CS7643l Quizl 3l (Latestl 2025/l 2026l Update)l Questionsl &l Answers|l Gradel A|l 100%l Correctl (Verifiedl Solutions)
Q:l AlexNetl Keyl Aspects
Answer:
-l Morel kernelsl inl eachl convolutionl layerl asl depthl isl increased -l ReLUl insteadl ofl sigmoid/tanh -l NORMl layers -l PCA-basedl datal augmentation -l Dropout -l Ensemblesl 7l NNl thatl werel trainedl randomlyl andl combinedl forl prediction
Q:l VGGl Keyl Aspects
Answer:
-l Usesl repeatedl modulesl inl sequencel (3l Xl 3l filters) -l 2l Xl 2l poolingl withl stridel ofl 2 -l Veryl largel numberl ofl parameters -l Usesl smallerl filtersl asl deeperl networkl maintainsl receptivel fieldl withl smallerl numberl ofl params -l Mostl memoryl usedl inl convolutionl layers -l Mostl paramsl inl FCl layers
Q:l Inceptionl Keyl Aspects
Answer:
-l Usesl parallell filters;l thatl is,l manyl filtersl ofl differentl sizesl inl thel samel layer.l Thisl allowsl forl featurel extractionl atl multiplel scalesl inl eachl layer.-l Alsol usesl 1l Xl 1l filters 1 / 3
-l Sincel increasingl depthl makesl optimizationl harder,l networkl addedl skip/residuall connections.l Thisl allowsl infol froml onel layerl tol propagatel tol anyl futurel layer.l Thisl improvesl learningl andl gradientl flow.
Q:l Transferl Learning
Answer:
-l Wel canl reusel featuresl learnedl froml al networkl trainedl onl al largel datasetl forl ourl specializedl task.l Thusl wel canl significantlyl reducel thel amountl ofl datal wel need.
1.l Trainl somel networkl onl largel dataset 2.l Initializel newl networkl withl weightsl learnedl inl (1) -l ifl wel havel al differentl numberl ofl outputl classes,l lastl layerl mayl needl tol bel altered.
3.l Trainl onl newl dataset
a.l Finel tune:l allowsl parametersl tol bel updated
b.l Freeze:l Onlyl weightsl ofl finall layerl arel updatedl (goodl whenl don'tl havel enoughl data)
Q:l Transferl Learningl Advantagesl andl Disadvantages
Answer:
-l Worksl welll evenl withl differentl numberl ofl outputl classes -l Generalizesl acrossl differentl tasks
-l Doesn'tl workl welll ifl thel sourcel datasetl isl veryl differentl froml thel targetl dataset -l Ifl youl havel enoughl data,l mayl getl betterl resultsl froml trainingl directlyl onl targetl task
Q:l Optimizationl Error
Answer:
Optimizationl algol isl unablel tol findl al goodl locall minimum
Q:l Estimationl Error
Answer: 2 / 3
Wel mayl bel overfittingl andl thusl notl generalizel well
Q:l Modelingl Error
Answer:
Therel mayl bel nol modell thatl canl modell thel reall world.l Thisl decreasesl withl morel depthl andl withl thel numberl ofl parametersl inl al model,l butl suchl alsol increasesl optimizationl andl estimationl error.
Q:l Effectivenessl ofl Morel Data
Answer:
Smalll Datal Regionl -l notl enoughl datal tol reducel error Power-lawl Regionl -l Generalizationl errorl decreasesl linearlyl withl thel logl ofl thel data Irreduciblel errorl Regionl -l Morel datal won'tl improvel modell (reducel error)
Q:l Distributedl Representation
Answer:
Nol onel Neuronl representsl al particularl feature,l thusl interpretationl isl difficult
Q:l Saliencyl Map
Answer:
Showsl gradientsl ofl lossl withl respectl tol eachl inputl pixel
Q:l Gradientl Basedl Visualizationl Uses
Answer:
-l Canl getl objectl segmentationl forl free -l Canl bel usedl tol detectl bias
- / 3