This paper introduces WhisperSeg, utilizing the Whisper Transformer pre-trained for Automatic Speech Recognition (ASR) for human and animal Voice Activity Detection (VAD). Contrary to traditional methods that detect human voice or animal vocalizations from a short audio frame and rely on careful threshold selection, WhisperSeg processes entire spectrograms of long audio and generates plain text representations of onset, offset, and type of voice activity. Processing a longer audio context with a larger network greatly improves detection accuracy from few labeled examples. We further demonstrate a positive transfer of detection performance to new animal species, making our approach viable in the data-scarce multi-species setting.1

The abstracts of scientific papers typically contain both premises (e.g., background and observations) and conclusions.Although conclusion sentences are highlighted in structured abstracts, in non-structured abstracts the concluding information is not explicitly marked, which makes the automatic segmentation of conclusions from scientific abstracts a challenging task.In this work, we explore Normalized Mutual Information (NMI) as a means for abstract segmentation.We consider each abstract as a recurrent cycle of sentences and place two segmentation boundaries by greedily optimizing the NMI score between the two segments, assuming that conclusions are strongly semantically linked with preceding premises.On nonstructured abstracts, our proposed unsupervised approach GreedyCAS achieves the best performance across all evaluation metrics; on structured abstracts, GreedyCAS outperforms all baseline methods measured by P k .The strong correlation of NMI to our evaluation metrics reveals the effectiveness of NMI for abstract segmentation.1

Event detection improves when events are captured by two different modalities rather than just one. But to train detection systems on multiple modalities is challenging, in particular when there is abundance of unlabelled data but limited amounts of labeled data. We develop a novel self-supervised learning technique for multi- modal data that learns (hidden) correlations between simultaneously recorded microphone (sound) signals and accelerometer (body vibration) signals. The key objective of this work is to learn useful embeddings associated with high performance in downstream event detection tasks when labeled data is scarce and the audio events of interest — songbird vocalizations — are sparse. We base our approach on deep canonical correlation analysis (DCCA) that suffers from event sparseness. We overcome the sparseness of positive labels by first learning a data sampling model from the labelled data and by applying DCCA on the output it produces. This method that we term balanced DCCA (b-DCCA) improves the performance of the unsupervised embeddings on the down-stream supervised audio detection task compared to classsical DCCA. Because data labels are frequently imbalanced, our method might be of broad utility in low-resource scenarios.

Animal vocalisations serve a wide range of vital functions. Although it is possible to record animal vocalisations with external microphones, more insights are gained from miniature sensors mounted directly on animals' backs. We present TinyBird-ML; a wearable sensor node weighing only 1.4 g for acquiring, processing, and wirelessly transmitting acoustic signals to a host system using Bluetooth Low Energy. TinyBird-ML embeds low-latency tiny machine learning algorithms for song syllable classification. To optimize battery lifetime of TinyBird-ML during fault-tolerant continuous recordings, we present an efficient firmware and hardware design. We make use of standard lossy compression schemes to reduce the amount of data sent over the Bluetooth antenna, which increases battery lifetime by 70% without negative impact on offline sound analysis. Furthermore, by not transmitting signals during silent periods, we further increase battery lifetime. One advantage of our sensor is that it allows for closed-loop experiments in the microsecond range by processing sounds directly on the device instead of streaming them to a computer. We demonstrate this capability by detecting and classifying song syllables with minimal latency and a syllable error rate of 7%, using a light-weight neural network that runs directly on the sensor node itself. Thanks to our power-saving hardware and software design, during continuous operation at a sampling rate of 16 kHz, the sensor node achieves a lifetime of 25 hours on a single size 13 zinc-air battery.

We introduce MemSum (Multi-step Episodic Markov decision process extractive SUMmarizer), a reinforcement-learning-based extractive summarizer enriched at each step with information on the current extraction history.When MemSum iteratively selects sentences into the summary, it considers a broad information set that would intuitively also be used by humans in this task: 1) the text content of the sentence, 2) the global text context of the rest of the document, and 3) the extraction history consisting of the set of sentences that have already been extracted.With a lightweight architecture, MemSum obtains state-of-the-art test-set performance (ROUGE) in summarizing long documents taken from PubMed, arXiv, and GovReport.Ablation studies demonstrate the importance of local, global, and history information.A human evaluation confirms the high quality and low redundancy of the generated summaries, stemming from MemSum's awareness of extraction history.

