Overconfidence in deep neural networks (DNN) reduces the model’s generalization performance and increases its risk. The deep ensemble method improves model robustness and generalization of the model by combining prediction results from multiple DNNs. However, training multiple DNNs for model averaging is a time-consuming and resource-intensive process. Moreover, combining multiple base learners (also called inducers) is hard to master, and any wrong choice may result in lower prediction accuracy than from a single inducer. We propose an approximation method for deep ensembles that can obtain ensembles of multiple DNNs without any additional costs. Specifically, multiple local optimal parameters generated during the training phase are sampled and saved by using an intelligent strategy. We use cycle learning rates starting at 75% of the training process and save the weights associated with the minimum learning rate in every iteration. Saved sets of the multiple model parameters are used as weights for a new model to perform forward propagation during the testing phase. Experiments on benchmarks of two different modalities, static images and dynamic videos, show that our method not only reduces the calibration error of the model but also improves the accuracy of the model.

For hand gesture recognition, recurrent neural networks and 3D convolutional neural networks are the most commonly used methods for learning the spatial–temporal features of gestures. The calculation of the hidden state of the recurrent neural network at a specific time is determined by both input at the current time and the output of the hidden state at the previous time, therefore limiting its parallel computation. The large number of weight parameters that need to be optimized leads to high computational costs associated with 3D convolution-based methods. We introduced a transformer-based network for hand gesture recognition, which is a completely self-attentional architecture without any convolution or recurrent layers. The framework classifies hand gestures by focusing on the sequence information of the whole gesture video. In addition, we introduced an adaptive sampling strategy based on the video content to reduce the input of gesture-free frames to the model, thus reducing computational consumption. The proposed network achieved 83.2% and 93.8% recognition accuracy on two publicly available benchmark datasets, NVGesture and EgoGesture datasets, respectively. The results of extensive comparison experiments show that our proposed approach outperforms the existing state-of-the-art gesture recognition systems.

In recent years, person re-identification based on deep learning approaches has made great progress and achieved good results. However, many of the latest network design methods, which usually deploy ResNet or SENet as the backbone network, were originally designed for classification tasks. Since the person re-identification task is essentially different from the classification task, the structure of the backbone network should be modified accordingly. In this paper, we propose a retrieval network based on a multi-scale backbone architecture, which is specifically suitable for the person re-identification task. By constructing hierarchical residual-like connections within a single residual block, the model learns multi-scale discriminative features of pedestrian images. Unlike many state-of-the-art methods that use complex network structures and concatenate multi-branch features, our proposed retrieval network is implemented using only global features, simple triplet loss, and softmax with cross-entropy loss. The results of extensive experiments show that the proposed network has stronger fine-grained pedestrian representation ability, leading to performance gains for person re-identification tasks. Our proposed network achieves a rank-1 accuracy of 96.03% on the Market-1501 and 92.11% on DukeMTMC-reID datasets while only using global features.

Discriminating out-of-distribution (OOD) input is critical for safely deploying deep learning models in reality. A primary challenge in OOD detection is that the model produces abnormally high prediction-confidence scores for both in-distribution (ID) and OOD samples. The loss function forces the logits to increase abnormally during training, leading to overconfident output in the model. Our work shows that a logit amplitude constraint and an advanced training strategy can mitigate this issue. By constraining the amplitude of the logits during training, the distribution of confidence scores between ID and OOD samples can be separated, thus improving detection efficiency. Experiments demonstrate that our method greatly improves the performance of existing OOD detection methods based on maximum softmax probability scores. FPR95 was reduced by an average of 4.86% in general benchmarks. The proposed approach does not require any model modifications or access to OOD instances and can easily be applied to existing OOD detection methods.