Ultra-slow fluctuations are a hallmark of spontaneous cortical activity. We examine the hypothesis that these dynamics arise from recurrent neuronal networks operating near a phase-transition point, a state marked by "critical slowing down". In such networks, a subtle shift toward criticality should selectively amplify slow fluctuations, providing a lever that can switch the cortex from quiet rest into self-generated behavior. Using a simple random recurrent network, we reproduce this amplification effect. The resulting spectra closely match intracranial electroencephalography from human visual cortex recorded during rest and during category-specific visual free recall. In particular, the model captures the experimentally observed enhancement of slow fluctuations during recall. These simulations reveal a parsimonious mechanism that explains spontaneous ultra-slow activity and enables rapid transitions between spontaneous states, suggesting that dynamic tuning toward criticality may be a general strategy by which cortical networks enter a generative mode.

Sensory deprivation has long been known to cause hallucinations or "phantom" sensations, the most common of which is tinnitus induced by hearing loss, affecting 10-20% of the population. An observable hearing loss, causing auditory sensory deprivation over a band of frequencies, is present in over 90% of people with tinnitus. Existing plasticity-based computational models for tinnitus are usually driven by homeostatic mechanisms, modeled to fit phenomenological findings. Here, we use an objective-driven learning algorithm to model an early auditory processing neuronal network, e.g., in the dorsal cochlear nucleus. The learning algorithm maximizes the network's output entropy by learning the feed-forward and recurrent interactions in the model. We show that the connectivity patterns and responses learned by the model display several hallmarks of early auditory neuronal networks. We further demonstrate that attenuation of peripheral inputs drives the recurrent network towards its critical point and transition into a tinnitus-like state. In this state, the network activity resembles responses to genuine inputs even in the absence of external stimulation, namely, it "hallucinates" auditory responses. These findings demonstrate how objective-driven plasticity mechanisms that normally act to optimize the network's input representation can also elicit pathologies such as tinnitus as a result of sensory deprivation.

Decoding brain activity for brain computer interfaces (BCI) is a challenging machine learning task. Many BCIs are designed in a synchronous setting, which allows to output a command from the brain to the computer at pre-determined, specific time points. The task becomes substantially more demanding when designing asynchronous BCIs, which allow control over the interface at any given time. Most BCIs use static classifiers, in which the readout is driven only by the current input. However, due to the complex nature of ongoing brain activity, dynamic classifiers, which capture the temporal dynamics of the input signal, are likely to be more suitable for the task. To examine this notion, we designed a classification framework based on Echo State Networks (ESNs), which have demonstrated strong performance in similar tasks in other domains. We evaluated the performance of this approach on pre-recorded EEG data of an asynchronous motor imagery (MI) BCI task, and compared our approach to a conventional method. ESN-based architectures exhibited superior performance in comparison. Our findings suggest that ESNs may prove useful in other BCI paradigms in virtue of their ability to reduce the detrimental effect of non-stationarity in brain signals.

The limited capacity of recent memory inevitably leads to partial memory of past stimuli. There is also evidence that behavioral and neural responses to novel or rare stimuli are dependent on one's memory of past stimuli. Thus, these responses may serve as a probe of different individuals' remembering and forgetting characteristics. Here, we utilize two lossy compression models of stimulus sequences that inherently involve forgetting, which in addition to being a necessity under many conditions, also has theoretical and behavioral advantages. One model is based on a simple stimulus counter and the other on the Information Bottleneck (IB) framework which suggests a more general, theoretically justifiable principle for biological and cognitive phenomena. These models are applied to analyze a novelty-detection event-related potential commonly known as the P300. The trial-by-trial variations of the P300 response, recorded in an auditory oddball paradigm, were subjected to each model to extract two stimulus-compression parameters for each subject: memory length and representation accuracy. These parameters were then utilized to estimate the subjects' recent memory capacity limit under the task conditions. The results, along with recently published findings on single neurons and the IB model, underscore how a lossy compression framework can be utilized to account for trial-by-trial variability of neural responses at different spatial scales and in different individuals, while at the same time providing estimates of individual memory characteristics at different levels of representation using a theoretically-based parsimonious model.

