Convolution neural network is successful in pervasive vision tasks, including label distribution learning, which usually takes the form of learning an injection from the nonlinear visual features to the well-defined labels. However, how the discrepancy between features is mapped to the label discrepancy is ambient, and its correctness is not guaranteed. To address these problems, we study the mathematical connection between feature and its label, presenting a general and simple framework for label distribution learning. We propose a so-called Triangular Distribution Transform (TDT) to build an injective function between feature and label, guaranteeing that any symmetric feature discrepancy linearly reflects the difference between labels. The proposed TDT can be used as a plug-in in mainstream backbone networks to address different label distribution learning tasks. Experiments on Facial Age Recognition, Illumination Chromaticity Estimation, and Aesthetics assessment show that TDT achieves on-par or better results than the prior arts. Code is available at https://github.com/redcping/TDT.

Efficient detectors for edge devices are often optimized for parameters or speed count metrics, which remain in weak correlation with the energy of detectors. However, some vision applications of convolutional neural networks, such as always-on surveillance cameras, are critical for energy constraints. This paper aims to serve as a baseline by designing detectors to reach tradeoffs between energy and performance from two perspectives: 1) We extensively analyze various CNNs to identify low-energy architectures, including selecting activation functions, convolutions operators, and feature fusion structures on necks. These underappreciated details in past work seriously affect the energy consumption of detectors; 2) To break through the dilemmatic energy-performance problem, we propose a balanced detector driven by energy using discovered low-energy components named FemtoDet. In addition to the novel construction, we improve FemtoDet by considering convolutions and training strategy optimizations. Specifically, we develop a new instance boundary enhancement (IBE) module for convolution optimization to overcome the contradiction between the limited capacity of CNNs and detection tasks in diverse spatial representations, and propose a recursive warm-restart (RecWR) for optimizing training strategy to escape the sub-optimization of light-weight detectors by considering the data shift produced in popular augmentations. As a result, FemtoDet with only 68.77k parameters achieves a competitive score of 46.3 AP50 on PASCAL VOC and 1.11 W & 64.47 FPS on Qualcomm Snapdragon 865 CPU platforms. Extensive experiments on COCO and TJUDHD datasets indicate that the proposed method achieves competitive results in diverse scenes.

Semi-supervised learning is a challenging problem which aims to construct a model by learning from limited labeled examples. Numerous methods for this task focus on utilizing the predictions of unlabeled instances consistency alone to regularize networks. However, treating labeled and unlabeled data separately often leads to the discarding of mass prior knowledge learned from the labeled examples. In this paper, we propose a novel method for semi-supervised semantic segmentation named GuidedMix-Net, by leveraging labeled information to guide the learning of unlabeled instances. Specifically, GuidedMix-Net employs three operations: 1) interpolation of similar labeled-unlabeled image pairs; 2) transfer of mutual information; 3) generalization of pseudo masks. It enables segmentation models can learning the higher-quality pseudo masks of unlabeled data by transfer the knowledge from labeled samples to unlabeled data. Along with supervised learning for labeled data, the prediction of unlabeled data is jointly learned with the generated pseudo masks from the mixed data. Extensive experiments on PASCAL VOC 2012, and Cityscapes demonstrate the effectiveness of our GuidedMix-Net, which achieves competitive segmentation accuracy and significantly improves the mIoU over 7$\%$ compared to previous approaches.

The SSAPRE algorithm for performing partial redundancy elimination based entirely on SSA form is presented. The algorithm is formulated based on a new conceptual framework, the factored redundancy graph, for analyzing redundancy, and representes the first sparse approach to the classical problem and on methods for its solution. With the algorithm description, theorems and their proofs are given showing that the algorithm produces the best possible code by the criteria of computational optimality and lifetime optimality of the introduced temporaries. In addition to the base algorithm, a practical implementation of SSAPRE that exhibits additional compile-time efficiencies is described. In closing, measurement statistics are provided that characterize the instances of the partial redundancy problem from a set of benchmark programs and compare optimization time spent by an implementation of SSAPRE aganist a classical partial redundancy elimination implementation. The data lend insight into the nature of partial redundancy elimination and demonstrate the expediency of this new approach.

An algorithm for register promotion is presented based on the observation that the circumstances for promoting a memory location’s value to register coincide with situations where the program exhibits partial redundancy between ac-cesses to the memory location. The recent SSAPRE al-gorithm for eliminating partial redundancy using a sparse SSA representation forms the foundation for the present al-gorithm to eliminate redundancy among memory accesses, enabling us to achieve both computational and live range op-timality in our register promotion results. We discuss how to effect speculative code motion in the SSAPRE framework. We present two different algorithms for performing specu-lative code motion: the conservative speculation algorithm used in the absence of profile data, and the the profile-driven

A new algorithm, SSAPRE, for performing partial redundancy elimination based entirely on SSA form is presented. It achieves optimal code motion similar to lazy code motion [KRS94a, DS93], but is formulated independently and does not involve iterative data flow analysis and bit vectors in its solution. It not only exhibits the characteristics common to other sparse approaches, but also inherits the advantages shared by other SSA-based optimization techniques. SSAPRE also maintains its output in the same SSA form as its input. In describing the algorithm, we state theorems with proofs giving our claims about SSAPRE. We also give additional description about our practical implementation of SSAPRE, and analyze and compare its performance with a bit-vector-based implementation of PRE. We conclude with some discussion of the implications of this work.

In this paper, we present a GSA-based technique that performs more efficient and more precise symbolic analysis of predicated assignments, recurrences and index arrays.The efficiency is improved by using a backward substitution scheme that performs resolution of assertions on-demand and uses heuristics to limit the number of substitution.The precision is increased by utilizing the gating predicate information embedded in the GSA and the control dependence information in the program flow graph.Examples from array privatization are used to illustrate how the technique aids loop parallelization. 1

this paper, we present an almost linear time algorithm to construct the GSA. The new algorithm is more efficient and simpler than the existing algorithms for GSA construction [BMO90, Hav93]. Since SSA is a special case of GSA, it can also be used as an efficient alternative algorithm for SSA construction. The existing algorithms for building the GSA follow two steps. The first step is the same

The scheduling problems for real-time systems with multiversion computations are studied. A computation is performance polymorphic if it has been implemented in several versions each with a different performance characteristics like the time needed to produce a result. Given a set of periodic or aperiodic jobs, each with multiple versions, an investigation is made of the scheduling problem which determines the execution time for each job, and a version is selected to optimize the overall system performance objective. The problems are modeled as resource sharing problems. Known techniques for the sharing problem can be used to allocate the time to each job. Several heuristic algorithms are studied for problems which are NP-complete, and their performances are compared.>