{"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/rectifying-the-data-bias-in-knowledge","title":"Rectifying the Data Bias in Knowledge Distillation","arxiv_id":null,"date":"2021-10-11","proceeding":"ICCV 2021 10","authors":["Boxiao Liu","Shenghan Zhang","Guanglu Song","Haihang You","Yu Liu"],"abstract":"Knowledge distillation is a representative technique for\r\nmodel compression and acceleration, which is important for\r\ndeploying neural networks on resource limited devices. The\r\nknowledge transferred from teacher to student is the mapping of teacher model, or represented by all the input-output\r\npairs. However, in practice the student model only learns\r\nfrom data pairs of the dataset that may be biased, and we\r\nthink this limits the performance of knowledge distillation.\r\nIn this paper, we first quantitatively define the uniformity\r\nof the sampled data for training, providing a unified view\r\nfor methods that learn from biased data. Then we evaluate\r\nthe uniformity on real world dataset and show that existing methods actually improve the uniformity of data. We\r\nfurther introduce two uniformity-oriented methods for rectifying the bias of data for knowledge distillation. Extensive experiments conducted on Face Recognition and Person Re-identification have shown the effectiveness of our\r\nmethod. Moreover, we analyze the sampled data on Face\r\nRecognition and show that better balance is achieved between races and between easy and hard samples. And this\r\neffect can be also confirmed in training the student model\r\nfrom scratch, resulting in a comparable performance with\r\nstandard knowledge distillation.","url_abs":"https://ieeexplore.ieee.org/document/9607686","url_pdf":"https://openaccess.thecvf.com/content/ICCV2021W/MFR/papers/Liu_Rectifying_the_Data_Bias_in_Knowledge_Distillation_ICCVW_2021_paper.pdf","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":[],"tasks":[{"task_slug":"face-recognition","task_name":"Face Recognition"},{"task_slug":"face-verification","task_name":"Face Verification"},{"task_slug":"knowledge-distillation","task_name":"Knowledge Distillation"},{"task_slug":"model-compression","task_name":"Model Compression"},{"task_slug":"person-re-identification","task_name":"Person Re-Identification"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/face-verification-on-ijb-c","task":"Face Verification","dataset":"IJB-C","model":"L2E+IS-sampling","rank_in_archive_order":9,"of":26,"metrics":{"TAR @ FAR=1e-3":"97.05%","TAR @ FAR=1e-4":"95.49%","TAR @ FAR=1e-5":"93.25%","model":"MobileFaceNet","training dataset":"MS1M V3"},"uses_additional_data":false}],"syntology":{"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}