We explore the suitability of self-attention models for character-level neural machine translation.We test the standard transformer model, as well as a novel variant in which the encoder block combines information from nearby characters using convolutions.We perform extensive experiments on WMT and UN datasets, testing both bilingual and multilingual translation to English using up to three input languages (French, Spanish, and Chinese).Our transformer variant consistently outperforms the standard transformer at the character-level and converges faster while learning more robust character-level alignments. 1

Abstractive summarization typically relies on large collections of paired articles and summaries. However, in many cases, parallel data is scarce and costly to obtain. We develop an abstractive summarization system that relies only on large collections of example summaries and non-matching articles. Our approach consists of an unsupervised sentence extractor that selects salient sentences to include in the final summary, as well as a sentence abstractor that is trained on pseudo-parallel and synthetic data, that paraphrases each of the extracted sentences. We perform an extensive evaluation of our method: on the CNN/DailyMail benchmark, on which we compare our approach to fully supervised baselines, as well as on the novel task of automatically generating a press release from a scientific journal article, which is well suited for our system. We show promising performance on both tasks, without relying on any article-summary pairs.

We introduce an image annotation approach for the analysis of volumetric electron microscopic imagery of brain tissue. The core task is to identify and link tubular objects (neuronal fibers) in images taken from consecutive ultrathin sections of brain tissue. In our approach an individual 'flies' through the 3D data at a high speed and maintains eye gaze focus on a single neuronal fiber, aided by navigation with a handheld gamepad controller. The continuous foveation on a fiber of interest constitutes an intuitive means to define a trace that is seamlessly recorded with a desktop eyetracker and transformed into precise 3D coordinates of the annotated fiber (skeleton tracing). In a participant experiment we validate the approach by demonstrating a tracing accuracy of about the respective radiuses of the traced fibers with browsing speeds of up to 40 brain sections per second.

We address the problem of separating two audio sources from a single channel mixture recording. A novel method called Multi Layered Random Forest (MLRF) that learns a binary mask for both the sources is presented. Random Forest (RF) classifiers are trained for each frequency band of a source spectrogram. A specialized set of linear transformations are applied to a local time-frequency (T-F) neighborhood of the mixture that captures relevant local statistics. A sampling method is presented that efficiently samples T-F training bins in each frequency band. We draw equal numbers of dominant (more power) training samples from the two sources for RF classifiers that estimate the Ideal Binary Mask (IBM). An estimated IBM in a given layer is used to train a RF classifier in the next higher layer of the MLRF hierarchy. On average, MLRF performs better than deep Recurrent Neural Networks (RNNs) and Non-Negative Sparse Coding (NNSC) in signal-to-noise ratio (SNR) of reconstructed audio, overall T-F bin classification accuracy, as well as PESQ and STOI scores. Additionally, we demonstrate the ability of the MLRF to correctly reconstruct T-F bins of the target even when the latter has lower power in that frequency band.

This letter addresses the problem of separating two speakers from a single microphone recording. Three linear methods are tested for source separation, all of which operate directly on sound spectrograms: (1) eigenmode analysis of covariance difference to identify spectro-temporal features associated with large variance for one source and small variance for the other source; (2) maximum likelihood demixing in which the mixture is modeled as the sum of two gaussian signals and maximum likelihood is used to identify the most likely sources; and (3) suppression-regression, in which autoregressive models are trained to reproduce one source and suppress the other. These linear approaches are tested on the problem of separating a known male from a known female speaker. The performance of these algorithms is assessed in terms of the residual error of estimated source spectrograms, waveform signal-to-noise ratio, and perceptual evaluation of speech quality scores. This work shows that the algorithms compare favorably to nonlinear approaches such as nonnegative sparse coding in terms of simplicity, performance, and suitability for real-time implementations, and they provide benchmark solutions for monaural source separation tasks.

Recently, there have been remarkable advances in modeling the relationships between the sensory environment, neuronal responses, and behavior. However, most models cannot encompass variable stimulus-response relationships such as varying response latencies and state or context dependence of the neural code. Here, we consider response modeling as a dynamic alignment problem and model stimulus and response jointly by a mixed pair hidden Markov model (MPH). In MPHs, multiple stimulus-response relationships (e.g., receptive fields) are represented by different states or groups of states in a Markov chain. Each stimulus-response relationship features temporal flexibility, allowing modeling of variable response latencies, including noisy ones. We derive algorithms for learning of MPH parameters and for inference of spike response probabilities. We show that some linear-nonlinear Poisson cascade (LNP) models are a special case of MPHs. We demonstrate the efficiency and usefulness of MPHs in simulations of both jittered and switching spike responses to white noise and natural stimuli. Furthermore, we apply MPHs to extracellular single and multi-unit data recorded in cortical brain areas of singing birds to showcase a novel method for estimating response lag distributions. MPHs allow simultaneous estimation of receptive fields, latency statistics, and hidden state dynamics and so can help to uncover complex stimulus response relationships that are subject to variable timing and involve diverse neural codes.

