Marine mammal communication is a complex field, hindered by the diversity of vocalizations and environmental factors. Accurate classification of these vocalizations is critical for understanding species behavior, monitoring population trends, and assessing the impact of human activities on marine life. However, current classification approaches face significant challenges due to the wide range of vocalization types and environmental noise. The Watkins Marine Mammal Sound Database (WMMD) constitutes a comprehensive labeled dataset employed in machine learning applications. Nevertheless, the methodologies for data preparation, preprocessing, and classification documented in the literature exhibit considerable variability and are typically not applied to the dataset in its entirety. This study initially undertakes a concise review of the state-of-the-art benchmarks pertaining to the dataset, with a particular focus on clarifying data preparation and preprocessing techniques. Subsequently, we explore the utilization of the Wavelet Scattering Transform (WST) and Mel spectrogram as preprocessing mechanisms for feature extraction. In this paper, we introduce WhaleNet (Wavelet Highly Adaptive Learning Ensemble Network), a sophisticated deep ensemble architecture for the classification of marine mammal vocalizations, leveraging both WST and Mel spectrogram for enhanced feature discrimination. By addressing the inconsistencies in data preparation and by utilizing advanced preprocessing techniques, our approach provides a more robust framework for classifying marine mammal vocalizations, which is essential for conservation efforts and behavioral research. By integrating the insights derived from WST and Mel representations, we achieved an improvement in classification accuracy by 8−10% over existing architectures, corresponding to a classification accuracy of 97.61%.
Federated learning (FL) enables the training of machine learning models across decentralized clients while preserving data privacy. However, the presence of anomalous or corrupted clients—such as those with faulty sensors or nonrepresentative data distributions—can significantly degrade model performance. Detecting such clients without accessing raw data remains a key challenge. We propose wavelet and fourier representations for FL), a detection algorithm that labels malicious clients before training, using locally computed compressed representations derived from either the wavelet scattering transform (WST) or the Fourier transform (FT). Both approaches provide low-dimensional, task-agnostic embeddings suitable for unsupervised client separation. A lightweight detector, trained on a distilled public dataset, performs the labeling with minimal communication and computational overhead. While both transforms enable effective detection, WST offers theoretical advantages, such as noninvertibility and stability to local deformations, that make it particularly well-suited to federated scenarios. Experiments on benchmark datasets demonstrate that our method improves both detection accuracy and downstream classification performance compared to existing FL anomaly detection algorithms, validating its effectiveness as an offline alternative to online detection strategies. Source code for this article is publicly available at https://github.com/davedleo/Waffle
Accurate and efficient humidity estimation is critical for various applications, particularly with the rise of IoT devices and smart sensors where computational resources are limited. A common issue with many conventional humidity sensors is their slow response dynamics, which restricts their performance in applications requiring rapid, real-time data. This work introduces a novel Machine Learning algorithm for fast humidity estimation, based on analyzing the voltage discharge dynamics across microelectrodes, computationally frugal yet accurate estimation method suitable for resource-constrained environments. We propose a Physics-Informed Dimensionality Reduction (PIDR) methodology that leverages an underlying physical model – specifically, anomalous diffusion governing the electric discharge between microelectrodes – to extract low-dimensional, physically meaningful features from high-dimensional sensor time series data. Neural Parameter Estimation (NPE), trained effectively on synthetically augmented data guided by limited experimental observations, maps the voltage discharge curves to the anomalous diffusion parameters of the physical model. These parameters, representing a low dimensional physical space, are then fed alongside temperature readings into a compact Artificial Neural Network (ANN) for final humidity prediction. This two-stage, physics-aware architecture significantly reduces model complexity. Experimental results demonstrate the effectiveness of the PIDR approach, achieving high prediction accuracy while demanding significantly less computational effort and training data than traditional parameter estimation techniques or purely data-driven deep learning models applied to raw data. Our study highlights the successful integration of physical principles with machine learning for developing efficient, interpretable, and robust AI solutions tailored for smart sensors and sustainable IoT applications.