Neural networks have been demonstrated to be trainable even with hundreds of layers, which exhibit remarkable improvement on expressive power and provide significant performance gains in a variety of tasks. However, the prohibitive computational cost has become a severe challenge for deploying them on resource-constrained platforms. Meanwhile, widely adopted deep neural network architectures, for example, ResNets or DenseNets, are manually crafted on benchmark datasets, which hamper their generalization ability to other domains. To cope with these issues, we propose an evolutionary algorithm-based method for shallowing deep neural networks (DNNs) at block levels, which is termed as ESNB. Different from existing studies, ESNB utilizes the ensemble view of block-wise DNNs and employs the multiobjective optimization paradigm to reduce the number of blocks while avoiding performance degradation. It automatically discovers shallower network architectures by pruning less informative blocks, and employs knowledge distillation to recover the performance. Moreover, a novel prior knowledge incorporation strategy is proposed to improve the exploration ability of the evolutionary search process, and a correctness-aware knowledge distillation strategy is designed for better knowledge transferring. Experimental results show that the proposed method can effectively accelerate the inference of DNNs while achieving superior performance when compared with the state-of-the-art competing methods.

Automatic esophageal lesion identification (ESEI) is of great importance to clinically aid the endoscopists with the early detection of esophageal cancer. However, accurate identification of esophageal lesion is challenging due to the varying shape, size, illumination condition, and complex background with artifacts in endoscopic images. Although deep neural network based approaches have considerably boosted the performance by automatically learning features from esophageal images, the configuration of the network architecture is highly dependent on domain expertise and is a daunting task to be manually tuned. In this article, we propose an evolutionary algorithm based approach to search for the optimal multitask network architecture for ESEI. Different from existing studies, we first design a multitask network search space, which considers the lesion identification as two steps including esophageal image classification and esophageal lesion segmentation. In particular, the input image resolution is covered in the search space, and the classification utilizes both downsampled and upsampled features. Besides, to avoid scratch training of sampled network architectures in the evolutionary algorithm, the one-shot supernet strategy is developed for searching the optimal network architecture. Results from the performed experiments on a collected sizeable clinical esophageal image dataset show that the proposed method improves on the state of the art on all measured metrics.

Generative Adversarial Networks (GANs) have achieved remarkable success in image translation tasks. However, its prohibitive computational overhead has been a major hurdle for deployment on resource constrained platforms. Due to the training instability and the complicated network architecture, existing model acceleration techniques cannot appropriately handle the GAN compression problem. To cope with these difficulties, we propose a multi-objective evolutionary algorithm to compress GAN models in the context of image translation tasks, which is termed as MEGC. Particularly, the conflict between the computational cost of the GAN model and the quality of generated image is explicitly modeled as a two-objective optimization problem, and the evolved Pareto set is utilized to guide the sampling process during supernet training, which can in turn divert the focus of the supernet training to well-performing compact subnets. Besides, an evaluation-free strategy is introduced to facilitate exploration in the search space while incurring no extra computational cost. Based on the above design, the proposed MEGC eliminates the requirement of subnet searching in the post-processing procedure. Experiments on image translation tasks under paired and unparied settings demonstrate the effectiveness of the proposed MEGC on reducing the computational cost of GANs while improving the quality of generated images compared to those of the full models.

Time series anomaly detection is crucial in various industrial applications to identify unusual behaviors within the time series data. Due to the challenges associated with annotating anomaly events, time series reconstruction has become a prevalent approach for unsupervised anomaly detection. However, effectively learning representations and achieving accurate detection results remain challenging due to the intricate temporal patterns and dependencies in real-world time series. In this paper, we propose a cross-dimension attentive feature fusion network for time series anomaly detection, referred to as CAFFN. Specifically, a series and feature mixing block is introduced to learn representations in 1D space. Additionally, a fast Fourier transform is employed to convert the time series into 2D space, providing the capability for 2D feature extraction. Finally, a cross-dimension attentive feature fusion mechanism is designed that adaptively integrates features across different dimensions for anomaly detection. Experimental results on real-world time series datasets demonstrate that CAFFN performs better than other competing methods in time series anomaly detection.