National audience

Graph distillation endeavors to transfer knowledge from large, complex teacher models, such as Graph Neural Networks (GNNs), to smaller, more efficient student models such as Multi-Layer Perceptrons (MLPs). Our work is motivated by a critical observation: as the complexity and depth of the teacher GNN increase, the performance of the distilled student MLP models tends to decline significantly. This issue is not limited to a specific distillation method but is a prevalent challenge across various GNN to MLP approaches. To address this gap, we introduce a novel prompt-driven graph distillation framework that enhances the student model’s input space by appending or integrating learned prompts with the original features. These prompts, derived from the teacher’s input, hidden and output knowledge, provide additional context that assists the MLP student in distilling information, thereby circumventing the need to directly capture any teacher’s knowledge into the simplest single-layer MLP. Our empirical experiment on benchmark graph datasets reveals a counter-intuitive phenomenon: the more capable the teacher, the greater the student’s ability to utilize prompts to outperform the teacher. This finding highlights the strength of our approach: a well-prompted student can indeed surpass its teacher, at achieving the best performance with an accuracy of 87.2% on the Cora standard split with a single-layer MLP, while also maintaining efficiency and robustness. In addition to testing on benchmark graph datasets, we applied our framework to the TUH EEG Epilepsy Corpus (TUEP), condensing complex GNN models to simple MLPs while achieving good performance and efficiency in epilepsy classification.

Early detection of mild cognitive impairment (MCI) is critical, as timely interventions during this transitional phase can slow or even prevent progression to Alzheimer’s disease. Eye-tracking (ET) signals recorded during virtual reality (VR)-based cognitive tasks present strong potential for MCI detection, as they integrate the immersive multisensory environment of VR with the rich temporal and behavioral information of ET signals. In this study, we built a VR-based system incorporating four Stroop-inspired tasks, proposed a corresponding ET signal dataset, and introduced a lightweight network for MCI detection. Using our system, a 38-subject MCI dataset was constructed, including 17 individuals with MCI and 21 healthy control individuals. However, the inherently non-stationary and redundant characteristics of ET signals often pose challenges for accurate classification. To overcome these issues, the proposed network, KWS-TA-CNN-PE, includes four key components: (1) Kymatio-based wavelet scattering transform (KWS) to calculate the wavelet scattering coefficients, generating time-robust features and minimizing memory requirements via a depth-first traversal strategy; (2) temporal attention (TA), which dynamically prioritizes the most informative time steps and suppresses noise in ET signals; (3) a one-dimensional convolutional neural network (CNN) that extracts localized temporal patterns; and (4) a probabilistic ensemble (PE) strategy that combines the outputs across the four tasks for final classification. Experimental results of our KWS-TA-CNN-PE network, using leave-one-subject-out cross-validation and a blind-test protocol, exhibit strong performance with accuracies of 0.8629 and 0.8621, respectively. The promising results highlight the clinical potential of the proposed system, dataset, and network.

Multivariate autoregressive modeling is widely considered in neuroscience, especially when effective connectivity is concerned. In high-dimensional space, the conventional least-squares estimation of the autoregressive coefficients is no more consistent, hence the interest in regularizing the solution. Therefore, regularized approaches have been developed such as the Least Absolute Shrinkage Selection Operator (LASSO) promoting sparsity, which performs well provided that it is combined with the extended Bayesian Information Criterion (eBIC) to jointly estimate the MVAR model order and penalty parameter. Unfortunately, this need for eBIC requires much more computation time, making such approaches unsuitable for identifying high-dimensional MVAR models. The method proposed in this letter, named SOCAR (Simultaneous identification of the Order and the Coefficients of multivariate AutoRegressive models), estimates both the order and the coefficients of the model by combining mixed-norm regularization with classical sparse priors (e.g. LASSO). Weighted penalties are minimized in order to drive more easily sparsity to the right place. Experiments carried out on simulated signals show that SOCAR offers a good compromise between performance and numerical complexity.

In the context of epilepsy, understanding the interactions among different brain regions during epileptic seizures is often crucial to identify brain regions involved in the seizure onset. In this letter, we propose a dictionary-based method to detect and reconstruct the relationships between brain regions during epileptic seizures based on electroencephalographic signals. To this end, a proximal alternating linearized minimization algorithm is used to quantify the relations between each of the dictionary candidates. Then, by introducing an adaptive threshold, an estimation of the involvement of the different brain regions in the seizure onset is performed, which proves to be consistent with the clinical expertise.

