Search

We are presenting the Belgian Federal COVID-19 corpus, nicknamed the BeCoS (Belgian CovidSign language) corpus. It consists of the entire archive of official press conferences from the BelgianFederal Government concerning the COVID-19 pandemic. The speakers speak mostly in Dutchor French and occasionally in German, and nearly all speech is accompanied by a deaf signer whoperforms live interpreting from what is being said. We have preprocessed the corpus with speakerdiarisation, applied Belgian Dutch ASR, and post-ASR language identification and punctuationprediction as well as signer diarisation, sign language identification and sign language keypointrecognition. The corpus is made publicly available. ispartof: Computational Linguistics in the Netherlands Journal vol:12 pages:7-17 status: published

Recently, unsupervised local learning, based on Hebb's idea that change in synaptic efficacy depends on the activity of the pre- and postsynaptic neuron only, has shown potential as an alternative training mechanism to backpropagation. Unfortunately, Hebbian learning remains experimental and rarely makes it way into standard deep learning frameworks. In this work, we investigate the potential of Hebbian learning in the context of standard deep learning workflows. To this end, a framework for thorough and systematic evaluation of local learning rules in existing deep learning pipelines is proposed. Using this framework, the potential of Hebbian learned feature extractors for image classification is illustrated. In particular, the framework is used to expand the Krotov-Hopfield learning rule to standard convolutional neural networks without sacrificing accuracy compared to end-to-end backpropagation. The source code is available at https://github.com/Joxis/pytorch-hebbian., Presented as a poster at the NeurIPS 2020 Beyond Backpropagation workshop

Photonic reservoir computing is a promising candidate for low-energy computing at high bandwidths. Despite recent successes, there are bounds to what one can achieve simply by making photonic reservoirs larger. Therefore, a switch from single-reservoir computing to multi-reservoir and even deep physical reservoir computing is desirable. Given that backpropagation can not be used directly to train multi-reservoir systems in our targeted setting, we propose an alternative approach that still uses its power to derive intermediate targets. In this work we report our findings on a conducted experiment to evaluate the general feasibility of our approach by training a network of 3 Echo State Networks to perform the well-known NARMA-10 task using targets derived through backpropagation. Our results indicate that our proposed method is well-suited to train multi-reservoir systems in a efficient way., Submitted to the IEEE International Conference on Rebooting Computing 2019; accepted as a poster, will not be presented though

In this work, we apply the CycleGAN image-to-image translation framework to Mel-scaled log-amplitude spectrograms, successfully realizing audio texture transfer between excerpts from two musically related genres. Such automatic musical transfer could provide music producers and DJs with new creative tools, e.g. to quickly prototype a remix of an existing song in another genre, or to use as an advanced effect during a live performance. We show that meaningful style transfer can be realized using only a limited amount of data and computational resources. A high-quality audio reconstruction is obtained from the generated amplitude spectrogram by simply using the phase of the original audio as an approximation for the phase of the generated spectrogram. This results in a significant quality improvement over traditional phase reconstruction methods.

We present a novel model architecture which leverages deep learning tools to perform exact Bayesian inference on sets of high dimensional, complex observations. Our model is provably exchangeable, meaning that the joint distribution over observations is invariant under permutation: this property lies at the heart of Bayesian inference. The model does not require variational approximations to train, and new samples can be generated conditional on previous samples, with cost linear in the size of the conditioning set. The advantages of our architecture are demonstrated on learning tasks that require generalisation from short observed sequences while modelling sequence variability, such as conditional image generation, few-shot learning, and anomaly detection., NIPS 2018

We present our latest results on silicon photonics neuromorphic information processing based on techniques like reservoir computing. We will discuss aspects like scalability, novel architectures for enhanced power efficiency, as well as all-optical readout. Additionally, we will touch upon new machine learning techniques to operate these integrated readouts.

