Federated Learning (FL) is a paradigm that enables collaborative training while keeping data localized, avoiding direct sharing with a central server. However, Non-IID (Non-Independent and Identically Distributed) data across clients presents a significant challenge for FL, compromising the global model performance. In this paper, we propose Entropy-based Loss Function Selection (ELFS), designed to enhance global model accuracy by selectively adapting the loss function based on the entropy of each client’s data distribution. ELFS leverages two core steps before local training to determine the appropriate loss function: 1) entropy calculation and 2) adaptive loss function selection. The entropy calculation step quantifies the data distribution of each client based on label frequencies. Subsequently, in the adaptive loss function selection step, each client selects an appropriate loss function to mitigate the impact of data imbalance. Our experimental results on Non-IID datasets, CIFAR-10 and CIFAR-100, demonstrate that ELFS improves global model accuracy by up to 16.13% compared to conventional FL methods, such as FedAvg, FedProx, and FedPer. By optimizing entropy thresholds, we further demonstrate the importance of fine-tuning hyperparameters to maximize accuracy. Moreover, ELFS offers flexibility for integrating additional loss functions, providing potential for further performance improvements in handling Non-IID data.

As SoC (System-on-Chip) complexity continues to increase, function/performance verification is required in the middle of design process (before tape-out) to reduce the possible risks ranging from over-design to non-compliance with the design specifications. In this paper, we propose a seamless SoC verification. The proposed methodology exploits a modern virtual platform (VP) technology which can combine high-level C++ firmware, timing-accurate SystemC models, and RTL (register-transfer level) designs. Thus, the full-chip level verification can be done at any design stages in the whole development process. With experimental results, this paper shows the benefits and lessons of using VPs.

This paper presents an new framework for homography-based analysis of pedestrian-vehicle activity in crowded scenes. Planar homography constraint is exploited to extract view-invariant object features including footage area and velocity of objects on the ground plane. Spatio-temporal relationships between people- and vehicle- tracks are represented by a semantic event. Context awareness of the situation is achieved by the estimated density distribution of objects and the anticipation of possible directions of near-future tracks using piecewise velocity history. Single-view and multi-view based homography mapping options are compared. Our framework can be used to enhance situational awareness for disaster prevention, human interactions in structured environments, and crowd movement analysis at wide regions

This paper presents a track-based system for human movement analysis and privacy protection. Our system is adaptive to environmental contexts such as illumination variations, complex moving cast shadows, different camera perspectives, and diverse sue scenarios. Most of outdoor surveillance systems have been targeting at specific environmental situation: i.e., specific time, place, and activity scenarios. We address that more general human movement analysis systems should be able to handle multiple heterogeneous situations in an adaptive manner. We introduce the concept of "spatio-temporal personal boundary" to represent different grouping patterns of human tracks, and we incorporate the concept with various site models. Experimental evaluations with extensive outdoor data show our system's robustness to environmental changes and effectiveness to properly handle various environmental contexts.

This paper presents a recognition system that classifies four kinds of human interactions: shaking hands, pointing at the opposite person, standing hand-in-hand, and an intermediate/transitional state between them. Our system achieves recognition by applying the K-nearest neighbor classifier to the parametric human-interaction model, which describes the interpersonal configuration with multiple features from gray scale images (i.e., binary blob, silhouette contour, and intensity distribution). Unlike the algorithms that use temporal information about motion, our system independently classifies each frame by estimating the relative poses of the interacting persons. The system provides a tool to detect the initiation and the termination of an interaction with no parsing procedure for sequential data. Experimental results are presented and illustrated.