AIHiranyagarbha · Class 11-12
Train, validation, test: AI ki imaandaar pariksha
Ek model test data par 98% sahi hai, par asli duniya mein fisal jaata hai. Galti kahan hui — model mein, ya pariksha ke tareeke mein?
Socho pehle… phir dekho jawab
Aksar pariksha mein. Agar test data ko baar-baar dekh kar model tune kiya, ya test ki jaankari training mein 'leak' ho gayi, toh score phoola hua ho jaata hai. Isliye data teen hisson mein: train (seekhna), validation (settings chunna), test (aakhri, ek baar).
Meher ne spam message pehchaanne ka model banaya. Validation par settings badalte-badalte usne test set bhi 'bas ek baar' dekhne ka socha. Phir ruki: 'Agar test se tune karungi, toh test bhi training ban jaayega.' Usne test set tab tak band rakha jab tak model final nahi hua — aur report mein wahi score likha jo nikla.
Bias-variance trade-off: expected error ≈ bias² + variance + irreducible noise. Bahut seedha model (high bias) pattern nahi pakad paata — underfitting. Bahut jatil model (high variance) training data ka shor bhi yaad kar leta hai — overfitting. Model jatil karte jaao toh training error ghat-ta rehta hai, par validation error ek point ke baad badhne lagta hai; wahi 'sweet spot' chuna jaata hai. Regularisation, zyada data aur k-fold cross-validation is santulan mein madad karte hain.
Data leakage aur distribution shift: agar ek hi vyakti ke message train aur test dono mein hain, ya future ki jaankari training mein aa gayi, toh score jhootha badhega. Aur agar asli duniya ka data training se alag hai (naye tarah ke spam, nayi bhasha), toh model girega — isliye deployment ke baad bhi nigrani zaroori hai.
Trade-off: test ke liye zyada data rakho toh score bharosemand, par seekhne ko kam data; k-fold dono ka santulan karta hai par compute zyada leta hai.
Ghar pe try karo
20 chhote vaakya likho — 10 'achhi review' aur 10 'buri review' (jaise kisi kitaab ke baare mein). 5 vaakya ek lifaafe mein band karo (test set). Baaki 15 se ek seedha niyam banao (jaise kuch shabdon ki suchi). Aakhir mein lifaafa kholo aur niyam ka asli score likho.
Wow!
Test set ki keemat uske 'anchhue' hone mein hai — jitni baar use dekh kar model sudhaaroge, utna hi woh training ban jaayega!
3 sawaal
- Validation set ka mukhya kaam kya hai?
Jawab dekho
Model ki settings (hyperparameters) chunna. Validation se settings chuni jaati hain; test aakhri, ek baar ki jaanch hai.
- High variance wala model aksar kya karta hai?
Jawab dekho
Overfit. Bahut jatil model training ka shor bhi yaad kar leta hai — overfitting.
- Ek hi vyakti ke message train aur test dono mein hon, toh kya samasya hai?
Jawab dekho
Data leakage — score jhootha badhta hai. Test mein train jaisi jaankari aa gayi, isliye asli kshamata se zyada score dikhta hai.
NCERT se jodo: CBSE Class 11-12 Artificial Intelligence (skill subject) — model evaluation; NCERT Class 11 Maths — statistics