报告人简介：李家康博士剑桥大学博士毕业，美国国家工程院院士，广东康云多维视觉智能科技有限公司联台创始人，CEO兼CTO。中国图像图形学学会第一位外籍会员，华南理工大学校外硕士生导师，广东省图像图形学学会计算机视觉专业委员会副主任，广东省图像图形学学会虚拟现实与智能交互专业委员会副主任。创办英国图形图像引擎公司，实现年收入1.4亿美金；个人拿下Motorola全球总订单5000万美金合同；发明专项70项。创立康云多维视觉智能科技有限公司，康云公司荣获2019年第八届中国创新创业大赛“人工智能与先进制造”成长组全国总冠军。报告摘要：Wealreadyknowthatsupervisedmodelscanperformextremelywellonnumeroustasksbuttheoutstandingaccuracyhasitsownprice.Nevertheless,AIresearchesdoagreatjobincreatingmoreandmorepowerfulmathematicalmodels,theoneswhoarefeedingthesemodelsoftenmisusethemhorriblyorthesemodelsevenaren’tmeanttomeettheexpectations.Algorithmsthatworkwellinresearchphasecanfailontotallyunexpectedproblems.Forexample,itturnedoutthatwidelyusedtodayconvolutionalneuralnetworks(CNNs)aresufferingfromadversarialattacksandtendtolearnnottheshapeofthevisualobjects,butrathertheirtexture(pictureandthelinkabove).Theproblem,probably,liesinthecoremathematicaloperationchosenforthemodel—convolutions,thatarenotrobustenoughastheyseemlike.GenerativeAIwilldemocratise3Dcontentcreationanddisplayinasimilarmannertowhathashappenedinothercreativedisciplinessuchasvideoproduction,photography,imageprocessingandgamedevelopment.AtLight&Magic,we’repassionateaboutbringingthesenew“automation”toolstolifeandlesseningthegapbetweenideaandcreation.WeusedGENERATIVEAI(embeddedinhardware&software)for3DSpatialRealityComputing.Providingultra-fastscanningandmappingsolutionsfor3Dmeasurement,documentation&realisation.