Erdrin Azemi

Ran Liu, Ellen L. Zippi, Hadi Pouransari et al.

Leveraging multimodal information from biosignals is vital for building a comprehensive representation of people's physical and mental states. However, multimodal biosignals often exhibit substantial distributional shifts between pretraining and inference datasets, stemming from changes in task specification or variations in modality compositions. To achieve effective pretraining in the presence of potential distributional shifts, we propose a frequency-aware masked autoencoder ($\texttt{bio}$FAME) that learns to parameterize the representation of biosignals in the frequency space. $\texttt{bio}$FAME incorporates a frequency-aware transformer, which leverages a fixed-size Fourier-based operator for global token mixing, independent of the length and sampling rate of inputs. To maintain the frequency components within each input channel, we further employ a frequency-maintain pretraining strategy that performs masked autoencoding in the latent space. The resulting architecture effectively utilizes multimodal information during pretraining, and can be seamlessly adapted to diverse tasks and modalities at test time, regardless of input size and order. We evaluated our approach on a diverse set of transfer experiments on unimodal time series, achieving an average of $\uparrow$5.5% improvement in classification accuracy over the previous state-of-the-art. Furthermore, we demonstrated that our architecture is robust in modality mismatch scenarios, including unpredicted modality dropout or substitution, proving its practical utility in real-world applications. Code is available at https://github.com/apple/ml-famae .

Vikramjit Mitra, Vasudha Kowtha, Hsiang-Yun Sherry Chien et al.

Pre-trained model representations have demonstrated state-of-the-art performance in speech recognition, natural language processing, and other applications. Speech models, such as Bidirectional Encoder Representations from Transformers (BERT) and Hidden units BERT (HuBERT), have enabled generating lexical and acoustic representations to benefit speech recognition applications. We investigated the use of pre-trained model representations for estimating dimensional emotions, such as activation, valence, and dominance, from speech. We observed that while valence may rely heavily on lexical representations, activation and dominance rely mostly on acoustic information. In this work, we used multi-modal fusion representations from pre-trained models to generate state-of-the-art speech emotion estimation, and we showed a 100% and 30% relative improvement in concordance correlation coefficient (CCC) on valence estimation compared to standard acoustic and lexical baselines. Finally, we investigated the robustness of pre-trained model representations against noise and reverberation degradation and noticed that lexical and acoustic representations are impacted differently. We discovered that lexical representations are more robust to distortions compared to acoustic representations, and demonstrated that knowledge distillation from a multi-modal model helps to improve the noise-robustness of acoustic-based models.

Vikramjit Mitra, Hsiang-Yun Sherry Chien, Vasudha Kowtha et al.

Estimating dimensional emotions, such as activation, valence and dominance, from acoustic speech signals has been widely explored over the past few years. While accurate estimation of activation and dominance from speech seem to be possible, the same for valence remains challenging. Previous research has shown that the use of lexical information can improve valence estimation performance. Lexical information can be obtained from pre-trained acoustic models, where the learned representations can improve valence estimation from speech. We investigate the use of pre-trained model representations to improve valence estimation from acoustic speech signal. We also explore fusion of representations to improve emotion estimation across all three emotion dimensions: activation, valence and dominance. Additionally, we investigate if representations from pre-trained models can be distilled into models trained with low-level features, resulting in models with a less number of parameters. We show that fusion of pre-trained model embeddings result in a 79% relative improvement in concordance correlation coefficient CCC on valence estimation compared to standard acoustic feature baseline (mel-filterbank energies), while distillation from pre-trained model embeddings to lower-dimensional representations yielded a relative 12% improvement. Such performance gains were observed over two evaluation sets, indicating that our proposed architecture generalizes across those evaluation sets. We report new state-of-the-art "text-free" acoustic-only dimensional emotion estimation $CCC$ values on two MSP-Podcast evaluation sets.

Vikramjit Mitra, Anirban Chatterjee, Ke Zhai et al.

The process of human speech production involves coordinated respiratory action to elicit acoustic speech signals. Typically, speech is produced when air is forced from the lungs and is modulated by the vocal tract, where such actions are interspersed by moments of breathing in air (inhalation) to refill the lungs again. Respiratory rate (RR) is a vital metric that is used to assess the overall health, fitness, and general well-being of an individual. Existing approaches to measure RR (number of breaths one takes in a minute) are performed using specialized equipment or training. Studies have demonstrated that machine learning algorithms can be used to estimate RR using bio-sensor signals as input. Speech-based estimation of RR can offer an effective approach to measure the vital metric without requiring any specialized equipment or sensors. This work investigates a machine learning based approach to estimate RR from speech segments obtained from subjects speaking to a close-talking microphone device. Data were collected from N=26 individuals, where the groundtruth RR was obtained through commercial grade chest-belts and then manually corrected for any errors. A convolutional long-short term memory network (Conv-LSTM) is proposed to estimate respiration time-series data from the speech signal. We demonstrate that the use of pre-trained representations obtained from a foundation model, such as Wav2Vec2, can be used to estimate respiration-time-series with low root-mean-squared error and high correlation coefficient, when compared with the baseline. The model-driven time series can be used to estimate $RR$ with a low mean absolute error (MAE) ~ 1.6 breaths/min.

