{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/efficient-harmonic-neural-networks-with","title":"Efficient Harmonic Neural Networks with Compound Discrete Cosine Transform filters and Shared Reconstruction filters","arxiv_id":null,"date":"2022-05-27","proceeding":"IEEE Transactions on Neural Networks and Learning Systems 2022 5","authors":["Yao Lu","Le Zhang","Xiaofei Yang","and Yicong Zhou"],"abstract":"The harmonic neural network (HNN) learns a\r\ncombination of discrete cosine transform (DCT) filters to obtain\r\nan integrated feature from all spectra in the frequency domain.\r\nHNN, however, faces two challenges in learning and inference\r\nprocesses. First, the spectrum feature learned by HNN is insufficient\r\nand limited because the number of DCT filters is much\r\nsmaller than that of feature maps. In addition, the number of\r\nparameters and the computation costs of HNN are significantly\r\nhigh because the intermediate spectrum layers are expanded\r\nmultiple times. These two challenges will severely harm the\r\nperformance and efficiency of HNN. To solve these problems,\r\nwe first propose the compound DCT (C-DCT) filters integrating\r\nthe nearest DCT filters to retrieve rich spectrum features to\r\nimprove the performance. To significantly reduce the model size\r\nand computation complexity for improving the efficiency, the\r\nshared reconstruction filter is then proposed to share and dynamically\r\ndrop the meta-filters in every frequency branch. Integrating\r\nthe C-DCT filters with the shared reconstruction filters, the\r\nefficient harmonic network (EH-Net) is introduced. Extensive\r\nexperiments on different datasets demonstrate that the proposed\r\nEH-Nets can effectively reduce the model size and computation\r\ncomplexity while maintaining the model performance. The code\r\nhas been released at https://github.com/zhangle408/EH-Nets.","url_abs":"https://ieeexplore.ieee.org/abstract/document/9783450","url_pdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9783450","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"efficient-harmonic-neural-networks-with","repo_url":"https://github.com/zhangle408/EH-Nets","is_official":0,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"pytorch","reach":null}],"tasks":[],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}