) plays a key role in short- and long-term plasticity within the auditory system and is the ultimate cause of NH and ST. SR has been found ubiquitous in neuroscience and refers to the phenomenon that sub-threshold, unperceivable signals can be transmitted by adding noise to sensor input. We argue that after HL, SR serves to lift signals above the increased hearing threshold, hence subsequently decreasing thresholds again. The increased amount of internal noise is the correlate of the NH, which finally leads to the development of ST, due to neuronal plasticity along the auditory pathway. We demonstrate the plausibility of our hypothesis by using a computational model and provide exemplarily findings of human and animal studies that are consistent with our model.

Successful detection of weak signals is a universal challenge for numerous technical and biological systems and crucially limits signal transduction and transmission. Stochastic resonance (SR) has been identified to have the potential to tackle this problem, namely to enable non-linear systems to detect small, otherwise sub-threshold signals by means of added non-zero noise. This has been demonstrated within a wide range of systems in physical, technological and biological contexts. Based on its ubiquitous importance, numerous theoretical and technical approaches aim at an optimization of signal transduction based on SR. Several quantities like mutual information, signal-to-noise-ratio, or the cross-correlation between input stimulus and resulting detector response have been used to determine optimal noise intensities for SR. The fundamental shortcoming with all these measures is that knowledge of the signal to be detected is required to compute them. This dilemma prevents the use of adaptive SR procedures in any application where the signal to be detected is unknown. We here show that the autocorrelation function (AC) of the detector response fundamentally overcomes this drawback. For a simplified model system, the equivalence of the output AC with the measures mentioned above is proven analytically. In addition, we test our approach numerically for a variety of systems comprising different input signals and different types of detectors. The results indicate a strong similarity between mutual information and output AC in terms of the optimal noise intensity for SR. Hence, using the output AC to adaptively vary the amount of added noise in order to maximize information transmission via SR might be a fundamental processing principle in nature, in particular within neural systems which could be implemented in future technical applications.