The relationships between neural activity at the single-cell and the population levels are of central importance for understanding neural codes. In many sensory systems, collective behaviors in large cell groups can be described by pairwise spike correlations. Here, we test whether in a highly specialized premotor system of songbirds, pairwise spike correlations themselves can be seen as a simple corollary of an underlying random process. We test hypotheses on connectivity and network dynamics in the motor pathway of zebra finches using a high-level population model that is independent of detailed single-neuron properties. We assume that neural population activity evolves along a finite set of states during singing, and that during sleep population activity randomly switches back and forth between song states and a single resting state. Individual spike trains are generated by associating with each of the population states a particular firing mode, such as bursting or tonic firing. With an overall modification of one or two simple control parameters, the Markov model is able to reproduce observed firing statistics and spike correlations in different neuron types and behavioral states. Our results suggest that song- and sleep-related firing patterns are identical on short time scales and result from random sampling of a unique underlying theme. The efficiency of our population model may apply also to other neural systems in which population hypotheses can be tested on recordings from small neuron groups.

Neurons can have rapidly changing spike train statistics dictated by the underlying network excitability or behavioural state of an animal. To estimate the time course of such state dynamics from single- or multi- ple neuron recordings, we have developed an algorithm that maximizes the likelihood of observed spike trains by optimizing the state lifetimes and the state-conditional interspike-interval (ISI) distributions. Our non- parametric algorithm is free of time-binning and spike-counting prob- lems and has the computational complexity of a Mixed-state Markov Model operating on a state sequence of length equal to the total num- ber of recorded spikes. As an example, we fit a two-state model to paired recordings of premotor neurons in the sleeping songbird. We find that the two state-conditional ISI functions are highly similar to the ones mea- sured during waking and singing, respectively.

The richness and complexity of recurrent cortical circuits is an inexhaustible source of inspiration for thinking about high-level biological computation. In past theoretical studies, constraints on the synaptic connection patterns of threshold-linear networks were found that guaranteed bounded network dynamics, convergence to attractive fixed points, and multistability, all fundamental aspects of cortical information processing. However, these conditions were only sufficient, and it remained unclear which were the minimal (necessary) conditions for convergence and multistability. We show that symmetric threshold-linear networks converge to a set of attractive fixed points if and only if the network matrix is copositive. Furthermore, the set of attractive fixed points is nonconnected (the network is multiattractive) if and only if the network matrix is not positive semidefinite. There are permitted sets of neurons that can be coactive at a stable steady state and forbidden sets that cannot. Permitted sets are clustered in the sense that subsets of permitted sets are permitted and supersets of forbidden sets are forbidden. By viewing permitted sets as memories stored in the synaptic connections, we provide a formulation of long-term memory that is more general than the traditional perspective of fixed-point attractor networks. There is a close correspondence between threshold-linear networks and networks defined by the generalized Lotka-Volterra equations.

There is strong anatomical and physiological evidence that neurons with large receptive fields located in higher visual areas are recurrently connected to neurons with smaller receptive fields in lower areas. We have previously described a minimal neuronal network architecture in which top-down attentional signals to large receptive field neurons can bias and selectively read out the bottom-up sensory information to small receptive field neurons (Hahnloser, Douglas, Mahowald, & Hepp, 1999). Here we study an enhanced model, where the role of attention is to recruit specific inter-areal feedback loops (e.g., drive neurons above firing threshold). We first illustrate the operation of recruitment on a simple example of visual stimulus selection. In the subsequent analysis, we find that attentional recruitment operates by dynamical modulation of signal amplification and response multistability. In particular, we find that attentional stimulus selection necessitates increased recruitment when the stimulus to be selected is of small contrast and of small distance away from distractor stimuli. The selectability of a low-contrast stimulus is dependent on the gain of attentional effects; for example, low-contrast stimuli can be selected only when attention enhances neural responses. However, the dependence of attentional selection on stimulus-distractor distance is not contingent on whether attention enhances or suppresses responses. The computational implications of attentional recruitment are that cortical circuits can behave as winner-take-all mechanisms of variable strength and can achieve close to optimal signal discrimination in the presence of external noise.

