In recent years, federated learning has attracted more and more attention as it could collaboratively train a global model without gathering the users' raw data. It has brought many challenges. In this paper, we proposed layer-based federated learning system with privacy preservation. We successfully reduced the communication cost by selecting several layers of the model to upload for global averaging and enhanced the privacy protection by applying local differential privacy. We evaluated our system in non independently and identically distributed scenario on three datasets. Compared with existing works, our solution achieved better performance in both model accuracy and training time.

The performance of the kernel-based learning algorithms, such as SVM, depends heavily on the proper choice of the kernel parameter. It is desirable for the kernel machines to work on the optimal kernel parameter that adapts well to the input data and the learning tasks. In this paper, we present a novel method for selecting Gaussian kernel parameter by maximizing a class separability criterion, which measures the data distribution in the kernel-induced feature space, and is invariant under any non-singular linear transformation. The experimental results show that both the class separability of the data in the kernel-induced feature space and the classification performance of the SVM classifier are improved by using the optimal kernel parameter.

A parametric linear filter for a linear observation model usually requires a parameter selection process so that the filter achieves a better filtering performance. Generally, criteria for the parameter selection need not only the filtered solution but also the filter itself with each candidate of the parameter. Obtaining the filter usually costs a large amount of calculations. Thus, an efficient algorithm for the parameter selection is required. In this paper, we propose a fast parameter selection algorithm for linear parametric filters that utilizes a joint diagonalization of two non-negative definite Hermitian matrices.

We previously proposed a robust hybrid edge detector which relaxes the trade off between robustess against noise and accurate localization of the edges. This hybrid detector separates the tasks of localization and noise suppresion between two sub-detectors. In this paper, we present an extension to this hybrid detector to determine its optimal parameters, independently of the scene. This extension defines a probabilistic cost function using for criteria the probability of missing an edge buried in noise and the probability of detecting false edges. The optimization of this cost function allows the automatic selection of the parameters of the hybrid edge detector given the height of the minimum edge to be detected and the variance of the noise, σ2n. The results were applied to the 2D case and the performance of the adaptive hybrid detector was compared to other detectors.

Within the framework of a previously proposed vision system, a new part-segmentation algorithm, that breaks an object defined by its contour into its constituent parts, is presented. The contour is assumed to be obtained using an edge detector. This decomposition is achieved in two stages. The first stage is a preprocessing step which consists of extracting the convex dominant points (CDPs) of the contour. For this aim, we present a new technique which relaxes the compromise that exists in most classical methods for the selection of the width of the Gaussian filter. In the subsequent stage, the extracted CDPs are used to break the object into convex parts. This is performed as follows: among all the points of the contour only the CDPs are moved along their normals nutil they touch another moving CDP or a point on the contour. The results show that this part-segmentation algorithm is invariant to transformations such as rotation, scaling and shift in position of the object, which is very important for object recognition. The algorithm has been tested on many object contours, with and without noise and the advantages of the algorithm are listed in this paper. Our results are visually similar to a human intuitive decomposition of objects into their parts.