Search

In this work, we propose a method for domain-incremental learning for audio classification from a sequence of datasets recorded in different acoustic conditions. Fine-tuning a model on a sequence of evolving domains or datasets leads to forgetting of previously learned knowledge. On the other hand, freezing all the layers of the model leads to the model not adapting to the new domain. In this work, our novel dynamic network architecture keeps the shared homogeneous acoustic characteristics of domains, and learns the domain-specific acoustic characteristics in incremental steps. Our approach achieves a good balance between retaining the knowledge of previously learned domains and acquiring the knowledge of the new domain. We demonstrate the effectiveness of the proposed method on incremental learning of single-label classification of acoustic scenes from European cities and Korea, and multi-label classification of audio recordings from Audioset and FSD50K datasets. The proposed approach learns to classify acoustic scenes incrementally with an average accuracy of 71.9% for the order: European cities -> Korea, and 83.4% for Korea -> European cities. In a multi-label audio classification setup, it achieves an average lwlrap of 47.5% for Audioset -> FSD50K and 40.7% for FSD50K -> Audioset., Comment: Accepted to ICASSP 2025

This paper investigates the feasibility of class-incremental learning (CIL) for Sound Event Localization and Detection (SELD) tasks. The method features an incremental learner that can learn new sound classes independently while preserving knowledge of old classes. The continual learning is achieved through a mean square error-based distillation loss to minimize output discrepancies between subsequent learners. The experiments are conducted on the TAU-NIGENS Spatial Sound Events 2021 dataset, which includes 12 different sound classes and demonstrate the efficacy of proposed method. We begin by learning 8 classes and introduce the 4 new classes at next stage. After the incremental phase, the system is evaluated on the full set of learned classes. Results show that, for this realistic dataset, our proposed method successfully maintains baseline performance across all metrics.

Excitable systems can exhibit a variety of dynamics with different complexity, ranging from a single, stable spiral to spiral defect chaos (SDC), during which spiral waves are continuously formed and destroyed. The corresponding reaction-diffusion models, including ones for cardiac tissue, can involve a large number of variables and can be time-consuming to simulate. Here we trained a deep-learning (DL) model using snapshots from a single variable, obtained by simulating a single quasi-periodic spiral wave and SDC using a generic cardiac model. Using the trained DL model, we predicted the dynamics in both cases, using timesteps that are much larger than required for the simulations of the underlying equations. We show that the DL model is able to predict the trajectory of a quasi-periodic spiral wave and that the SDC activaton patterns can be predicted for approximately one Lyapunov time. Furthermore, we show that the DL model accurately captures the statistics of termination events in SDC, including the mean termination time. Finally, we show that a DL model trained using a specific domain size is able to replicate termination statistics on larger domains, resulting in significant computational savings.

In this paper, we propose a method for online domain-incremental learning of acoustic scene classification from a sequence of different locations. Simply training a deep learning model on a sequence of different locations leads to forgetting of previously learned knowledge. In this work, we only correct the statistics of the Batch Normalization layers of a model using a few samples to learn the acoustic scenes from a new location without any excessive training. Experiments are performed on acoustic scenes from 11 different locations, with an initial task containing acoustic scenes from 6 locations and the remaining 5 incremental tasks each representing the acoustic scenes from a different location. The proposed approach outperforms fine-tuning based methods and achieves an average accuracy of 48.8% after learning the last task in sequence without forgetting acoustic scenes from the previously learned locations., Comment: Accepted to EUSIPCO 2024