Epilepsy is one of the most common neurological diseases, which can seriously affect the patient's psychological well-being and quality of life. An accurate and reliable seizure prediction system can generate alarm before epileptic seizures to provide patients and their caregivers with sufficient time to take appropriate action. This study proposes an efficient seizure prediction system based on deep learning in order to anticipate the onset of the seizure as early as possible. Handcrafted features extracted based on the prior knowledge and hidden deep features are complementarily fused through the feature fusion module, and then the hybrid features are fed into the multiplicative long short-term memory (MLSTM) to explore the temporal dependency in EEG signals. A one-dimensional channel attention mechanism is implemented to emphasize the more representative information in the multi-channel output of the MLSTM. Finally, a transfer learning strategy is proposed to transfer the weights of the base model trained on the EEG data of all patients to the target patient model, and the latter is then continuously trained using the EEG data of the target patient. The proposed method achieves an average sensitivity of 95.56% and a false positive rate (FPR) of 0.27/h on the SWEC-ETHZ intracranial EEG data. For the more challenging CHB-MIT scalp EEG database, an average sensitivity of 89.47% and a FPR of 0.34/h are obtained. Experimental results demonstrate that the proposed method has good robustness and generalization ability in both intracranial and scalp EEG signals.

In this paper, our aim is to investigate the control of bias accumulation when estimating mutual information from nearest neighbors non-parametric approach with continuously distributed random data. Using a multidimensional Taylor series expansion, a general relationship between the estimation bias and neighborhood size for plug-in entropy estimator is established without any assumption on the data for two different norms. When applied with the maximum norm, our theoretical analysis explains experimental simulation tests drawn in existing literature. In the experiments, two different strategies are tested and compared to estimate mutual information on independent and dependent simulated signals.

The context of this work is the analysis of depth electroencephalographic signals recorded with depth electrodes during seizures in patients with drug-resistant epilepsy. Usually, different phases are observed during the seizure process and we aim to determine how cerebral structures get involved during these phases, in particular whether some structures can “drive” other ones. To this end, we consider a pair of signals and use transfer entropy which needs beforehand to choose efficiently the size of two conditioning vectors built on the past values of these signals. In this contribution, we extend a partial mutual information based technique, first developed for monochannel prediction models, to the case of two channels. Experimental results on signals generated either by a linear autoregressive model or by a physiology-based model of coupled neuronal populations support the relevance of the proposed approach.

Phase slope index is a measure which aims at detecting causal relation of interdependence in multivariate time series. One drawback of this approach relies in its incapability to distinguish the direct and indirect relations. So, in order to identify only direct relations, we propose to replace the ordinary coherence function used in the phase slope index with the partial coherence. Furthermore, we consider and compare two estimators of the coherence functions, the first one based on Fourier transform and the second one on an autoregressive model. In order to cope with the difficult issue of bidirectional flow, which cannot be addressed by the coherence based phase slope index, we propose another index based on the directed transfer function. Experimental results support the relevance of the new indices, both based on autoregressive modeling, in multivariate time series.

For subjective assessment of speech quality in codecs, a reference impairment system is required to introduce controlled degradations to calibrate subjective evaluation. A reference system provides a convenient means for making meaningful comparisons between subjective test results across laboratories and can be viewed as a scale on which mean opinion scores are projected, this scale being supposed to cover the whole range of quality. Nowadays, standardized anchor systems do not fit any more the degradations brought by the present codecs. This paper aims at offering new reference signals simulating the defaults of codecs currently used on telecommunication networks. Twenty wideband codecs are compared through dissimilarity tests. A multidimensional scaling technique allows us to define a four-dimensional perceptive space that appears stable for male and female talkers. A verbalization task suggests qualifying the degradations perceived by the listeners with the following attributes: muffle, background noise, noise on speech, and hiss, each conveyed by one dimension. These dimensions are correlated with objective measures such as spectral centroid, energy in the silent part in the high frequency sub-band, ratio of brightness between deterministic part and residual part of the signal and spectral correlation coefficient. New reference signals are produced and a phase of validation suggests a perceptive space quite coherent with the original one.

Synchronization of a neuronal response to a given periodic stimulus is usually measured by Goldberg and Brown's vector strength metric (1969). The limitations of this index are discussed and two alternative indexes are proposed: the corrected vector strength index and the phase variance index. Using a standard data set, the three metrics are compared and the results indicate that the vector strength index should be interpreted with great care. The proposed new indexes are both more relevant to 'synchronization'. Moreover, these two indexes can be tuned to show a large variation of behavior depending on the user's idea of synchronization.

