Large Vision-Language models (VLMs) such as Contrastive Language-Image Pretraining (CLIP) have transformed open world image recognition. Nevertheless, few-shot classification, particularly in the extremely low-shot regime, requires not only high accuracy but also reliably calibrated uncertainty for decisions with high confidence. Existing training-free CLIP adapters are primarily designed to increase accuracy and efficiency; integrate the zero-shot text logits with the few-shot feature caches, but not definitely model predictive uncertainty and therefore often exhibit considerable miscalibration and weak selective performance. Bayesian adapters move in the direction of probabilistic modeling by placing priors over adapter parameters and employing task-specific variational training; however, this requires gradient-based optimization for every new task, increases computational costs, and becomes fragile when only one or two labeled examples per class are available. Starting from this observation, we introduce a training-free posterior-predictive Likelihood Ratio Adapter(LiRA-CLIP) for few-shot CLIP classification, which directly addresses probabilistic reliability under strict low-shot and deployment constraints. LiRA-CLIP extends the frozen CLIP head by a text-conditioned generative model in feature space that produces heavy-tailed posterior-predictive likelihood ratios, fused with the CLIP logits via a small, reliability-driven calibration layer. This layer is optimized in order to minimize the negative log-likelihood under an explicit accuracy side constraint, which leads to calibrated probabilities and dependable selective decisions without any gradient-based task-specific training. Extensive experiments show that LiRA-CLIP matches or slightly surpasses strong CLIP adapters in top-1 accuracy, while reducing calibration error by roughly 40%–50% and significantly increasing 95% and 99% reliable coverage in the low-shot regime, and thus establishes a new state of the art with respect to probabilistic reliability for training-free few-shot CLIP models.

Cardiovascular disease is the leading cause of death globally. This disease causes loss of heart muscles and is also responsible for the death of heart cells, sometimes damaging their functionality. A person’s life may depend on receiving timely assistance as soon as possible. Thus, minimizing the death ratio can be achieved by early detection of heart attack (HA) symptoms. In the United States alone, an estimated 610,000 people die from heart attacks each year, accounting for one in every four fatalities. However, by identifying and reporting heart attack symptoms early on, it is possible to reduce damage and save many lives significantly. Our objective is to devise an algorithm aimed at helping individuals, particularly elderly individuals living independently, to safeguard their lives. To address these challenges, we employ deep learning techniques. We have utilized a vision transformer (ViT) to address this problem. However, it has a significant overhead cost due to its memory consumption and computational complexity because of scaling dot-product attention. Also, since transformer performance typically relies on large-scale or adequate data, adapting ViT for smaller datasets is more challenging. In response, we propose a three-in-one steam model, the Multi-Head Attention Vision Hybrid (MHAVH). This model integrates a real-time posture recognition framework to identify chest pain postures indicative of heart attacks using transfer learning techniques, such as ResNet-50 and VGG-16, renowned for their robust feature extraction capabilities. By incorporating multiple heads into the vision transformer to generate additional metrics and enhance heart-detection capabilities, we leverage a 2019 posture-based dataset comprising RGB images, a novel creation by the author that marks the first dataset tailored for posture-based heart attack detection. Given the limited online data availability, we segmented this dataset into gender categories (male and female) and conducted testing on both segmented and original datasets. The training accuracy of our model reached an impressive 99.77%. Upon testing, the accuracy for male and female datasets was recorded at 92.87% and 75.47%, respectively. The combined dataset accuracy is 93.96%, showcasing a commendable performance overall. Our proposed approach demonstrates versatility in accommodating small and large datasets, offering promising prospects for real-world applications.