Winner-take-all networks have been proposed to underlie many of the brain's fundamental computational abilities. However, not much is known about how to extend the grouping of potential winners in these networks beyond single neuron or uniformly arranged groups of neurons. We show that competition between arbitrary groups of neurons can be realized by organizing lateral inhibition in linear threshold networks. Given a collection of potentially overlapping groups (with the exception of some degenerate cases), the lateral inhibition results in network dynamics such that any permitted set of neurons that can be coactivated by some input at a stable steady state is contained in one of the groups. The information about the input is preserved in this operation. The activity level of a neuron in a permitted set corresponds to its stimulus strength, amplified by some constant. Sets of neurons that are not part of a group cannot be coactivated by any input at a stable steady state. We analyze the storage capacity of such a network for random groups--the number of random groups the network can store as permitted sets without creating too many spurious ones. In this framework, we calculate the optimal sparsity of the groups (maximizing group entropy). We find that for dense inputs, the optimal sparsity is unphysiologically small. However, when the inputs and the groups are equally sparse, we derive a more plausible optimal sparsity. We believe our results are the first steps toward attractor theories in hybrid analog-digital networks.

Integration in the head-direction system is a computation by which hor- izontal angular head velocity signals from the vestibular nuclei are in- tegrated to yield a neural representation of head direction. In the thala- mus, the postsubiculum and the mammillary nuclei, the head-direction representation has the form of a place code: neurons have a preferred head direction in which their firing is maximal [Blair and Sharp, 1995, Blair et al., 1998, ?]. Integration is a difficult computation, given that head-velocities can vary over a large range. Previous models of the head-direction system relied on the assumption that the integration is achieved in a firing-rate-based attractor network with a ring structure. In order to correctly integrate head-velocity signals during high-speed head rotations, very fast synaptic dynamics had to be assumed. Here we address the question whether integration in the head-direction system is possible with slow synapses, for example excitatory NMDA and inhibitory GABA(B) type synapses. For neural networks with such slow synapses, rate-based dynamics are a good approximation of spik- ing neurons [Ermentrout, 1994]. We find that correct integration during high-speed head rotations imposes strong constraints on possible net- work architectures.

Ascribing computational principles to neural feedback circuits is an important problem in theoretical neuroscience. We study symmet(cid:173) ric threshold-linear networks and derive stability results that go beyond the insights that can be gained from Lyapunov theory or energy functions. By applying linear analysis to subnetworks com(cid:173) posed of coactive neurons, we determine the stability of potential steady states. We find that stability depends on two types of eigen(cid:173) modes. One type determines global stability and the other type determines whether or not multistability is possible. We can prove the equivalence of our stability criteria with criteria taken from quadratic programming. Also, we show that there are permitted sets of neurons that can be coactive at a steady state and forbid(cid:173) den sets that cannot. Permitted sets are clustered in the sense that subsets of permitted sets are permitted and supersets of forbidden sets are forbidden. By viewing permitted sets as memories stored in the synaptic connections, we can provide a formulation of long(cid:173) term memory that is more general than the traditional perspective of fixed point attractor networks. A Lyapunov-function can be used to prove that a given set of differential equations is convergent. For example, if a neural network possesses a Lyapunov-function, then for almost any initial condition, the outputs of the neurons converge to a stable steady state. In the past, this stability-property was used to construct attractor networks that associatively recall memorized patterns. Lyapunov theory applies mainly to symmetric networks in which neurons have monotonic activation functions [1, 2]. Here we show that the restriction of activation functions to threshold-linear ones is not a mere limitation, but can yield new insights into the computational behavior of recurrent networks (for completeness, see also [3]). We present three main theorems about the neural responses to constant inputs. The first theorem provides necessary and sufficient conditions on the synaptic weight ma(cid:173) trix for the existence of a globally asymptotically stable set of fixed points. These conditions can be expressed in terms of copositivity, a concept from quadratic pro(cid:173) gramming and linear complementarity theory. Alternatively, they can be expressed in terms of certain eigenvalues and eigenvectors of submatrices of the synaptic weight matrix, making a connection to linear systems theory. The theorem guarantees that the network will produce a steady state response to any constant input. We regard this response as the computational output of the network, and its characterization is the topic of the second and third theorems. In the second theorem, we introduce the idea of permitted and forbidden sets. Under certain conditions on the synaptic weight matrix, we show that there exist sets of neurons that are "forbidden" by the recurrent synaptic connections from being coactivated at a stable steady state, no matter what input is applied. Other sets are "permitted," in the sense that they can be coactivated for some input. The same conditions on the synaptic weight matrix also lead to conditional multistability, meaning that there exists an input for which there is more than one stable steady state. In other words, forbidden sets and conditional multistability are inseparable concepts. The existence of permitted and forbidden sets suggests a new way of thinking about memory in neural networks. When an input is applied, the network must select a set of active neurons, and this selection is constrained to be one of the permitted sets. Therefore the permitted sets can be regarded as memories stored in the synaptic connections. Our third theorem states that there are constraints on the groups of permitted and forbidden sets that can be stored by a network. No matter which learning algorithm is used to store memories, active neurons cannot arbitrarily be divided into permitted and forbidden sets, because subsets of permitted sets have to be permitted and supersets of forbidden sets have to be forbidden. 1 Basic definitions Our theory is applicable to the network dynamics dx· - ' + x · = b· + "W· ·x · 1 dt