This paper proposes an approach to model the conversational speech quality and an optimisation of this approach to some practical cases. A new subjective test has been designed to study the relationship between conversational speech quality and talking, listening and interaction qualities, when facing echo and delay. The results show that, in these conditions of degradation, the subjective conversational mean opinion scores (MOS) given by subjects can be estimated from the talking and listening quality scores by a multiple linear regression, which coefficients are calculated to minimize the mean squared error (MSE) between subjective and estimated conversational scores. We show the validity of the proposed method in predicting the conversational quality scores for the conditions assessed in this subjective test. For this, a comparison between subjective and estimated conversational quality scores is performed, by means of correlation coefficient and mean absolute error

This letter deals with the problem of speech detection in the context of hands-free telecommunications. The principle of the detectors studied is based on the coherence function (ordinary or partial coherence), which is usually computed from the periodogram (first approach). The present work is aimed at proposing a new approach to compute the coherence from the eigenspectra with a view to reduce the bias. A statistical analysis is carried out to estimate the bias of both coherence estimators. Both approaches are compared experimentally, and results show the superiority of the new estimator.

This paper describes a nonlinear acoustic echo cancellation algorithm, mainly focused on loudspeaker distortions. The proposed system is composed of two distinct modules organized in a cascaded structure: a nonlinear module based on polynomial Volterra filters models the loudspeaker, and a second module of standard linear filtering identifies the impulse response of the acoustic path. The tracking of the overall system model is achieved by a modified normalized-least mean square algorithm for which equations are derived. Stability conditions are given, and particular attention is placed on the transient behavior of cascaded filters. Finally, results of real data recorded with Alcatel GSM material are presented.

The modern telecommunications field is concerned with freedom and, in this context, hands-free systems offer subscribers the possibility of talking more naturally, without using a handset. This new type of use leads to new problems which were negligible in traditional telephony, namely the superposition of noise and echo on the speech signal. To solve these problems and provide a quality that is sufficient for telecommunications, combined reduction of these disturbances is required. This paper presents a summary of the solutions retained for this dual reduction in the context of mono-channel and two-channel sound pick-ups.

Our concern is the development of speech enhancement techniques for hands-free audio terminals, including two major problems: noise reduction and acoustic echo cancellation. The objective is to find a joint structure to get a near-end speech signal with minimum distortion and low levels of echo and noise. To solve the two problems, a new promising technique is investigated and tested in real conditions.

This correspondence deals with the problem of noise reduction for hands-free communications when two microphones are in use. We consider two methods usually adopted in the case of uncorrelated noises. The first one is based on the coherence function and the second on Wiener filtering. We propose to modify the previous methods by including a cross power spectrum estimation to take the presence of some correlated noise components into account. Objective results show a significant improvement in comparison with the basic structures.

In hands-free communications, the speech signal to be transmitted is disturbed by ambient noise and acoustic echo. So, processing to reduce these disturbances must be performed before transmission. Classical solutions are cascaded structures where the acoustic echo canceller precedes or follows the noise reduction system. We have proposed a new joint system where noise reduction preprocessing allows one to improve the performance of the acoustic echo canceller. This preprocessing reduces the noise but distorts the original echo. This paper presents an optimization of the preprocessing. Objective results in terms of the echo return loss enhancement and gain are presented.

The problem of speech enhancement and mainly noise reduction in speech remains a key-point of hand-free telecommunications. A great number of techniques have been already put forward and for a few years an auditory model has been investigated in noise reduction. A new approach for enhancing a speech signal degraded by uncorrelated stationary additive noise is developed. In this approach the simultaneous masking effect of the human ear is exploited. Two states "noise masked/noise unmasked" are derived from a noise masking threshold computed using a rough estimate of the speech signal. Then a speech signal estimator is proposed as a weighted sum of the individual estimators in each state. The gain in the signal to noise ratio (SNR) and a distortion measure indicate some improvement in real noise conditions. Subjectively this improvement is noticeable only at high input SNRs.

This letter deals with the problem of speech enhancement for mobile radio communications when two noisy observations are available. The basic idea consists of using the coherence function as a pertinent criterion to distinguish the useful signal from the disturbing noises when noises are decorrelated. There, we propose an optimization of a noise-reduction method based on the coherence function and compare the two methods objectively and by listening tests.>

The estimation of a signal when two observations (signal+noise) are available is addressed. The target application is the enhancement of noisy speech for mobile radio applications. Noises are considered to be spatially decorrelated. A maximum likelihood problem is formulated to estimate the parameters required to cancel the noise, which is solved by the estimate-maximize algorithm. In order to control and enhance this iterative technique, the authors apply spectral constraints (line spectrum pairs) to improve speech quality. Their method is tested on real noisy speech signals recorded in a car. Results relative to the gain on the signal-to-noise ratio and cepstral distances are presented to evaluate the proposed method.>

The authors' concern is the estimation of a signal when two observations signal+noise are available. The intended application is the enhancement of the noisy speech recorded in a car. They present three methods, each consisting of two stages. The first stage identifies a transfer function between the signals or between noises. The second stage acts as a noise canceller. A comparative study between the first and the second method is made to evaluate the constraints for each. An evaluation of the performances of the three methods is made on two observations of noisy speech recorded in a car.>