A major bottleneck in building autonomous robots is the lack of suitable power supplies. Consequently, researchers aim for more energy-efficient locomotion. One way to achieve this is by using compliant materials in the robot bodies. In this work, we introduce Tigrillo, a small, low-cost quadruped robot platform with modular legs. The robot’s legs consist of a simple two-segmented construction which can be configured stiff or passive compliant. Each leg is actuated at the hip with one motor. By thorough optimization of the open loop control signals we demonstrate that compliance not only results in gaits that are more insensitive to parameter variations, but also leads to more energy-efficient gaits.

Introduction: Temporal lobe epilepsy (TLE) is the most common type of pharmaco-resistant epilepsy in adults. During pre-surgical evaluation, EEG is recorded to identify pathological activity such as interictal spikes that can help to identify the epileptogenic zone. However, sometimes no spikes are recorded in the scalp EEG. The characterization of brain network dysfunction in the absence of visible scalp epileptic activity could be very important to improve diagnostic accuracy and treat these patients. Here, we aimed to build a classification system that uses EEG-based directed functional connectivity patterns to assign a patient to one of three classes: left TLE (LTLE), right TLE (RTLE) or healthy control. Methods: Sixty subjects underwent a resting-state high-density EEG recording: 20 LTLE, 20 RTLE and 20 healthy controls. For each subject sixty 1-sec epochs free of artifacts or interictal spikes were selected. The cortical source activity was obtained for 82 regions of interest (ROIs) using an individual head model and distributed linear inverse solution. The whole-brain directed functional connectivity was estimated in the theta, alpha and beta frequency bands using Granger-causal modeling (weighted Partial Directed Coherence[1-3]) applied to the source signal of all ROIs. The summed outflow from each ROI was computed for each frequency band (total of 246 connectivity values per subject). Details of this analysis can be found in [4]. Here, we used logistic regression classifiers which take a linear combination of these connectivity values and compare it with a threshold in order to assign the subject to one of the three given classes. A leave-one-out procedure was used to compute the mean classification accuracy on subjects that are not used as example to build the classifier system. Due to the low amount of recorded subjects compared to the number of connectivity values, a subset was selected for classification purpose. We selected a maximum of six connectivity values making use of a greedy forward selection algorithm: gradually adding connectivity values to improve the classification accuracy. Finally, three classifiers were built: ‘Control vs. LTLE’, ‘Control vs. RTLE’ and ‘LTLE vs. RTLE’. In the final classification system, a new subject is assigned to the class that was most voted for by these three individual classifiers. Subjects for which every class is voted once were classified as undefined. Results: The selected brain regions and corresponding frequency bands for the three classifiers are depicted in Figure 1. The ‘Control vs. RTLE’ classifier achieved an accuracy of 92.5% (sensitivity: 95.0%, specificity 90.0%), the ‘Control vs. LTLE’ classifier an accuracy of 87.5% (sensitivity 90.0%, specificity 85.0%) and the ‘LTLE vs. RTLE’ classifier an accuracy of 80.0% (sensitivity 85.0%, specificity 75.0%). Combining these individual classifiers in one system resulted in the confusion matrix shown in Figure 2. Conclusion: We had previously found significant differences in the summed outflow from regions included in the Default-Mode Network and also known to be important in TLE in patients vs healthy controls using these EEG periods without interictal spikes [4]. Here, the high accuracy achieved in classifying a subject as LTLE, RTLE or healthy control in periods of EEG without visible interictal pathological events further demonstrates the potential of resting-state EEG-based directed functional connectivity for the diagnosis and lateralization of TLE. This could thus constitute a new important clinical biomarker in surgical candidates or even earlier in the course of the disease, especially to lateralize the epilepsy.

