報(bào)告題目：A new encoding computation method for Distributed multi-task learning

報(bào)告時(shí)間：2024年9月6日下午15:30-16:30

報(bào)告地點(diǎn)：kaiyun開(kāi)云官方網(wǎng)站犀浦校區(qū)7教7510

報(bào)告人：程民權(quán)

摘要： Distributed multi-task learning (MTL) enables the joint training of multiple models by leveraging correlations between tasks, which can lead to better generalization performance. However, MTL in distributed settings faces significant communication bottlenecks, particularly in large-scale learning scenarios involving numerous tasks. This talk considers a distributed multi-task learning (MTL) system where distributed workers aim to learn different models orchestrated by a central server. To mitigate the communication bottlenecks in both the uplink and downlink, we propose a novel approach leveraging matrix multiplication. We have presented the results of achieving communication loads for both uplink and downlink under arbitrary layouts, which outperform known results and are close to the theoretical optimum. Specifically, when N = K, we achieve the optimal communication loads for both uplink and downlink as prescribed by information theory, marking a perfect attainment of the theoretical minimum.

報(bào)告人簡(jiǎn)介：程民權(quán)，廣西師范大學(xué)計(jì)算機(jī)科學(xué)與工程學(xué)院教授、博士生導(dǎo)師，2012年博士畢業(yè)于（日本）筑波大學(xué)，并獲得筑波大學(xué)校長(zhǎng)特別表彰?，F(xiàn)為首批廣西高校引進(jìn)海外高層次人才，廣西杰出青年基金獲得者，中國(guó)保密協(xié)會(huì)隱私保護(hù)專(zhuān)業(yè)委員會(huì)委員，主持多項(xiàng)國(guó)家和省部級(jí)項(xiàng)目。主要研究領(lǐng)域?yàn)榫幋a密碼理論、組合設(shè)計(jì)及其在分布式系統(tǒng)、信息安全與隱私、多媒體版權(quán)保護(hù)、機(jī)器學(xué)習(xí)等領(lǐng)域的應(yīng)用。目前已發(fā)表SCI學(xué)術(shù)論文55篇，其中11篇發(fā)表在信息論國(guó)際頂級(jí)期刊《IEEE Transactions on Information Theory》上（6篇曾入選該雜志月度前 50 熱門(mén)文章），1篇發(fā)表在計(jì)算機(jī)網(wǎng)絡(luò)領(lǐng)域國(guó)際頂級(jí)期刊《IEEE/ACMTransactions on Networking》，18篇發(fā)表通信領(lǐng)域頂級(jí)期刊《IEEE Transactions on Communications》上，7篇發(fā)表在《Designs Codes and Cryptography》上。共主持9個(gè)基金項(xiàng)目，其中國(guó)家自然科學(xué)基金2項(xiàng)，廣西自然科學(xué)基金4項(xiàng)。