Responsible AI deployment practices: Foundations — Hinglish deep dive

6.4.15 · D1 · AI-ML › AI Safety & Alignment › Responsible AI deployment practices

Parent note (Responsible AI deployment practices) bahut saare symbols bahut tezi se throw karta hai: $p$ , $α$ , $n$ , $(1 - p)^{n}$ , $ln$ , $D_{K L}$ , $H$ , $lo g$ , $p_{i}$ . Agar inme se koi bhi magic squiggles jaisa lagta hai, toh yeh page aapka ground floor hai. Hum har ek ko zero se build karte hain, use ek picture se anchor karte hain, aur tabhi parent use karta hai.

Hum order mein jaate hain — har idea pehle wale par lean karta hai.

0. "Model" aur "prediction" kya hota hai?

Kisi bhi symbol se pehle, woh object jiske baare mein sab kuch baat karta hai.

Picture: ek box jisme ek arrow andar jaata hai (input) aur ek arrow bahar aata hai (prediction). Is chapter mein sab kuch ya toh (a) yeh check karna hai ki arrow-out safe hai, ya (b) yeh notice karna hai ki woh safe rehna kab band ho jaata hai.

1. Probability $p$ — "yeh kitni baar hota hai?"

Picture: 1000 marbles ka ek bag socho jisme sirf 1 lal hai. Blindly andar haath daalo — lal ka chance hai $p = 1/1000 = 0.001$ . Yahi woh "0.1% failure rate" hai jiske baare mein parent baat karta hai.

Percent ko $p$ mein baadalna: 100 se divide karo. Toh $5% = 0.05$ , $0.1% = 0.001$ , $100% = 1$ .

2. Ek trial ko $n$ baar repeat karna, aur product $(1 - p)^{n}$

Ab key move: parent poochh raha hai "agar failure probability $p$ se hoti hai, toh kitne tries $n$ mein hum shayad kam se kam ek baar dekh chuke honge?"

Picture: wahi marble bag. Lal = failure with $p = 0.001$ . Not-lal = $1 - p = 0.999$ . Ek pie ka slice: ek patla lal sliver ( $p$ ) aur ek bada grey hissa ( $1 - p$ ).

Figure — Responsible AI deployment practices

Multiply kyun karte hain? Figure dekho. Har try ek coin hai jo "safe" land karta hai probability $1 - p$ se. Do safe tries ek ke baad ek: $(1 - p) \times (1 - p)$ . Teen: $(1 - p)^{3}$ . Chota sa superscript $n$ sirf shorthand hai "ise $n$ baar khud se multiply karo" ke liye. Jaise $n$ badhta hai, yeh product $0$ ki taraf shrink hota hai — matlab eventually aap almost sure ho jaate ho ki failure ko kam se kam ek baar pakad liya. Woh shrinking curve hi poori wajah hai ki staged rollouts kaam karte hain.

3. Logarithm $ln$ — "humne kitni baar multiply kiya?"

Hamare paas hai $(1 - p)^{n} \leq α$ aur hum $n$ solve karna chahte hain. Lekin $n$ exponent mein fase hua hai. Woh tool jo exponent ko zameen par kheench laata hai woh hai logarithm.

Picture: $ln$ ko ek ruler ki tarah socho jo "number of doublings/multiplications" measure karta hai. Ek bada number feed karo toh ek modest number nikalta hai; woh output hi chhupa hua exponent hai.

Loop close karna: derivation, ek line ek time mein

Formula

$(1 - p)^{n} \leq α$ se sample-size rule tak Pehle domain (edge cases jo ise legal banate hain): hume chahiye $0 < p < 1$ taaki $1 - p$ ek proper probability ho strictly $0$ aur $1$ ke beech (iska log defined aur negative ho), aur $0 < α < 1$ taaki $ln (α)$ defined aur negative ho. Agar $p = 0$ toh detect karne ko kuch nahi; agar $p = 1$ toh failure pehle try par certain hai; agar $α \geq 1$ toh hum bilkul koi confidence nahi maang rahe. Teeno degenerate hain aur exclude kiye gaye hain.

Step 1 — shuru karo wahan se jo hum chahte hain. Hum demand karte hain ki "bug miss hua" probability hamare tolerated risk se zyada na ho: $(1 - p)^{n} \leq α$ Kaisa dikhta hai: figure s01 ka shrinking yellow curve pink risk line se neeche dip karna chahiye.

