Deploying language models often requires navigating accuracy vs. performance trade-offs to meet latency constraints while preserving utility. Traditional model distillation reduces size but incurs substantial costs through training separate models. We introduce ModularStarEncoder (MoSE), a 1-billion-parameter multi-exit encoder for code retrieval and classification that employs a novel Self-Distillation mechanism. This approach significantly enhances lower-layer representations, enabling flexible deployment of different model portions with favorable performance trade-offs. Our architecture improves text-to-code and code-to-code search by targeting specific encoder layers as exit heads, where higher layers guide earlier ones during training, thereby improving intermediate representations at minimal additional cost. We further enhance MoSE with a repository-level contextual loss that maximizes training context window utilization. Additionally, we release a new dataset created through code translation that extends text-to code benchmarks with cross-language code-to-code pairs. Evaluations demonstrate the effectiveness of Self-Distillation as a principled approach to trading inference cost for accuracy across various code understanding tasks.

This paper introduces Trajectory-Embedded Matryoshka Representation Learning (TE-MRL).This novel framework synergies the capabilities of trajectory representation learning with the adaptability and efficiency of Matryoshka Representation Learning (MRL).TE-MRL is engineered to generate adaptive, multi-granular embeddings that efficiently capture the spatial-temporal dynamics inherent in trajectory data.We evaluate TE-MRL on the Porto dataset, focusing on trajectory similarity and k-nearest trajectory similarity tasks.Our findings demonstrate that TE-MRL preserves critical features such as travel semantics and temporal regularities while it can significantly reduce computational time and memory footprint.The proposed approach matches existing methods' accuracy and efficiency but demonstrates robust adaptability under varying computational constraints.Furthermore, we proposed a two-stage retrieval pipeline to enhance computational time while maintaining the same precision.We reduced the computation time by 8× while maintaining state-of-the-art precision.The effectiveness of TE-MRL in handling the complexity of the Porto dataset underlines its potential for broader applications in urban computing and mobility analytics.

Capturing the intricate dynamics of complex systems often requires integrating time-series data from multiple sources. However, effectively unifying and analyzing these heterogeneous streams poses significant methodological challenges. This work proposes a contrastive dual-encoder framework that learns a shared latent representation across distinct timeseries modalities. By mapping semantically related segments of different data views to proximate regions in the latent space, our approach captures cross-modal dependencies in a manner that traditional single-view models often overlook. We illustrate the framework in a motorcycle riding application, where rider behavior signals (e.g., gear, throttle, braking) are embedded jointly with vehicle dynamics (e.g., speed, lean angle, acceleration). We used real-world data to demonstrate the framework’s utility in uncovering how rider inputs translate into specific vehicle states and vice versa. Beyond validating cross-modal alignment, the learned embeddings facilitate robust downstream analyses, including feature importance via SHAP, rider profiling based on latent-space clustering, and mission profile characterization that links riding objectives to observed patterns. These findings underscore the effectiveness of dual-encoder architectures in enhancing the analysis and interpretation of complex multimodal time-series data.

In the automotive environment, understanding the thermal behavior of the engine is crucial: a fault in measuring the temperature can reduce reliability and poor fuel efficiency. Internal combustion engine temperature is usually measured using a device called coolant temperature sensor. The sensor may fail, and having an indirect system to get that value can avoid unpleasant conditions. This work addresses a novel solution to estimate the coolant temperature based on machine learning models trained using riding data. This solution can back up the physical sensor and monitor abnormal behavior. We also focused on developing a virtual sensor that would behave well in case of thermal shock, which is essential for ensuring safety, improving the engine work, and making the components last longer. We showed how an approach based on the prediction of the temperature delta is better for dealing with this kind of phenomenon. We obtained an RMSE of 2.79°C training a Long short-term memory (LSTM) on more than 1060$h$recorded from Ducati Multistrada V4.

This paper presents a novel approach to body gesture recognition for powered wheelchair users, leveraging inertial data from wrist-mounted sensors to facilitate movement and enhance autonomy in Adapted Physical Activity (APA). Gesture recognition technology interprets human gestures to allow non-direct communication with devices, enhancing human-machine interaction across various fields. APA fosters inclusion and well-being through tailored physical engagement. Our model not only identifies known gestures with high accuracy, as indicated by a mean Average Precision (mAP) score of 0.92 and a Recall@1 score of 0.983, but also demonstrates the ability to recognize gestures not included in the training set. This research contributes to the field of human-robot interaction by offering a more dynamic and inclusive form of interaction for individuals reliant on powered mobility aids.