Department Publications

Accurate sleep stage classification is significant for sleep health assessment. In recent years, several machine-learning based sleep staging algorithms have been developed, and in particular, deep-learning based algorithms have achieved performance on par with human annotation. Despite improved performance, a limitation of most deep-learning based algorithms is their black-box behavior, which have limited their use in clinical settings. Here, we propose a cross-modal transformer, which is a transformer-based method for sleep stage classification. The proposed cross-modal transformer consists of a cross-modal transformer encoder architecture along with a multi-scale one-dimensional convolutional neural network for automatic representation learning. The performance of our method is on-par with the state-of-the-art methods and eliminates the black-box behavior of deep-learning models by utilizing the …

The upcoming Sixth Generation (6G) networks aim for fully automated, intelligent network functionalities and services. Therefore, Machine Learning (ML) is essential for these networks. Given stringent privacy regulations, future network architectures should use privacy-preserved ML for their applications and services. Federated Learning (FL) is expected to play an important role as a popular approach for distributed ML, as it protects privacy by design. However, many practical challenges exist before FL can be fully utilized as a key technology for these future networks. We consider the vision of a 6G layered architecture to evaluate the applicability of FL-based distributed intelligence. In this paper, we highlight the benefits of using FL for 6G and the main challenges and issues involved. We also discuss the existing solutions and the possible future directions that should be taken toward more robust and trustworthy FL …

This paper presents several machine learning methods and highlights the most effective one for detecting elephant rumbles in infrasonic seismic signals. The design and implementation of electronic circuitry to amplify, filter, and digitize the seismic signals captured through geophones are presented. The process converts seismic rumbles to a spectrogram and the existing methods of spectrogram feature extraction and appropriate machine learning algorithms are compared on their merit for automatic seismic rumble identification. A novel method of denoising the spectrum that leads to enhanced accuracy in identifying seismic rumbles is presented. It is experimentally found that the combination of the Mel-frequency cepstral coefficient (MFCC) feature extraction method and the ridge classifier machine learning algorithm give the highest accuracy of 97% in detecting infrasonic elephant rumbles hidden in seismic signals. The trained machine learning algorithm can run quite efficiently on general-purpose embedded hardware such as a Raspberry Pi, hence the method provides a cost-effective and scalable platform to develop a tool to remotely localize elephants, which would help mitigate the human–elephant conflict.

The neonatal period is the most vulnerable time for the development of seizures. Seizures in the immature brain lead to detrimental consequences, therefore require early diagnosis. The gold -standard for neonatal seizure detection currently relies on continuous video-EEG monitoring; which involves recording multi-channel electroencephalogram (EEG) alongside real-time video monitoring within a neonatal intensive care unit (NICU). However, video-EEG monitoring technology requires clinical expertise and is often limited to technologically advanced and resourceful settings. Cost-effective new techniques could help the medical fraternity make an accurate diagnosis and advocate treatment without delay. In this work, a novel explainable deep learning model to automate the neonatal seizure detection process with a reduced EEG montage is proposed, which employs convolutional nets, graph attention layers, and …

Cardiovascular diseases rank among the leading causes of mortality worldwide and the early identification of diseases is of paramount importance. This work focuses on developing a novel machine learning-based framework for early detection and classification of heart murmurs by analysing phonocardiogram signals. Our heart murmur detection and classification pipeline encompasses three classification settings. We first develop a set of methods based on transfer learning to determine the existence of heart murmurs and categorize them as present, absent, or unknown. If a murmur is present it will be classified as normal or abnormal based on its clinical outcome by using 1D convolution and audio spectrogram transformers. Finally, we use Wav2Vec encoder with raw audio data and AdaBoost abstain classifier for heart murmur quality identification. Heart murmurs are categorized based on their specific attributes …