Step 2 — dono taraf ka $ln$ lo. $ln$ increasing hai, isliye $\leq$ preserve hoti hai: $ln ((1 - p)^{n}) \leq ln (α)$

Step 3 — exponent ko neeche slide karo $ln (a^{n}) = n ln (a)$ use karke: $n ln (1 - p) \leq ln (α)$

Step 4 — $ln (1 - p)$ se divide karo. Kyunki $0 < p < 1$ , hamare paas $ln (1 - p) < 0$ hai, isliye isse divide karne par inequality flip ho jaati hai: $n \geq \frac{l n ( α )}{l n ( 1 - p )}$ $ln (α)$ aur $ln (1 - p)$ dono negative hain, isliye unka ratio positive hai — ek sensible number of samples. Loop band ho gaya.

4. Confidence $1 - α$ aur risk $α$

Picture: 20-slot ka ek spinner jisme 1 lal slot hai. Lal par land karna ( $α = 1/20 = 0.05$ ) = hum badkismat rahe aur bug miss kar diya. 19 grey slots = hum ne pakad liya. Chhota $α$ = kam lal slots = zyada samples $n$ chahiye.

Ab parent ka formula plain English mein padha jaata hai:

$kitne samples n \geq \frac{l n ( miss karne ka risk α )}{l n ( 1 - failure rate p )}$

" $1 - α$ confident hone ke liye ki ek failure jo probability $p$ se hoti hai use pakad liya, kam se kam itne samples dekho." (Same sample-size logic experiments par apply hoti hai, dekho 5.2.4-A-B-testing.)

5. Distributions $P$ — "jo aa raha hai uski poori shape"

Runtime monitoring training data ko live data se compare karta hai. Yeh karne ke liye hume ek single object chahiye jo describe kare "inputs kaisa dikhte hain."

Picture: ek histogram — bars ki ek row, har outcome ke liye ek, heights $1$ mein sum karte hain. $P_{train}$ woh histogram hai jis par model ne seekha; $P_{prod}$ woh histogram hai jo production mein actually aa raha hai. Agar woh do shapes drift apart ho jaayein, toh model ek aisi duniya dekh raha hai jiske baare mein use sikhaya nahi gaya tha.

6. $\sum$ symbol — "saare boxes mein add karo"

Picture: bar 1 par point karo, kuch compute karo, likh lo; bar 2 par point karo, compute karo, likh lo; … phir sab total karo. $\sum$ ek "har bar ke liye, aur add karo" instruction hai — isse zyada daraaun kuch nahi.

Example: $\sum_{i} P (x_{i}) = 1$ simply kehta hai "saari bar heights add karo aur $1$ milta hai," jo ek distribution ka defining rule hai.

7. KL divergence $D_{K L}$ — "model kitna surprised hai?"

Ab hum parent ka drift formula padh sakte hain (yaad raho: $lo g$ yahan $ln$ matlab hai, hamare page convention ke mutabiq).

$D_{K L} (P_{train} ∥ P_{prod}) = \sum_{i} P_{train} (x_{i}) lo g \frac{P _{train} ( x _{i} )}{P _{prod} ( x _{i} )}$

Ise piece by piece padhna, un symbols ka use karke jo ab hamare paas hain:

$\frac{P _{train} ( x _{i} )}{P _{prod} ( x _{i} )}$ — har outcome ke liye, ratio ki training ne use kitni baar expect kiya vs. woh actually kitni baar show up karta hai. Ratio $= 1$ ⇒ koi surprise nahi.
$lo g (ratio)$ — natural log "twice as common" aur "half as common" ko equal-sized opposite surprises mein turn karta hai, aur ratio $1$ ko exactly $0$ deta hai.
$P_{train} (x_{i})$ se multiply karo — har surprise ko us outcome ki training ke liye importance se weight karo.
$\sum_{i}$ — har outcome par weighted surprises ko add karo.

