YE KYUN MATTER KARTA HAI: MSE aur cross-entropy do random formulas nahi hain jo yaad karni hain — ye usi recipe (NLL) ke do alag noise models (Gaussian vs. Categorical) pe applications hain.
Har standard loss kiska negative log-likelihood hai?
Data ka, ek assumed noise/probability model ke under.
MSE kis noise distribution ko targets pe assume karke aata hai?
Gaussian (normal) noise, ε∼N(0,σ2).
MSE formula likho.
n1∑i(yi−y^i)2.
MSE mein error ko square kyun karte hain?
Use positive banane ke liye aur bade errors ko zyada penalize karne ke liye.
21(y−y^)2 ka y^ ke w.r.t. gradient?
y^−y (error hi).
Cross-entropy kis distribution ko assume karke aati hai?
Bernoulli (binary) / Categorical (multi-class).
Binary cross-entropy likho.
−[ylogy^+(1−y)log(1−y^)].
One-hot y ke saath multi-class cross-entropy likho.
−∑kyklogy^k=−logy^true.
Softmax+CE ka logit zk ke w.r.t. gradient?
y^k−yk.
Classification ke liye MSE kyun nahi?
Sigmoid+MSE gradient saturate (vanish) ho jaata hai jab confidently galat ho; CE gradient y^−y rakhta hai.
BCE jab y^=0.5 aur y=1?
ln2≈0.693 (coin-flip baseline).
CE ke log ka input kya hona chahiye?
Ek valid probability (0,1) mein, toh pehle softmax/sigmoid lagao.
Recall Feynman: 12 saal ke bache ko samjhao
Ek dartboard socho. Loss hai kitni door tumhara dart bullseye se gaya. Koi number guess karne ke liye (jaise temperature), hum distance measure karte hain aur square karte hain — do baar door jaana chaar baar bura hai (ye MSE hai). Koi category guess karne ke liye (cat vs dog), hum distance nahi maapte; hum poochte hain "sahi jawaab se tum kitne surprised ho?" Agar tumne kaha "99% dog" aur woh dog tha, bahut kam surprise → bahut kam loss. Agar tumne kaha "1% dog" aur woh dog tha, bahut zyada surprise → bahut bada loss (ye cross-entropy hai). Training ka matlab bas apne throws ko nudge karna hai taaki loss shrink ho.
