Effectively steering hearable devices requires understanding the acoustic environment around the user. In the computational analysis of sound scenes, foundation models have emerged as the state of the art to produce high-performance, robust, multi-purpose audio representations. We introduce and release Deep Evaluation of Audio Representations (DEAR), the first dataset and benchmark to evaluate the efficacy of foundation models in capturing essential acoustic properties for hearables. The dataset includes 1,158 audio tracks, each 30 seconds long, created by spatially mixing proprietary monologues with commercial, high-quality recordings of everyday acoustic scenes. Our benchmark encompasses eight tasks that assess the general context, speech sources, and technical acoustic properties of the audio scenes. Through our evaluation of four general-purpose audio representation models, we demonstrate that the BEATs model significantly surpasses its counterparts. This superiority underscores the advantage of models trained on diverse audio collections, confirming their applicability to a wide array of auditory tasks, including encoding the environment properties necessary for hearable steering. The DEAR dataset and associated code are available at https://dear-dataset.github.io.

This work introduces T-Loss, a novel and robust loss function for medical image segmentation. T-Loss is derived from the negative log-likelihood of the Student-t distribution and excels at handling noisy masks by dynamically controlling its sensitivity through a single parameter. This parameter is optimized during the backpropagation process, obviating the need for additional computations or prior knowledge about the extent and distribution of noisy labels. We provide in-depth analysis of this parameter behavior during training and revealing its adaptive nature and its role in preventing noisy memorization. Our extensive experiments demonstrate that T-Loss significantly outperforms traditional loss functions in terms of dice scores on two public medical datasets, specifically for skin lesion and lung segmentation. Moreover, T-Loss exhibits remarkable resilience to various types of simulated label noise, which mimics human annotation errors. Our results provide strong evidence that T-Loss is a promising alternative for medical image segmentation where high levels of noise or outliers in the dataset are a typical phenomenon in practice. The project website, including code and additional resources, can be found at: https://robust-tloss.github.io/.

Benchmark datasets in computer vision often contain off-topic images, near duplicates, and label errors, leading to inaccurate estimates of model performance.In this paper, we revisit the task of data cleaning and formalize it as either a ranking problem, which significantly reduces human inspection effort, or a scoring problem, which allows for automated decisions based on score distributions.We find that a specific combination of context-aware self-supervised representation learning and distance-based indicators is effective in finding issues without annotation biases.This methodology, which we call SelfClean, surpasses state-of-the-art performance in detecting off-topic images, near duplicates, and label errors within widely-used image datasets, such as ImageNet-1k, Food-101N, and STL-10, both for synthetic issues and real contamination.We apply the detailed method to multiple image benchmarks, identify up to 16% of issues, and confirm an improvement in evaluation reliability upon cleaning.The official implementation can be found at: https://github.com/Digital-Dermatology/SelfClean.

Insufficient data and data lacking the diversity to represent the general public is a common challenge when modelling diagnosis prediction. We consider a much larger and more diverse database of commercial Electronic Health Records than what is prevalent in the literature. We formulate a simplified version of diagnosis prediction that focuses on major developments in medical histories of patients. To this end, we leverage Auto-Regressive Self-Attention models that have seen promising applications in language modelling and extend them to incorporate ontological representations of medical codes. Additionally, we include time-intervals between diagnoses into the attention calculation. We evaluate models and baselines at different levels of diagnostic granularity and our results suggest that using very detailed clinical classifications does not significantly degrade performance, possibly allowing their use in practice. Our model outperforms all baselines and we suggest that leveraging the ontology for generating diagnosis representations is mostly helpful for rare diagnoses.