5 May 2022 in Computational Intelligence and Neuroscience
Channel pruning has been demonstrated as a highly effective approach to compress large convolutional neural networks. Existing differentiable channel pruning methods usually use deterministic soft masks to scale the channelwise outputs and explore an appropriate threshold on the masks to remove unimportant channels, which sometimes causes unexpected damage to the network accuracy when there are no sweet spots that clearly separate important channels from redundant ones. In this article, we introduce a new differentiable channel pruning method based on polarization of probabilistic channelwise soft masks (PPSMs). We use variational inference to approximate the posterior distributions of the masks and simultaneously exploit a polarization regularization to push the probabilistic masks towards either 0 or 1; thus, the channels with near-zero masks can be safely eliminated with little hurt on network accuracy. Our method significantly relieves the difficulty faced by the existing methods to find an appropriate threshold on the masks. The joint inference and polarization of probabilistic soft masks enable PPSM to yield better pruning results than the state of the arts. For instance, our method prunes 65.91% FLOPs of ResNet50 on the ImageNet dataset with only 0.7% model accuracy degradation.
1 November 2021 in Journal of Physics: Conference Series
The rotary cold extrusion forming process is a plastic forming process with very low material loss, especially in the production of hollow screw rods with equal wall thickness. In this work, the rotary cold extrusion forming process of a hollow T2 copper screw rod with a wall thickness of 4 mm and solid T2 copper screw rod was verified by experimental method. The finite element simulation software Deform-3D was also used to simulate the rotary cold extrusion forming process of the screw rod. The effects of the die with different heights of the working belt and the different wall thickness of the billet on the eccentricity, extrusion force, and forming torque in the forming process of the screw rod were studied. The results show that it is feasible to process hollow and solid T2 copper screw rods with equal wall thickness by rotary cold extrusion. With the increase of die working belt height, the eccentricity of the screw rod becomes smaller and closer to the ideal eccentricity. With the increase of the wall thickness of the billet, the amplitude fluctuation range of the eccentricity of the screw rod gradually decreases. The higher the height of the die working belt, the greater the extrusion force and torque required in the extrusion process, and the more intense the change of torque. These results also provide theoretical guidance for the production practice and lightweight transformation of the screw pump rotor.
3 October 2025 in Biomedical Physics & Engineering Express
This paper proposes a novel gesture prediction method for accurately predicting hand gesture types from raw sEMG signals in real time. First, we utilize a linear combination of the mean and standard deviation of sEMG signals within a sliding window to define a new information index in the time domain. Based on this information index, we introduce a new motion detection algorithm that more accurately captures the start and end times of hand gesture motions. Second, we design a new LCNN model, in which LSTM is integrated into the middle of the encoder, allowing for the direct fusion of multi-scale features to prevent the separation of local and temporal features. An ablation study demonstrates that each functional module of the proposed LCNN model positively contributes to the performance of sEMG pattern recognition. The evaluation of the proposed hand gesture prediction method was conducted by comparing it with existing methods using two publicly available datasets. In the experiment involving the dataset Zhang et al (2020 Sensors, 20 3994), the average prediction accuracy for 21 gestures reaches 92.4%. In the experiment with the dataset Krilova et al (2018 UCI Machine Learn. Repo. doi: 10.24432/C5ZP5C), the average prediction accuracy for six hand gestures reaches 82.7%. The results of this study indicate that our motion detection algorithm significantly outperforms the threshold method based on a single time-domain information standard deviation (92.4%, p = 0.0136). Furthermore, our LCNN model also surpasses GRU, LSTM, and other models in terms of prediction accuracy and real-time performance. The research results of this paper highlights the superiority in accuracy and real-time performance of our proposed hand gesture prediction method, which holds great potential for practical applications.