Time-series analysis is confounded by nonlinear time warping of the data. Traditional methods for joint alignment do not generalize: after aligning a given signal ensemble, they lack a mechanism, that does not require solving a new optimization problem, to align previously-unseen signals. In the multi-class case, they must also first classify the test data before aligning it. Here we propose the Diffeomorphic Temporal alignment Net (DTAN), a learning-based method for time-series joint alignment. Via flexible temporal transformer layers, DTAN learns and applies an input-dependent nonlinear time warping to its input signal. Once learned, DTAN easily aligns previously-unseen signals by its inexpensive forward pass. In a single-class case, the method is unsupervised: the ground-truth alignments are unknown. In the multi-class case, it is semi-supervised in the sense that class labels (but not the ground-truth alignments) are used during learning; in test time, however, the class labels are unknown. As we show, DTAN not only outperforms existing joint-alignment methods in aligning training data but also generalizes well to test data. Our code is available at https://github.com/BGU-CS-VIL/dtan.

The maximum entropy principle is a well established approach to unsupervised optimization. Entropy maximization learning algorithms for single-layered neural networks already exist for the cases in which the number of output neurons is greater or equal to the number of input neurons. These models were successfully employed in various applications, most notably for independent component analysis. In this work, we generalize the maximum entropy principle to a single-layered neural network with fewer output than input neurons. The proposed learning algorithm finds a low-dimensional representation of the data and identifies the independent components within it. In general, such a model must incorporate some prior knowledge of the input distribution; however, we overcome this difficulty using a variational approach. We illustrate the performance of the model through several examples and compare it to other algorithms. While our model achieves similar results to the state-of-the-art algorithm for overdetermined independent component analysis within a similar convergence time, its main advantage lies in its ability to be learned efficiently on-line.

The finding of power law scaling in neural recordings lends support to the hypothesis of critical brain dynamics. However, power laws are not unique to critical systems and can arise from alternative mechanisms. Here, we investigate whether a common time-varying external drive to a set of Poisson units can give rise to neuronal avalanches and exhibit apparent criticality. To this end, we analytically derive the avalanche size and duration distributions, as well as additional measures, first for homogeneous Poisson activity, and then for slowly varying inhomogeneous Poisson activity. We show that homogeneous Poisson activity cannot give rise to power law distributions. Inhomogeneous activity can also not generate perfect power laws, but it can exhibit approximate power laws with cutoffs that are comparable to those typically observed in experiments. The mechanism of generating apparent criticality by time-varying external fields, forces or input may generalize to many other systems like dynamics of swarms, diseases or extinction cascades. Here, we illustrate the analytically derived effects for spike recordings in vivo and discuss approaches to distinguish true from apparent criticality. Ultimately, this requires causal interventions, which allow separating internal system properties from externally imposed ones.

Body language of robot arms, have rarely been explored as a medium of conveying robot intentions. An explorative study was done focusing on two questions: one, if robot arm postures can convey robot intentions, and two, if participants coming in contact with this robot arm for the first time can associate any meaning to the postures without watching the robot in action, or working with it. Thirty five participants took part in this exploratory study. Results show significant effect of the design of postures on the interpretation of its meaning. Postures designed in tempered form of human/animal-like body language could successfully convey the desired intention to a significant number of people in the following categories: ‘Robot giving object in a friendly manner’; ‘Robot is saying Hi!’, ‘Robot has been told not to disturb’. Developers may predefine such postures in a robot to enhance its acceptability.

Synaesthesia is an unusual perceptual experience in which an inducer stimulus triggers a percept in a different domain in addition to its own. To explore the conditions under which synaesthesia evolves, we studied a neuronal network model that represents two recurrently connected neural systems. The interactions in the network evolve according to learning rules that optimize sensory sensitivity. We demonstrate several scenarios, such as sensory deprivation or heightened plasticity, under which synaesthesia can evolve even though the inputs to the two systems are statistically independent and the initial cross-talk interactions are zero. Sensory deprivation is the known causal mechanism for acquired synaesthesia and increased plasticity is implicated in developmental synaesthesia. The model unifies different causes of synaesthesia within a single theoretical framework and repositions synaesthesia not as some quirk of aberrant connectivity, but rather as a functional brain state that can emerge as a consequence of optimising sensory information processing.

