Abstract

Abstract
Riding a motorcycle involves risks that can be minimized through advanced sensing and response systems to assist the rider. The use of camera-collected images to monitor road conditions can aid in the development of tools designed to enhance rider safety and prevent accidents. This paper proposes a method for developing deep learning models designed to operate efficiently on embedded systems like the Raspberry Pi, facilitating real-time decisions that consider the road condition. Our research tests and compares several state-of-the-art convolutional neural network architectures, including EfficientNet and Inception, to determine which offers the best balance between inference time and accuracy. Specifically, we measured top-1 accuracy and inference time on a Raspberry Pi, identifying EfficientNetV2 as the most suitable model due to its optimal trade-off between performance and computational demand. The model's top-1 accuracy significantly outperformed other models while maintaining competitive inference speeds, making it ideal for real-time applications in traffic-dense urban settings.

Abstract

Abstract
Text readability is vital for effective communication and learning, especially for those with lower information literacy. This research aims to assess Llama 3's ability to grade readability and compare its alignment with established metrics. For that purpose, we create a new dataset of article lead sections from English and Simple English Wikipedia, covering nine categories. The model is prompted to rate the readability of the texts on a grade-level scale, and an in-depth analysis of the results is conducted. While Llama 3 correlates strongly with most metrics, it may underestimate text grade levels.

Abstract
The main challenges limiting the adoption of deep learning-based solutions in medical workflows are the availability of annotated data and the lack of interpretability of such systems. Concept Bottleneck Models (CBMs) tackle the latter by constraining the model output on a set of predefined and human-interpretable concepts. However, the increased interpretability achieved through these concept-based explanations implies a higher annotation burden. Moreover, if a new concept needs to be added, the whole system needs to be retrained. Inspired by the remarkable performance shown by Large Vision-Language Models (LVLMs) in few-shot settings, we propose a simple, yet effective, methodology, CBVLM, which tackles both of the aforementioned challenges. First, for each concept, we prompt the LVLM to answer if the concept is present in the input image. Then, we ask the LVLM to classify the image based on the previous concept predictions. Moreover, in both stages, we incorporate a retrieval module responsible for selecting the best examples for in-context learning. By grounding the final diagnosis on the predicted concepts, we ensure explainability, and by leveraging the few-shot capabilities of LVLMs, we drastically lower the annotation cost. We validate our approach with extensive experiments across four medical datasets and twelve LVLMs (both generic and medical) and show that CBVLM consistently outperforms CBMs and task-specific supervised methods without requiring any training and using just a few annotated examples. More information on our project page: https://cristianopatricio.github.io/CBVLM/.

Abstract