Motor control systems in the brain humans and mammals are hierarchically organised, with each level controlling increasingly complex motor actions. Each level is controlled by the higher levels and also receives sensory and/or proprioceptive feedback. Through learning, this hierarchical structure adapts to its body, its sensors and the way these interact with the environment. An even more integrated view is taken in morphological or embodied computation. On the one hand, there is both biological and mechanical (robotics) evidence that a properly chosen body morphology can drastically facilitate control when the body dynamics naturally generate low level motion primitives. On the other hand, several papers have used physical bodies as reservoirs in a reservoir computing setup. In some cases, reservoir computing was used as an easy way to obtain robust linear feedback controllers for locomotion. In other cases, the body dynamics of soft robots were shown to perform general computations in response to some input stimulation. In general, very specific highly compliant bodies were used. We present recent results on two open questions regarding the way morphological computation could be exploited in biological motor control. Generally, when reservoir computing has been used to exploit body dynamics for computation, the desired output signals were known. Clearly, in biological locomotion, the learning does not enforce specific muscle actuation signals. Instead, it rewards desirable forms of motion and penalizes undesirable ones. We show how a biologically plausible learning rule, reward modulated Hebbian learning, can enable the incorporation of compliant body dynamics into the control hierarchy, resulting in robust motor control. Despite the many successes with using physical bodies as reservoirs, the relationship between compliance and computational power has hardly been investigated. Although biological bodies are partially compliant, they also have a very specific structure and many rigid parts. It therefore remains unclear to what extent this type of bodies can help in motor control. In our research, we use compliant four legged robots to address this issue. We present first results that indicate that for such robots, linear feedback of proprioceptive signals alone is often not sufficient to result in stable gait control. In addition, a first comparison of different levels of compliance indicate that a well chosen level of compliance can drastically simplify motor control, compared to both, too little and too much compliance, and that the body should therefore be considered as an integral part of the control.

Motor control systems in the brain of humans and mammals are hierarchically organised, with each level controlling increasingly complex motor actions. Each level is controlled by the higher levels and also receives sensory and/or proprioceptive feedback. Through learning, this hierarchical structure adapts to its body, its sensors and the way these interact with the environment. An even more integrated view is taken in morphological or embodied computation. On the one hand, there is both biological and mechanical (robotics) evidence that a properly chosen body morphology can drastically facilitate control when the body dynamics naturally generate low level motion primitives. On the other hand, several papers have used robot bodies as reservoirs in a reservoir computing setup. In some cases, reservoir computing was used as an easy way to obtain robust linear feedback controllers for locomotion. In other cases, the body dynamics of soft robots were shown to perform general computations in response to some input stimulation. In general, very specific highly compliant bodies were used. At Ghent University’s Reservoir Lab, we have previously used reservoir computing to generate locomotion on quite different robot platforms: the highly compliant tensegrity robot Recter and the far less compliant quadruped robot Oncilla and a new low cost modular quadruped puppy robot. In all cases, we succeeded in generating stable gaits. However, not surprisingly, not all robot bodies are equally suitable to help generating their own motor actuations. As a result, the reservoir computing principle alone was not always sufficient. We present an overview of our experience with these different robot platforms and give practical guidelines for applying physical reservoir computing to new robots. We finally discuss some perspectives on a more systematic evaluation between body morphology, compliance and the complexity of generating stable gaits for locomotion.

Recently there has been a lot of interest in density modeling with Deep Generative Models. So far, these models have arguably been relatively successful on tasks such as modeling handwritten digits (MNIST), small image patches (e.g., 8 by 8 pixels) and other low-dimensional datasets. However, convincingly modeling higher dimensional data such as small images (e.g., 32 by 32 pixels and higher) is still a big unsolved problem. In this work we will extend and apply Deep Gaussian Mixture Models (DGMMs) to this task, by introducing locally connected transformations. Similarly to convolutions in deep neural networks, local connectivity in DGMMs allow us to train faster and with less overfitting than fully connected networks when applied to images. Our experiments show the benefits of using locally-connected Deep GMMs and give new insights on modeling higher dimensional images.