Common mistake Support condition — jahan KL blow up ho sakta hai

Ratio $\frac{P _{train} ( x _{i} )}{P _{prod} ( x _{i} )}$ mein $P_{prod} (x_{i})$ se divide hota hai. Isliye hamare paas zaroorat hai $P_{prod} (x_{i}) > 0$ jahan bhi $P_{train} (x_{i}) > 0$ ho — production ko har us outcome ko kuch probability deni chahiye jiske baare mein training ne parwah ki. Agar training ne ek outcome expect kiya ( $P_{train} > 0$ ) jo production kabhi produce nahi karta ( $P_{prod} = 0$ ), toh aap zero se divide karte ho aur $D_{K L}$ infinite ho jaata hai — total drift ka ek legitimate signal, lekin ek number jise aapka dashboard gracefully handle karna chahiye (practice mein hum har $P_{prod}$ bar mein ek tiny $ε$ add karte hain ise finite rakhne ke liye). Dhyan do ki $D_{K L}$ symmetric nahi hai: train aur prod swap karna generally ek alag value deta hai.

8. Entropy $H$ — "ek prediction kitni unsure hai?"

Same log machinery, ek prediction ek time mein (phir se, $lo g = ln$ ).

$H = - \sum_{i = 1}^{C} p_{i} lo g p_{i}$

Picture (figure dekho): do bar charts. Ek confident prediction ek tall spike hai — low entropy. Ek unsure prediction equal height ki flat bars hai — high entropy. Kai predictions par average $H$ ko climb karte dekho, aur tum dekh rahe ho model apna footing khota ja raha hai (parent ka "entropy 0.3 → 0.55" alarm).

Foundations topic ko kaise feed karte hain

Neeche wala figure wahi map hai jo board par draw kiya gaya hai — ise padho agar code block render na ho.

Har arrow ek "yeh pehle chahiye, woh baad mein" hai. Do rivers — sample-size river (left) aur monitoring river (right) — responsible deployment par milti hain. Related downstream topics: 6.4.1-Adversarial-examples (attacks jo entropy spike karte hain), 6.4.8-Fairness-metrics (per-group monitoring), 6.4.12-Explainability-methods, aur 6.4.14-AI-governanceframeworks.

Equipment checklist

Right side cover karo aur zor se jawab do. Agar koi stumps kare, toh us section ko dobara padho.

p

ka kya matlab hai aur percent se use kaise nikalte ho?

Ek probability

[0, 1]

mein; percent ko 100 se divide karo (toh

0.1% = 0.001

1 - p

kya hai?

Woh probability ki event nahi hoti (iska complement).

(1 - p)^{n}

kyun

n

tries mein zero failures ka chance hai?

Har try independently failure se bachti hai prob

1 - p

se;

n

ko multiply karna all-safe deta hai.

Kaunsa tool ek variable ko exponent mein fase se free karta hai, aur kyun?

Logarithm, kyunki

ln (a^{n}) = n ln (a)

exponent ko plain multiplier tak neeche slide karta hai.

Sample-size formula valid hone ke liye

p

aur

α

ka kya domain hona chahiye?

0 < p < 1

aur

0 < α < 1

, taaki dono logs defined aur negative hon.

Sample-size formula mein inequality

\geq

mein kyun flip hoti hai?

Kyunki

ln (1 - p)

negative hai, aur ek negative number se inequality divide karne par direction flip ho jaata hai.

α

aur

1 - α

kya represent karte hain?

α

= bug miss hone ka accepted chance;

1 - α

= ise pakadne ki confidence.

Ek distribution

P

kya hai aur kaunsi shape ise picture karti hai?

Har outcome ke liye ek probability (milake 1 hoti hai); ek histogram.

\sum_{i}

aapko kya karne ka instruction deta hai?

Har outcome

i

ke liye, expression compute karo aur sab ko add kar do.

Is page par saare

lo g

symbols ka base kya hai?

Natural log (

ln

, base

e

) — poore page mein

ln

jaisa hi.

D_{K L}

kya measure karta hai aur kab zero hota hai?

Do distributions kitni alag hain; zero jab woh identical hon.

D_{K L}

ke liye kaunsi support condition chahiye, aur agar fail ho toh kya hota hai?

P_{prod} (x_{i}) > 0

jahan bhi

P_{train} (x_{i}) > 0

ho; warna zero se divide hota hai aur infinite ho jaata hai.

Entropy

H

kya measure karti hai, aur low value ka kya matlab hai?

Ek prediction ki uncertainty; low

H

= ek confident spike.

H = - \sum p_{i} lo g p_{i}

mein leading minus sign kyun hai?

Kyunki

p_{i} < 1

ke liye

lo g p_{i}

negative hota hai; minus uncertainty ko positive banata hai.