Systematic review of training-free neural architecture search: research progress, core challenges, and future directions

{"Abstract":[0],"Neural":[1,35],"Architecture":[2,36],"Search":[3,37],"(NAS)":[4],"serves":[5],"as":[6,104,185],"a":[7,85,105,112,210],"performance":[8],"engine":[9],"for":[10,204,213],"automated":[11],"machine":[12],"learning,":[13],"driving":[14],"intelligent":[15],"advancements":[16],"across":[17],"various":[18],"domains.":[19],"However,":[20],"its":[21,67,214],"development":[22],"has":[23],"been":[24],"significantly":[25],"constrained":[26],"by":[27,42,165],"the":[28,56,63,92,119,124,133,147,151,161,186,205],"prohibitively":[29],"high":[30],"computational":[31],"costs":[32],"involved.":[33],"Training-Free":[34],"(TFNAS)":[38],"addresses":[39],"this":[40,140,143,177],"bottleneck":[41],"enabling":[43],"rapid":[44],"architecture":[45],"evaluation":[46,154],"through":[47],"extremely":[48],"low-cost":[49],"proxy":[50,128],"metrics,":[51,129],"where":[52],"“training-free\\"":[53],"refers":[54],"to":[55,131,173],"absence":[57],"of":[58,73,89,98,121,127,136,153,207],"candidate":[59],"network":[60],"training":[61],"during":[62],"search":[64,157],"loop,":[65],"and":[66,95,107,123,156,168,192,201,216],"broader":[68],"interpretation":[69],"permits":[70],"offline":[71],"pre-training":[72],"lightweight":[74],"predictors":[75],"that":[76,160,183],"do":[77],"not":[78],"constitute":[79],"search-stage":[80],"training.":[81],"This":[82,196],"paper":[83],"presents":[84],"systematic,":[86],"critical":[87],"review":[88,178,197],"TFNAS.":[90],"First,":[91],"core":[93,181],"concepts":[94],"practical":[96,190],"advantages":[97],"TFNAS":[99],"are":[100],"revisited,":[101],"defining":[102],"it":[103],"distinct":[106],"emerging":[108],"research":[109],"paradigm.":[110],"Second,":[111],"two-dimensional":[113],"taxonomy":[114],"is":[115,171],"proposed":[116],"based":[117],"on":[118],"source":[120],"information":[122],"construction":[125],"methodology":[126],"aiming":[130],"clarify":[132],"developmental":[134],"trajectory":[135],"state-of-the-art":[137],"techniques":[138],"in":[139],"field.":[141],"Furthermore,":[142],"work":[144],"delves":[145],"into":[146],"evolutionary":[148],"mechanism":[149],"involving":[150],"co-development":[152],"metrics":[155],"strategies,":[158],"positing":[159],"joint":[162],"optimization":[163],"characterized":[164],"dynamic":[166],"feedback":[167],"mutual":[169],"reinforcement":[170],"key":[172],"maximizing":[174],"efficacy.":[175],"Finally,":[176],"reveals":[179],"four":[180],"contradictions":[182],"persist":[184],"technology":[187],"progresses":[188],"toward":[189],"application":[191],"outlines":[193],"future":[194],"directions.":[195],"offers":[198],"valuable":[199],"references":[200],"conceptual":[202],"inspiration":[203],"maturation":[206],"TFNAS,":[208],"laying":[209],"solid":[211],"foundation":[212],"reliable":[215],"widely":[217],"applicable":[218],"real-world":[219],"deployment.":[220]}