With applications ranging from metabolomics to histopathology, quantitative phase microscopy (QPM) is a powerful label-free imaging modality. Despite significant advances in fast multiplexed imaging sensors and deep-learning-based inverse solvers, the throughput of QPM is currently limited by the pixel-rate of the image sensors. Complementarily, to improve throughput further, here we propose to acquire images in a compressed form so that more information can be transferred beyond the existing hardware bottleneck of the image sensor. To this end, we present a numerical simulation of a learnable optical compression-decompression framework that learns content-specific features. The proposed differentiable quantitative phase microscopy (∂-QPM) first uses learnable optical processors as image compressors. The intensity representations produced by these optical processors are then captured by the …

Neural networks can learn features from millions of histopathology images. However, curating high-quality annotations for training is laborious. In this work, we show how such models can be trained with only 1-10% of annotated data.

Frequency-modulated continuous-wave (FMCW) multiple-input multiple-output (MIMO) long-range radars currently employed for autonomous robotics have limited maximum range. By employing transmit beamforming and beam scanning, the range can be increased. However, the beam scanning time reduces the achievable velocity resolution. In this paper, we propose an FMCW MIMO radar, operating at 77 GHz, with transmit (TX) beamforming with subarrays to increase the range. We employ four TX subarrays, each having three antennas, with analog beamforming. Compared to TX beamforming as one TX array, our approach provides sufficiently wide beams alleviating beam scanning, further, without substantially reducing the virtual elements of the MIMO radar. We implement the radar signal processing pipeline on a ZYNQ Ultrascale+ ZCU106 FPGA to achieve real-time processing. Furthermore, we employ a …

Few-shot learning (FSL) aims to enable models to recognize novel objects or classes with limited labeled data. Feature generators, which synthesize new data points to augment limited datasets, have emerged as a promising solution to this challenge. This paper investigates the effectiveness of feature generators in enhancing the embedding process for FSL tasks. To address the issue of inaccurate embeddings due to the scarcity of images per class, we introduce a feature generator that creates visual features from class-level textual descriptions. By training the generator with a combination of classifier loss, discriminator loss, and distance loss between the generated features and true class embeddings, we ensure the generation of accurate same-class features and enhance the overall feature representation. Our results show a significant improvement in accuracy over baseline methods, with our approach …

With applications in quantitative metabolomics and label-free digital pathology, Quantitative Phase Microscopy (QPM) measures refractive index maps of thin transparent specimens like live cells or tissue sections. In QPM, refractive index maps are usually reconstructed from interference measurements of the object’s light field with respect to a reference field. To this end, many previous works focused on designing stable full-field interferometers from the bottom up. In this work, we present an alternative strategy to design a QPM system top-down, starting from the desired measurement outcomes, with no explicit knowledge about interferometry. We call our inverse design strategy Differentiable Microcopy. To this end, our Differentiable Microcopy approach designed a range of Fourier-filter-based QPM systems that do not require computational post-reconstruction. Our designs are superior to existing similar designs in …

Recently, various optical neural architectures have been designed to perform imaging and all-optical image processing. These designs consist of several optical masks with optimized transmission coefficients for the task. Designing sparse optical masks for them is important as it can promote different aspects of the design such as ease of fabrication and power efficiency of the design. To this end, inspired by the sparse filter designs, we propose training optical neural architectures with a regularization promoting sparsity in the masks. As preliminary results, we demonstrate a D2NN design for QPI achieving a sparsity of 33% with a performance degradation of only 12%.

Wi-Fi has become a cost-effective sensing mechanism due to its widespread availability. Channel State Information (CSI) from the OFDM-based physical layer of Wi-Fi contains rich information for sensing static and dynamic channel conditions, and hence for sensing activities within the channel. Numerous applications which combine CSI-based sensing with machine learning (ML) available for both indoor and outdoor environments. We extend the use of CSI- based Wi-Fi sensing to a unique setting: a confined environment. An industrial beverage cooler is specifically considered. We address the challenges of sensing device placement in such confined environments, where line-of-sight obstruction and near-field effects play a key role in CSI stability. We compensate for these effects by introducing novel combinations of data preprocessing techniques and input feature definitions. We demonstrate an accuracy of 89 …