Jingping Nie, Ran Liu, Behrooz Mahasseni et al.

Acoustic signals are crucial for health monitoring, particularly heart sounds which provide essential data like heart rate and detect cardiac anomalies such as murmurs. This study utilizes a publicly available phonocardiogram (PCG) dataset to estimate heart rate using model-driven methods and extends the best-performing model to a multi-task learning (MTL) framework for simultaneous heart rate estimation and murmur detection. Heart rate estimates are derived using a sliding window technique on heart sound snippets, analyzed with a combination of acoustic features (Mel spectrogram, cepstral coefficients, power spectral density, root mean square energy). Our findings indicate that a 2D convolutional neural network (\textbf{\texttt{2dCNN}}) is most effective for heart rate estimation, achieving a mean absolute error (MAE) of 1.312 bpm. We systematically investigate the impact of different feature combinations and find that utilizing all four features yields the best results. The MTL model (\textbf{\texttt{2dCNN-MTL}}) achieves accuracy over 95% in murmur detection, surpassing existing models, while maintaining an MAE of 1.636 bpm in heart rate estimation, satisfying the requirements stated by Association for the Advancement of Medical Instrumentation (AAMI).

Christopher Sandino, Sayeri Lala, Geeling Chau et al.

Self-supervised learning (SSL) offers a promising approach for learning electroencephalography (EEG) representations from unlabeled data, reducing the need for expensive annotations for clinical applications like sleep staging and seizure detection. While current EEG SSL methods predominantly use masked reconstruction strategies like masked autoencoders (MAE) that capture local temporal patterns, position prediction pretraining remains underexplored despite its potential to learn long-range dependencies in neural signals. We introduce PAirwise Relative Shift or PARS pretraining, a novel pretext task that predicts relative temporal shifts between randomly sampled EEG window pairs. Unlike reconstruction-based methods that focus on local pattern recovery, PARS encourages encoders to capture relative temporal composition and long-range dependencies inherent in neural signals. Through comprehensive evaluation on various EEG decoding tasks, we demonstrate that PARS-pretrained transformers consistently outperform existing pretraining strategies in label-efficient and transfer learning settings, establishing a new paradigm for self-supervised EEG representation learning.

Vikramjit Mitra, Jingping Nie, Erdrin Azemi

Representations derived from models such as BERT (Bidirectional Encoder Representations from Transformers) and HuBERT (Hidden units BERT), have helped to achieve state-of-the-art performance in dimensional speech emotion recognition. Despite their large dimensionality, and even though these representations are not tailored for emotion recognition tasks, they are frequently used to train large speech emotion models with high memory and computational costs. In this work, we show that there exist lower-dimensional subspaces within the these pre-trained representational spaces that offer a reduction in downstream model complexity without sacrificing performance on emotion estimation. In addition, we model label uncertainty in the form of grader opinion variance, and demonstrate that such information can improve the models generalization capacity and robustness. Finally, we compare the robustness of the emotion models against acoustic degradations and observed that the reduced dimensional representations were able to retain the performance similar to the full-dimensional representations without significant regression in dimensional emotion performance.

Ching Fang, Christopher Sandino, Behrooz Mahasseni et al.

Many healthcare applications are inherently multimodal, involving several physiological signals. As sensors for these signals become more common, improving machine learning methods for multimodal healthcare data is crucial. Pretraining foundation models is a promising avenue for success. However, methods for developing foundation models in healthcare are still in early exploration and it is unclear which pretraining strategies are most effective given the diversity of physiological signals. This is partly due to challenges in multimodal health data: obtaining data across many patients is difficult and costly, there is a lot of inter-subject variability, and modalities are often heterogeneously informative across downstream tasks. Here, we explore these challenges in the PhysioNet 2018 dataset. We use a masked autoencoding objective to pretrain a multimodal model. We show that the model learns representations that can be linearly probed for a diverse set of downstream tasks. We hypothesize that cross-modal reconstruction objectives are important for successful multimodal training, as they encourage the model to integrate information across modalities. We demonstrate that modality dropout in the input space improves performance across downstream tasks. We also find that late-fusion models pretrained with contrastive learning objectives are less effective across multiple tasks. Finally, we analyze the model's representations, showing that attention weights become more cross-modal and temporally aligned with our pretraining strategy. The learned embeddings also become more distributed in terms of the modalities encoded by each unit. Overall, our work demonstrates the utility of multimodal foundation models with health data, even across diverse physiological data sources. We further argue that explicit methods for inducing cross-modality may enhance multimodal pretraining strategies.

