To implement the parallel acceleration of convolution operation of Convolutional Neural Networks (CNNs) on field programmable gate array (FPGA), large quantities of the logic resources will be consumed, expecially DSP cores. Many previous researches fail to make a well balance between DSP and LUT6. For better resource efficiency, a typical convolution structure is implemented with LUT6s in this paper. Besides, a novel convolution structure is proposed to further reduce the LUT6 resource consumption by modifying the typical convolution structure. The equations to evaluate the LUT6 resource consumptions of both structures are presented and validated. The theoretical evaluation and experimental results show that the novel structure can save 3.5-8% of LUT6s compared with the typical structure.

In this paper, we propose a hybrid fine-tuned multi-objective memetic algorithm hybridizing different solution fitness evaluation methods for global exploitation and exploration. To search across all regions in objective space, the algorithm uses a widely diversified set of weights at each generation, and employs a simulated annealing to optimize each utility function. For broader exploration, a grid-based technique is adopted to discover the missing nondominated regions on existing tradeoff surface, and a Pareto-based local perturbation is performed to reproduce incrementing solutions trying to fill up the discontinuous areas. Additional advanced feature is that the procedure is made dynamic and adaptive to the online optimization conditions based on a function of improvement ratio to obtain better stability and convergence of the algorithm. Effectiveness of our approach is shown by applying it to multi-objective 0/1 knapsack problem (MOKP).

This paper addresses the scheduling problem of a class of automated manufacturing systems with multiple resource requests. In the automated manufacturing system model, a set of jobs is to be processed and each job requires a sequence of operations. Each operation may need more than one resource type and multiple identical units with the same resource type. Upon the completion of an operation, resources needed in the next operation of the same job cannot be released and the remaining resources cannot be released until the start of the next operation. The scheduling problem is formulated by Timed Petri nets model under which the scheduling goal consists in sequencing the transition firing sequence in order to avoid the deadlock situation and to minimize the makespan. In the proposed genetic algorithm with deadlock-free constraint, Petri net transition sequence is coded and a deadlock detection method based on D-siphon technology is proposed to reschedule the sequence of transitions. The enabled transitions should be fired as early as possible and thus the quality of solutions can be improved. In the fitness computation procedure, a penalty item for the infeasible solution is involved to prevent the search process from converging to the infeasible solution. The method proposed in this paper can get a feasible scheduling strategy as well as enable the system to achieve good performance. Numerical results presented in the paper show the efficiency of the proposed algorithm.

Deep Graphical Model (DGM) based on Generative Adversarial Nets (GANs) has shown promise in image generation and latent variable inference. One of the typical models is the Iterative Adversarial Inference model (GibbsNet), which learns the joint distribution between the data and its latent variable. We present RGNet (Re-inference GibbsNet) which introduces a re-inference chain in GibbsNet to improve the quality of generated samples and inferred latent variables. RGNet consists of the generative, inference, and discriminative networks. An adversarial game is cast between the generative and inference networks and the discriminative network. The discriminative network is trained to distinguish between (i) the joint inference-latent/data-space pairs and re-inference-latent/data-space pairs and (ii) the joint sampled-latent/generated-data-space pairs. We show empirically that RGNet surpasses GibbsNet in the quality of inferred latent variables and achieves comparable performance on image generation and inpainting tasks.