AIPushan · Class 9-10
Ratta maarne wali machine: overfitting
Jo dost pichhle saal ke paper ke jawab ratke poore number laata hai, kya woh naye sawaal par bhi utna hi achha karega? Machine ke saath bhi yahi sawaal hai.
Socho pehle… phir dekho jawab
Shayad nahi. Jab model training data ko itna 'yaad' kar le ki uske sanyog aur shor bhi pakad le, toh naye data par fisalta hai. Ise overfitting kehte hain. Lakshya hai generalisation — naye examples par bhi sahi hona.
Naina ne science mele ke liye ek model banaya jo photo se aam aur amrood alag karta tha. Training photos par poora sahi! Par naye photos par baar-baar galat. Pata chala, saare aam ki photo laal kapde par kheenchi gayi thi — model ne 'laal kapda = aam' seekh liya tha. Naina chahti toh sirf training wala score dikha deti; usne naye test ka asli score bhi poster par likha aur photos alag-alag background par dobara kheenchi.
Mechanism: data ko do hisson mein baanto — training set (seekhne ke liye) aur test set (jise model ne kabhi nahi dekha). Training score bahut achha aur test score kharab = overfitting. Dono kharab = underfitting (model itna seedha ki pattern hi nahi pakda). Ilaaj: zyada aur vividh data, seedha model, aur test set ko alag rakhna.
So what: ratta-padhai mein bhi yahi hota hai — asli samajh woh hai jo naye sawaal par kaam aaye. Aage Class 11-12 mein yahi idea validation set aur bias-variance trade-off ban jaata hai.
Ghar pe try karo
Ghar ki 8 cheezein lo (chamcha, katori, pencil, rubber, rumaal, aadi) aur unhe 'rasoi' aur 'padhai' ke do dher mein baanto. Ek lamba niyam likho jo in 8 ki har chhoti baat yaad rakhe (rang, size), aur ek chhota saaf niyam (kis kaam aati hai). Ab 4 nayi cheezein lo — kaun sa niyam naye par bhi kaam karta hai?
Wow!
Model aksar sabse aasaan pattern pakadta hai — chahe woh background ka rang hi kyun na ho! Isliye test data alag aur vividh rakhna zaroori hai.
3 sawaal
- Training par bahut achha aur test par kharab score kya dikhata hai?
Jawab dekho
Overfitting. Model ne training data yaad kar liya, naye data par generalise nahi kiya.
- Test set ki sabse zaroori khaasiyat kya hai?
Jawab dekho
Model ne use seekhte waqt kabhi na dekha ho. Naye, andekhe data par hi asli seekh naapi jaati hai.
- Naina ke model ne galti se kya seekh liya tha?
Jawab dekho
Laal kapda = aam. Saare aam laal kapde par the, isliye model ne background ko hi pehchaan bana liya.
NCERT se jodo: CBSE Class 9-10 Artificial Intelligence (skill subject) — AI project cycle: data, modelling aur evaluation