In this paper, we propose a method for class-incremental learning of potentially overlapping sounds for solving a sequence of multi-label audio classification tasks. We design an incremental learner that learns new classes independently of the old classes. To preserve knowledge about the old classes, we propose a cosine similarity-based distillation loss that minimizes discrepancy in the feature representations of subsequent learners, and use it along with a Kullback-Leibler divergence-based distillation loss that minimizes discrepancy in their respective outputs. Experiments are performed on a dataset with 50 sound classes, with an initial classification task containing 30 base classes and 4 incremental phases of 5 classes each. After each phase, the system is tested for multi-label classification with the entire set of classes learned so far. The proposed method obtains an average F1-score of 40.9% over the five phases, ranging from 45.2% in phase 0 on 30 classes, to 36.3% in phase 4 on 50 classes. Average performance degradation over incremental phases is only 0.7 percentage points from the initial F1-score of 45.2%., Comment: Accepted to ICASSP 2024

In this paper, we propose a method for incremental learning of two distinct tasks over time: acoustic scene classification (ASC) and audio tagging (AT). We use a simple convolutional neural network (CNN) model as an incremental learner to solve the tasks. Generally, incremental learning methods catastrophically forget the previous task when sequentially trained on a new task. To alleviate this problem, we propose independent learning and knowledge distillation (KD) between the timesteps in learning. Experiments are performed on TUT 2016/2017 dataset, containing 4 acoustic scene classes and 25 sound event classes. The proposed incremental learner first solves the ASC task with an accuracy of 94.0%. Next, it learns to solve the AT task with an F1 score of 54.4%. At the same time, its performance on the previous ASC task decreases only by 5.1 percentage points due to the additional learning of the AT task., Comment: Accepted to DCASE2023 Workshop

Spiral waves are ubiquitous spatiotemporal patterns that occur in various excitable systems. In cardiac tissue, the formation of these spiral waves is associated with life-threatening arrhythmias, and, therefore, it is important to study the dynamics of these waves. Tracking the trajectory of a spiral-wave tip can reveal important dynamical features of a spiral wave, such as its periodicity, and its vulnerability to instabilities. We show how to employ the data-driven spectral-decomposition method, called dynamic mode decomposition (DMD), to detect a spiral tip trajectory (TT) in three settings: (1) a homogeneous medium; (2) a heterogeneous medium; and (3) with external noise. We demonstrate that the performance of DMD-based TT (DMDTT) is either comparable to or better than the conventional tip-tracking method called the isopotential-intersection method (IIM) in the cases (1)-(3): (1) Both IIM and DMDTT capture TT patterns at small values of the image-sampling interval $\tau$; however, IIM is more sensitive than DMDTT to the changes in $\tau$. (2) In a heterogeneous medium, IIM yields TT patterns, but with a background of scattered noisy points, which is suppresed in DMDTT. (3) DMDTT is more robust to external noise than IIM. We show, finally, that DMD can be used to reconstruct, and hence predict, the spatiotemporal evolution of spiral waves in the models we study.

Sound event detection (SED) in machine listening entails identifying the different sounds in an audio file and identifying the start and end time of a particular sound event in the audio. SED finds use in various applications such as audio surveillance, speech recognition, and context-based indexing and retrieval of data in a multimedia database. However, in real-life scenarios, the audios from various sources are seldom devoid of any interfering noise or disturbance. In this paper, we test the performance of the You Only Hear Once (YOHO) algorithm on noisy audio data. Inspired by the You Only Look Once (YOLO) algorithm in computer vision, the YOHO algorithm can match the performance of the various state-of-the-art algorithms on datasets such as Music Speech Detection Dataset, TUT Sound Event, and Urban-SED datasets but at lower inference times. In this paper, we explore the performance of the YOHO algorithm on the VOICe dataset containing audio files with noise at different sound-to-noise ratios (SNR). YOHO could outperform or at least match the best performing SED algorithms reported in the VOICe dataset paper and make inferences in less time., Comment: 7 pages, 1 figure, 3 tables, Efficient Natural Language and Speech Processing Workshop, NeurIPS 2021