It has long been known that lateral inhibition in neural networks can lead to a winner-take-all competition, so that only a single neuron is active at a steady state. Here we show how to organize lateral inhibition so that groups of neurons compete to be active. Given a collection of poten(cid:173) tially overlapping groups, the inhibitory connectivity is set by a formula that can be interpreted as arising from a simple learning rule. Our analy(cid:173) sis demonstrates that such inhibition generally results in winner-take-all competition between the given groups, with the exception of some de(cid:173) generate cases. In a broader context, the network serves as a particular illustration of the general distinction between permitted and forbidden sets, which was introduced recently. From this viewpoint, the computa(cid:173) tional function of our network is to store and retrieve memories as per(cid:173) mitted sets of coactive neurons. In traditional winner-take-all networks, lateral inhibition is used to enforce a localized, or "grandmother cell" representation in which only a single neuron is active [1, 2, 3, 4]. When used for unsupervised learning, winner-take-all networks discover representations similar to those learned by vector quantization [5]. Recently many research efforts have focused on unsupervised learning algorithms for sparsely distributed representations [6, 7]. These algorithms lead to networks in which groups of multiple neurons are coactivated to represent an object. Therefore, it is of great interest to find ways of using lateral inhibition to mediate winner-take-all competition between groups of neurons, as this could be useful for learning sparsely distributed representations. In this paper, we show how winner-take-all competition between groups of neurons can be learned. Given a collection of potentially overlapping groups, the inhibitory connectivity is set by a simple formula that can be interpreted as arising from an online learning rule. To show that the resulting network functions as advertised, we perform a stability analysis. If the strength of inhibition is sufficiently great, and the group organization satisfies certain conditions, we show that the only sets of neurons that can be coactivated at a stable steady state are the given groups and their subsets. Because of the competition between groups, only one group can be activated at a time. In general, the identity of the winning group depends on the initial conditions of the network dynamics. If the groups are ordered by the aggregate input that each receives, the possible winners are those above a cutoff that is set by inequalities to be specified. 1 Basic definitions Let m groups of neurons be given, where group membership is specified by the matrix fl = {I if the ith neuron is in the ath group , ° otherwise (1) We will assume that every neuron belongs to at least one group l, and every group contains at least one neuron. A neuron is allowed to belong to more than one group, so that the groups are potentially overlapping. The inhibitory synaptic connectivity of the network is defined in terms of the group membership, Ji ' = lIm (1 _ ~a ~'!) = {o

We use neural networks with pointer map architectures to provide simple attentional processing in a robotic task. A pointer map comprises a map of neurons that encode a stimulus. Besides global feedback inhibition, the map receives feedback excitation via a small group of pointer neurons that encode the location of a salient stimulus on the map as a vectorial representation. The pointer neurons are able to apply selective processing to a particular region of the network. The robot uses these properties to manoeuver in relation to an attended object. We implemented a controller composed of two pointer maps, and a motor map. The first pointer map reports the direction of a salient obstacle in a one-dimensional map of distance derived from infrared sensors. The second pointer map reports the direction to potential obstacles in a two-dimensional edge-enhanced image derived from a forward looking CCD-camera. These outputs are applied to a motor map, where they bias the motor control signals issued to the robots wheels, according to navigational intentions.