Existing explainability approaches for convolutional neural networks (CNNs) are mainly applied after training (post-hoc) which is generally unreliable. Ante-hoc explainers trained simultaneously with the CNN are more ...
Existing explainability approaches for convolutional neural networks (CNNs) are mainly applied after training (post-hoc) which is generally unreliable. Ante-hoc explainers trained simultaneously with the CNN are more reliable. However, current ante-hoc explanation methods mainly generate inexplicit concept-based explanations tailored to specific tasks. To address these limitations, we propose a task-agnostic ante-hoc framework that can generate interpretation maps to visually explain any CNN. Our framework simultaneously trains the CNN and a weighting network - an explanation generation module. The generated maps are self-explanatory, eliminating the need for manual identification of concepts. We demonstrate that our method can interpret classification, facial landmark detection, and image captioning tasks. We show that our framework is explicit, faithful, and stable through experiments. To the best of our knowledge, this is the first ante-hoc CNN explanation strategy that produces visual explanations generic to CNNs for different tasks.
Differential Ant Stigmergy Algorithm (DASA) is a recent meta-heuristic method which represents an adaptation of Ant Colony Optimization (ACO) to continuous optimization problems. Other adaptations of ACO to continuous...
详细信息
Differential Ant Stigmergy Algorithm (DASA) is a recent meta-heuristic method which represents an adaptation of Ant Colony Optimization (ACO) to continuous optimization problems. Other adaptations of ACO to continuous optimization exist but DASA seems to be very efficient for the class of highdimension real-parameters optimization problems, becoming a competitor to more classical methods such as Particle Swarm Optimization (PSO). The PSO is a meta-heuristic which is also inspired from insects' life as ACO. Even both methods use a population of entities, the memory of PSO is larger as those of DASA. DASA uses just one current solution around which the population search and the difference that generated last improvement. PSO stores the previous best position and velocity of every particle. This paper attempts to examine if an elitist version of DASA has at least the same effectiveness (finding the true global solution) as PSO, but with significantly better efficiency (less function evaluation). The performance comparison of DASA and PSO is implemented using a set of six test functions well known for their difficulty.
The, so-called, adaptive super-twist controller is considered here. It contains the adaptive gain-parameter which adjusts the level of a scalar control action on-line based on direct measurements the "equivalent ...
详细信息
ISBN:
(纸本)9781612848006
The, so-called, adaptive super-twist controller is considered here. It contains the adaptive gain-parameter which adjusts the level of a scalar control action on-line based on direct measurements the "equivalent control" obtained by a first-order low-pass filter. It is shown that the problems of the finite time convergence and keeping a small chattering amplitude can be handled simultaneously if the gain magnitude is reduced to a minimal admissible level defined by the conditions for sliding mode to exist. The suggested methodology is compared numerically with another schemes of a gain-adaptation (including sigma-adaptation) showing a high effectiveness under a significant level of uncertainties and external disturbances.
Repetitive processes are a distinct class of 2D linear systems with applications in areas ranging from long-wall coal cutting and metal rolling operations through to iterative learning control schemes. The main featur...
详细信息
Repetitive processes are a distinct class of 2D linear systems with applications in areas ranging from long-wall coal cutting and metal rolling operations through to iterative learning control schemes. The main featur...
详细信息
Repetitive processes are a distinct class of 2D linear systems with applications in areas ranging from long-wall coal cutting and metal rolling operations through to iterative learning control schemes. The main feature which makes them distinct from other classes of 2D linear systems is that information propagation in one of the two independent directions only occurs over a finite duration. This, in turn, means that a distinct systems theory must be developed for them, which can then be translated into efficient routinely applicable controller design algorithms for applications domains. In this paper, we give the first significant results on a positive realness based approach to the analysis of these processes.
暂无评论