Recurrent connections play an important role in cortical function, yet their exact contribution to the network computation remains unknown. The principles guiding the long-term evolution of these connections are poorly understood as well. Therefore, gaining insight into their computational role and into the mechanism shaping their pattern would be of great importance. To that end, we studied the learning dynamics and emergent recurrent connectivity in a sensory network model based on a first-principle information theoretic approach. As a test case, we applied this framework to a model of a hypercolumn in the visual cortex and found that the evolved connections between orientation columns have a "Mexican hat" profile, consistent with empirical data and previous modeling work. Furthermore, we found that optimal information representation is achieved when the network operates near a critical point in its dynamics. Neuronal networks working near such a phase transition are most sensitive to their inputs and are thus optimal in terms of information representation. Nevertheless, a mild change in the pattern of interactions may cause such networks to undergo a transition into a different regime of behavior in which the network activity is dominated by its internal recurrent dynamics and does not reflect the objective input. We discuss several mechanisms by which the pattern of interactions can be driven into this supercritical regime and relate them to various neurological and neuropsychiatric phenomena.

Understanding how populations of neurons encode sensory information is a major goal of systems neuroscience. Attempts to answer this question have focused on responses measured over several hundred milliseconds, a duration much longer than that frequently used by animals to make decisions about the environment. How reliably sensory information is encoded on briefer time scales, and how best to extract this information, is unknown. Although it has been proposed that neuronal response latency provides a major cue for fast decisions in the visual system, this hypothesis has not been tested systematically and in a quantitative manner. Here we use a simple 'race to threshold' readout mechanism to quantify the information content of spike time latency of primary visual (V1) cortical cells to stimulus orientation. We find that many V1 cells show pronounced tuning of their spike latency to stimulus orientation and that almost as much information can be extracted from spike latencies as from firing rates measured over much longer durations. To extract this information, stimulus onset must be estimated accurately. We show that the responses of cells with weak tuning of spike latency can provide a reliable onset detector. We find that spike latency information can be pooled from a large neuronal population, provided that the decision threshold is scaled linearly with the population size, yielding a processing time of the order of a few tens of milliseconds. Our results provide a novel mechanism for extracting information from neuronal populations over the very brief time scales in which behavioral judgments must sometimes be made.

Population rate models provide powerful tools for investigating the principles that underlie the cooperative function of large neuronal systems. However, biophysical interpretations of these models have been ambiguous. Hence, their applicability to real neuronal systems and their experimental validation have been severely limited. In this work, we show that conductance-based models of large cortical neuronal networks can be described by simplified rate models, provided that the network state does not possess a high degree of synchrony. We first derive a precise mapping between the parameters of the rate equations and those of the conductance-based network models for time-independent inputs. This mapping is based on the assumption that the effect of increasing the cell's input conductance on its f-I curve is mainly subtractive. This assumption is confirmed by a single compartment Hodgkin-Huxley type model with a transient potassium A-current. This approach is applied to the study of a network model of a hypercolumn in primary visual cortex. We also explore extensions of the rate model to the dynamic domain by studying the firing-rate response of our conductance-based neuron to time-dependent noisy inputs. We show that the dynamics of this response can be approximated by a time-dependent second-order differential equation. This phenomenological single-cell rate model is used to calculate the response of a conductance-based network to time-dependent inputs.

The principle of maximizing mutual information is applied to learning overcomplete and recurrent representations. The underlying model con(cid:173) sists of a network of input units driving a larger number of output units with recurrent interactions. In the limit of zero noise, the network is de(cid:173) terministic and the mutual information can be related to the entropy of the output units. Maximizing this entropy with respect to both the feed(cid:173) forward connections as well as the recurrent interactions results in simple learning rules for both sets of parameters. The conventional independent components (ICA) learning algorithm can be recovered as a special case where there is an equal number of output units and no recurrent con(cid:173) nections. The application of these new learning rules is illustrated on a simple two-dimensional input example.