Spiral waves of excitation in cardiac tissue are associated with life-threatening cardiac arrhythmias. It is, therefore, important to study the electrophysiological factors that affect the dynamics of these spiral waves. By using an electrophysiologically detailed mathematical model of a myocyte (cardiac cell), we study the effects of cellular parameters, such as membrane-ion-channel conductances, on the properties of the action-potential (AP) of a myocyte. We then investigate how changes in these properties, specifically the upstroke velocity and the AP duration (APD), affect the frequency $\omega$ of a spiral wave in the mathematical model that we use for human-ventricular tissue. We find that an increase (decrease) in this upstroke-velocity or a decrease (increase) in the AP duration increases (decreases) $\omega$. We also study how other intercellular factors, such as the fibroblast-myocyte coupling, diffusive coupling strength, and the effective number of neighboring myocytes, modulate $\omega$. Finally, we demonstrate how a spiral wave can drift to a region with a high density of fibroblasts. Our results provide a natural explanation for the anchoring of spiral waves in highly fibrotic regions in fibrotic hearts.

Infarction- or ischaemia-induced cardiac fibrosis can be arrythmogenic. We use mathematcal models for diffuse fibrosis ($\mathcal{DF}$), interstitial fibrosis ($\mathcal{IF}$), patchy fibrosis ($\mathcal{PF}$), and compact fibrosis ($\mathcal{CF}$) to study patterns of fibrotic cardiac tissue that have been generated by new mathematical algorithms. We show that the fractal dimension $\mathbb{D}$, the lacunarity $\mathcal{L}$, and the Betti numbers $\beta_0$ and $\beta_1$ of such patterns are \textit{fibrotic-tissue markers} that can be used to characterise the arrhythmogenicity of different types of cardiac fibrosis. We hypothesize, and then demonstrate by extensive \textit{in silico} studies of detailed mathematical models for cardiac tissue, that the arrhytmogenicity of fibrotic tissue is high when $\beta_0$ is large and the lacunarity parameter $b$ is small., Comment: Manuscript: 6 pages 7 figures Supplemental Material: 4 pages 6 figures

Unbroken and broken spiral waves, in partial-differential-equation (PDE) models for cardiac tissue, are the mathematical analogs of life-threatening cardiac arrhythmias, namely, ventricular tachycardia (VT) and ventricular-fibrillation (VF). We develop a (a) deep-learning method for the detection of unbroken and broken spiral waves and (b) the elimination of such waves, e.g., by the application of low-amplitude control currents in the cardiac-tissue context. Our method is based on a convolutional neural network (CNN) that we train to distinguish between patterns with spiral waves S and without spiral waves NS. We obtain these patterns by carrying out extensive direct numerical simulations (DNS) of PDE models for cardiac tissue in which the transmembrane potential V, when portrayed via pseudocolor plots, displays patterns of electrical activation of types S and NS. We then utilize our trained CNN to obtain, for a given pseudocolor image of V, a heat map that has high intensity in the regions where this image shows the cores of spiral waves. Given this heat map, we show how to apply low-amplitude Gaussian current pulses to eliminate spiral waves efficiently. Our in silico results are of direct relevance to the detection and elimination of these arrhythmias because our elimination of unbroken or broken spiral waves is the mathematical analog of low-amplitude defibrillation.

We compare and contrast spiral- and scroll-wave dynamics in five different mathematical models for cardiac tissue. The first is the TP06 model, due to ten Tusscher and Panfilov, which is based on the Hodgkin-Huxley formalism; the remaining four are Markov-state models, MM1 WT and MM2 WT, for the wild-type (WT) Na channel, and MM1 MUT and MM2 MUT, for the mutant Na channel. Our results are based on extensive direct numerical simulations of waves of electrical activation in these models, in two- and three-dimensional (2D and 3D) homogeneous simulation domains and also in domains with localised heterogeneities, either obstacles with randomly distributed inexcitable regions or mutant cells in a wild-type background. Our study brings out the sensitive dependence of spiral- and scroll-wave dynamics on these five models and the parameters that define them. We also explore the control of spiral-wave turbulence in these models., Comment: 26 pages, 12 figures and 2 Tables