Catastrophic forgetting — Hinglish notes

4.3.12 · AI-ML › Pretraining & Fine-Tuning LLMs

YEH KYUN hota hai? (First principles)

Ek neural network jo kuch bhi jaanta hai woh sab ek shared set of weights $θ$ mein store hota hai. Har task ke liye koi alag "memory slot" nahi hota — knowledge distributed aur overlapping hoti hai.

Jab tum ek naye dataset $D_{new}$ par fine-tune karte ho, tum ek aisi loss minimize karte ho jo sirf naye data ka zikr karti hai:

$L_{new} (θ) = \frac{1}{∣ D _{new} ∣} \sum_{(x, y) \in D_{new}} ℓ (f_{θ} (x), y)$

Gradient descent yeh steps leta hai:

$θ \leftarrow θ - η \nabla_{θ} L_{new} (θ)$

Is update mein purane task ki loss $L_{old}$ ka koi reference nahi hota. Toh optimizer khushi se $θ$ ko us region se bahar le jaata hai jo purane task ke liye achha tha, jab tak naya task improve hota rahe. Purani skill collateral damage ban jaati hai.

YEH DIKHAI KAISE deta hai?

Do loss landscapes imagine karo ek hi weight vector $θ$ par:

$L_{old} (θ)$ — ek valley jiska bottom "task A mein achha" region hai.
$L_{new} (θ)$ — task B ke liye ek alag valley.

Inke minima zyada tar alag-alag jagah hote hain. Pure fine-tuning $θ$ ko naye valley mein slide kar deti hai, purani valley se bahar nikalte hue.

FIX KAISE derive karte hain? (EWC scratch se)

Goal: naye task par train karo bina purane task ke achhe region se bahar nikalte hue. Hum chahte hain ek aisa $θ$ jo naye par achha ho aur purane optimum $θ^{*}$ ke paas rahe, lekin sirf un directions mein "stiff" ho jo purane task ke liye matter karte the.

Step 1 — Bayesian view. Purana task seekhne ke baad hamare paas posterior $p (θ ∣ D_{old})$ hai. Jab naya data aata hai, Bayes kehta hai:

$lo g p (θ ∣ D_{old}, D_{new}) = lo g p (D_{new} ∣ θ) + lo g p (θ ∣ D_{old}) - lo g p (D_{new})$

Yeh step kyun? Purane data ka influence poori tarah posterior $lo g p (θ ∣ D_{old})$ mein capture ho jaata hai — purana data paas rakhne ki zaroorat nahi, bas yaad rakho ki usne $θ$ ke baare mein kya bataya.

Step 2 — Purane posterior ko Gaussian se approximate karo (Laplace approximation) $θ^{*}$ par centered, jisme precision Fisher information $F$ se milti hai:

$lo g p (θ ∣ D_{old}) \approx - \frac{1}{2} \sum_{i} F_{i} (θ_{i} - θ_{i}^{*})^{2} + const$

Yeh step kyun? Kisi bhi minimum ke paas smooth loss quadratic dikhti hai; curvature (2nd derivative) hume batati hai ki har weight kitna important tha. $F_{i}$ bada ⇒ weight $i$ bahut mattered ⇒ use mat hilao.

Step 3 — Ek trainable loss mein combine karo (Elastic Weight Consolidation):

$L_{EWC} (θ) = L_{new} (θ) + \frac{λ}{2} i \sum F_{i} (θ_{i} - θ_{i}^{*})^{2}$

Doosri mitigation strategies (80/20)

Woh 20% jo tumhe zaroor pata hona chahiye:

Method	Core idea	WHY it works
Rehearsal / Replay	Training mein purane data ka thoda sa hissa (ya generated pseudo-data) mix karo	Purani loss ab objective mein hai, toh gradients uski raksha karte hain
EWC / regularization	Important weights ko move karne par penalty do	$θ$ ko key directions mein purane optimum ke paas rehne par restrict karta hai
Parameter isolation (LoRA/adapters)	Base weights freeze karo; chote naye trainable modules add karo	Purane weights literally change nahi ho sakte ⇒ forget nahi ho sakte
Lower LR + fewer epochs	$θ^{*}$ se chote steps lo	Purani valley se kam drift

Steel-manned mistakes

Recall Feynman: 12-saal ke bacche ko samjhao

Socho tumhara brain mitti ka ek lump hai aur jo bhi skill tum seekhte ho woh uspe ek shape dabata hai. Agar tum cycle chalana seekhte ho, phir usi jagah swimming seekhne ke liye bahut zor se dabate ho, toh bike-wali shape flat ho sakti hai aur tum use bhool sakte ho! Catastrophic forgetting exactly yahi hai — computer ek hi "clay weights" ka lump sab cheez ke liye use karta hai, toh naye kaam par zor se dabane se purana kaam erase ho sakta hai. Ise fix karne ke liye ya toh hum (a) purani cheez thodi practice karte rehte hain (rehearsal), ya (b) mitti ke un hisson par chote "mujhe rakho!" springs lagate hain jo important purani shapes hold karte hain (EWC), ya (c) naya skill ek nayi mitti ke tedon mein seekhte hain jo side mein lagayi ho (adapters/LoRA).

Connections

Fine-Tuning LLMs — jahan forgetting sabse zyada kaatti hai.
LoRA and Adapters — parameter-isolation fix.
Fisher Information — EWC mein "importance" $F_{i}$ .
Laplace Approximation — Gaussian posterior kaise aata hai.
Continual Learning — broader problem setting.
Multi-task Learning — "sab kuch jointly train karo" ideal jise EWC/rehearsal approximate karte hain.
Learning Rate & Optimization — chota LR drift kam karta hai.

Flashcards

Catastrophic forgetting kya hai?

Pehle seekhi hui knowledge ka abrupt loss jab network ko naye task par train kiya jaata hai, kyunki gradient updates shared weights ko overwrite kar dete hain jo purane task ko encode karte the.

Naye data par fine-tuning se forgetting kyun hoti hai?

Nayi loss

L_{new}

mein purane task ka koi term nahi hota, toh gradient descent freely

θ

ko purane task ke achhe region se bahar le jaata hai.

EWC loss likho.

L_{EWC} = L_{new} (θ) + \frac{λ}{2} \sum_{i} F_{i} (θ_{i} - θ_{i}^{*})^{2}

EWC mein Fisher information

F_{i}

kya represent karta hai?

Purane task ke liye weight

i

ki importance/curvature; bada

F_{i}

⇒ woh weight matter karta tha ⇒ use hilane par penalty do.

EWC saari learning kyun block nahi karta?

Purane task ke liye unimportant weights ka

F_{i} \approx 0

hota hai, toh woh change karne ke liye free hain — learning spare capacity mein hoti hai.

Teen families of mitigation batao.

Rehearsal/replay, regularization (EWC), aur parameter isolation (LoRA/adapters).

LoRA by design forgetting kyun avoid karta hai?

Yeh base weights freeze kar deta hai aur sirf chote added modules train karta hai, toh original knowledge weights literally overwrite nahi ho sakti.

Steel-man: zyada epochs forgetting kyun badhaate hain?

Naye task par extra epochs

θ

ko purane optimum

θ^{*}

se aur door push karte hain, purane task ki valley se drift badh jaati hai.

EWC quadratic penalty mathematically kahan se aati hai?

Purane-task posterior

p (θ ∣ D_{old})

ka Laplace (Gaussian) approximation

θ^{*}

ke around, jisme curvature

F_{i}

milti hai.