Jingping Nie, Dung T. Tran, Karan Thakkar et al.

Auscultation, particularly heart sound, is a non-invasive technique that provides essential vital sign information. Recently, self-supervised acoustic representation foundation models (FMs) have been proposed to offer insights into acoustics-based vital signs. However, there has been little exploration of the extent to which auscultation is encoded in these pre-trained FM representations. In this work, using a publicly available phonocardiogram (PCG) dataset and a heart rate (HR) estimation model, we conduct a layer-wise investigation of six acoustic representation FMs: HuBERT, wav2vec2, wavLM, Whisper, Contrastive Language-Audio Pretraining (CLAP), and an in-house CLAP model. Additionally, we implement the baseline method from Nie et al., 2024 (which relies on acoustic features) and show that overall, representation vectors from pre-trained foundation models (FMs) offer comparable performance to the baseline. Notably, HR estimation using the representations from the audio encoder of the in-house CLAP model outperforms the results obtained from the baseline, achieving a lower mean absolute error (MAE) across various train/validation/test splits despite the domain mismatch.

Vikramjit Mitra, Amrit Romana, Dung T. Tran et al.

Spontaneous speech emotion data usually contain perceptual grades where graders assign emotion score after listening to the speech files. Such perceptual grades introduce uncertainty in labels due to grader opinion variation. Grader variation is addressed by using consensus grades as groundtruth, where the emotion with the highest vote is selected. Consensus grades fail to consider ambiguous instances where a speech sample may contain multiple emotions, as captured through grader opinion uncertainty. We demonstrate that using the probability density function of the emotion grades as targets instead of the commonly used consensus grades, provide better performance on benchmark evaluation sets compared to results reported in the literature. We show that a saliency driven foundation model (FM) representation selection helps to train a state-of-the-art speech emotion model for both dimensional and categorical emotion recognition. Comparing representations obtained from different FMs, we observed that focusing on overall test-set performance can be deceiving, as it fails to reveal the models generalization capacity across speakers and gender. We demonstrate that performance evaluation across multiple test-sets and performance analysis across gender and speakers are useful in assessing usefulness of emotion models. Finally, we demonstrate that label uncertainty and data-skew pose a challenge to model evaluation, where instead of using the best hypothesis, it is useful to consider the 2- or 3-best hypotheses.

Wenhui Cui, Christopher Sandino, Hadi Pouransari et al.

Hand gesture classification using high-quality structured data such as videos, images, and hand skeletons is a well-explored problem in computer vision. Leveraging low-power, cost-effective biosignals, e.g. surface electromyography (sEMG), allows for continuous gesture prediction on wearables. In this paper, we demonstrate that learning representations from weak-modality data that are aligned with those from structured, high-quality data can improve representation quality and enables zero-shot classification. Specifically, we propose a Contrastive Pose-EMG Pre-training (CPEP) framework to align EMG and pose representations, where we learn an EMG encoder that produces high-quality and pose-informative representations. We assess the gesture classification performance of our model through linear probing and zero-shot setups. Our model outperforms emg2pose benchmark models by up to 21% on in-distribution gesture classification and 72% on unseen (out-of-distribution) gesture classification.

Joseph Y. Cheng, Hanlin Goh, Kaan Dogrusoz et al.

Datasets for biosignals, such as electroencephalogram (EEG) and electrocardiogram (ECG), often have noisy labels and have limited number of subjects (<100). To handle these challenges, we propose a self-supervised approach based on contrastive learning to model biosignals with a reduced reliance on labeled data and with fewer subjects. In this regime of limited labels and subjects, intersubject variability negatively impacts model performance. Thus, we introduce subject-aware learning through (1) a subject-specific contrastive loss, and (2) an adversarial training to promote subject-invariance during the self-supervised learning. We also develop a number of time-series data augmentation techniques to be used with the contrastive loss for biosignals. Our method is evaluated on publicly available datasets of two different biosignals with different tasks: EEG decoding and ECG anomaly detection. The embeddings learned using self-supervision yield competitive classification results compared to entirely supervised methods. We show that subject-invariance improves representation quality for these tasks, and observe that subject-specific loss increases performance when fine-tuning with supervised labels.