Ibajẹ iwuwo ati adanwo Regularization L2

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

1. Kini L2 deede ṣe afikun si iṣẹ isonu naa?
   • Awọn nọmba ti fẹlẹfẹlẹ ni awọn nẹtiwọki
   • Ijiya lori oṣuwọn ẹkọ
   • Awọn kika ti ikẹkọ apẹẹrẹ
   • Ijiya ti o ni ibamu si apao awọn iwuwo onigun mẹrin
2. Kini iṣoro akọkọ ti ibajẹ iwuwo ṣe iranlọwọ fun idena?
   • Ikojọpọ data ti o lọra
   • Overfitting, nibiti awoṣe ṣe akori ariwo ni data ikẹkọ
   • Hardware ikuna
   • Awọn akole ti ko tọ
3. Ninu itọsọna wo ni ibajẹ iwuwo nfa awọn iwuwo awoṣe?
   • Si ọna ailopin
   • Si ọna iwuwo ti o tobi julọ
   • Si ọna odo
   • Si ọna awọn iye laileto
4. Kini idi ti AdamW ṣe ṣafihan?
   • Lati decouple àdánù ibajẹ lati imudojuiwọn gradient adaptive ki ibajẹ huwa ti o tọ
   • Lati yọkuro gradient patapata
   • Lati ṣe ikẹkọ losokepupo lori idi
   • Lati yọkuro iwulo fun oṣuwọn ikẹkọ
5. Fun isọkalẹ sitokasitik itele, bawo ni isọdọtun L2 ati ibajẹ iwuwo ṣe ni ibatan?
   • Wọn jẹ idakeji nigbagbogbo
   • Wọn jẹ deede mathematiki
   • Ibajẹ iwuwo ṣe alaabo gradient
   • L2 nikan ṣiṣẹ fun SGD pẹlu ipa
6. Kini iṣakoso olùsọdipúpọ lambda ni isọdọtun L2?
   • Iwọn ipele naa
   • Awọn nọmba ti epochs
   • Iṣẹ imuṣiṣẹ
   • Agbara ti ijiya lori awọn iwuwo nla
7. Why can adding an L2 term to Adam’s loss differ from applying decoupled weight decay?
   • Adaptive per-parameter gradient scaling can distort the coupled penalty
   • Adam cannot train with any regularization
   • The two methods use different training labels
   • Decoupled decay removes the data loss
8. In a simple decoupled update, what changes the fractional shrinkage of a weight?
   • Only the number of classes
   • Only the model’s input dimension
   • The learning rate and decay coefficient together
   • Only the validation-set size