State-of-the-art CNN based recognition models are often computationally prohibitive to deploy on low-end devices. A promising high level approach tackling this limitation is knowledge distillation, which let small student model mimic cumbersome teacher model's output to get improved generalization. However, related methods mainly focus on simple task of classification while do not consider complex tasks like object detection. We show applying the vanilla knowledge distillation to detection model gets minor gain. To address the challenge of distilling knowledge in detection model, we propose a fine-grained feature imitation method exploiting the cross-location discrepancy of feature response. Our intuition is that detectors care more about local near object regions. Thus the discrepancy of feature response on the near object anchor locations reveals important information of how teacher model tends to generalize. We design a novel mechanism to estimate those locations and let student model imitate the teacher on them to get enhanced performance. We first validate the idea on a developed lightweight toy detector which carries simplest notion of current state-of-the-art anchor based detection models on challenging KITTI dataset, our method generates up to 15% boost of mAP for the student model compared to the non-imitated counterpart. We then extensively evaluate the method with Faster R-CNN model under various scenarios with common object detection benchmark of Pascal VOC and COCO, imitation alleviates up to 74% performance drop of student model compared to teacher. Codes released at https://github.com/twangnh/ Distilling-Object-Detectors

Spiking neural P systems (SN P systems) have been well established as a novel class of distributed parallel computing models. Some features that SN P systems possess are attractive to fault diagnosis. However, handling fuzzy diagnosis knowledge and reasoning is required for many fault diagnosis applications. The lack of ability is a major problem of existing SN P systems when applying them to the fault diagnosis domain. Thus, we extend SN P systems by introducing some new ingredients (such as three types of neurons, fuzzy logic and new firing mechanism) and propose the fuzzy reasoning spiking neural P systems (FRSN P systems). The FRSN P systems are particularly suitable to model fuzzy production rules in a fuzzy diagnosis knowledge base and their reasoning process. Moreover, a parallel fuzzy reasoning algorithm based on FRSN P systems is developed according to neuron's dynamic firing mechanism. Besides, a practical example of transformer fault diagnosis is used to demonstrate the feasibility and effectiveness of the proposed FRSN P systems in fault diagnosis problem.

Spiking neural P systems (SN P systems) are a new class of computing models inspired by the neurophysiological be-havior of biological spiking neurons. In order to make SN P sys-tems capable of representing and processing fuzzy and uncertain knowledge, we propose a new class of spiking neural P systems in this paper called weighted fuzzy spiking neural P systems (WFSN P systems). New elements, including fuzzy truth value, certain factor, weighted fuzzy logic, output weight, threshold, new firing rule, and two types of neurons, are added to the original definition of SN P systems. This allows WFSN P systems to adequately characterize the features of weighted fuzzy production rules in a fuzzy rule-based system. Furthermore, a weighted fuzzy backward reasoning algorithm, based on WFSN P systems, is developed, which can ac-complish dynamic fuzzy reasoning of a rule-based system more flexibly and intelligently. In addition, we compare the proposed WFSN P systems with other knowledge representation methods, such as fuzzy production rule, conceptual graph, and Petri nets, to demonstrate the features and advantages of the proposed tech-niques.

No abstract