混合信号:面向 EEG 与 ECG 的权重空间编程

研究论文 Hugging Face Papers 原文语言:中文

EPFL 的研究显示,前沿语言模型(如 Claude Opus 5.5)可以直接读写生物信号模型的权重:仅凭架构、任务和少量受试者就能从头写出全部权重,用 6 倍更少的训练受试者匹配梯度下降训练效果,并在同样 5 名受试者上超过传统 EEG 流水线,但模型无法改进已训练好的网络。

全文

Weight-spaceprogrammingallowslanguagemodelstoreadatrainedneuralnetwork'sweightsandeditthem.Thisopensthedoorto``training''byreasoningratherthanthroughgradientdescent.Priorworkhasdemonstrateditsfunctionalityandpromiseforgame-playingnetworksandfixingspecificfailuresinlanguagemodels,evenextendingtodiffusionmodels.Inthiswork,wepresentitsapplicabilitytomodelstrainedonbiosignaldata,specificallyEEGandECGdata.Weshowthatfrontiermodelsarenotonlyabletodebugspecificfaultsplantedinbiosignal-basedmodelswithoutbeingtoldwhatthesefaultsare,butareevencapableofwritingalltheweightsofanetworkfromscratch,givenonlythearchitecture,thetask,andaminimalsetoftrainingsubjects.Inparticular,inonesetofexperiments,alanguagemodelwasabletomatchstandardgradient-descenttrainingusing6timesfewertrainingsubjects,throughreasoningaloneandwithoutanygradientdescentstep.NetworkswrittenbyClaudeOpus5.5alsoscoredhigherthanclassichand-craftedEEGpipelines(alphabandpower,andcommonspatialpatternswithlineardiscriminantanalysis)whenbuiltfromthesame5subjects.Wealsohighlightsomeofthelimitations,whichincludetheinabilityoflanguagemodelstoimprovealreadywell-trainedclassificationmodels,andtheirincapacitytoreliablyrecognize,fromtheweightsalone,thetaskanetworkwastrainedon.Overall,ourresultsindicatethatweight-spaceprogrammingcanbeusefulforsmallnetworks,suchasthosedeployedonedgedevicesandwearables,buttheconceptstillrequiresfurtherinvestigationbeforeitcanbeusedinpractice.

View arXiv pageView PDF Project page Add to collection

正文以来源语言显示。

参考来源

← 返回每日简报