Responsible AI deployment practices — Hinglish notes

6.4.15 · AI-ML › AI Safety & Alignment

Core Philosophy: Defense in Depth

Responsible AI deployment ka matlab hai multiple safeguards ko layer karna taaki koi bhi single point of failure catastrophic harm na kar sake. Jaise airplane ke redundant systems hote hain, hum assume karte hain ki components fail honge aur accordingly design karte hain.

Teen Pillars

Pre-Deployment Validation — Users ke dekhne se pehle safety prove karo
Runtime Monitoring — Problems emerge hote hi detect karo
Human Oversight Architecture — Rapid human intervention enable karo

1. Pre-Deployment Validation

Staged Kyun? Long Tail Ka Curse

Aapke test set mein 10,000 examples hain. Production mein roz 10 million requests aate hain. Rare edge cases jo 0.01% time (1 in 10,000) occur karte hain, testing mein nazar nahi aayenge — lekin production mein roz 1,000 baar honge.

Staged rollouts aapko ye rare failures limited blast radius ke saath discover karne dete hain.

Shadow Mode Testing

Shadow mode kyun? Real traffic distribution ≠ test distribution. Users creative tarike se systems tod dete hain (adversarial examples, out-of-distribution queries). Shadow mode aapko production-scale data deta hai bina production risk ke.

2. Runtime Monitoring

Deployment "set and forget" nahi hai. Models time ke saath degrade hote hain concept drift ki wajah se (duniya badal jaati hai), data drift ki wajah se (input distribution shift ho jaata hai), aur adversarial attacks ki wajah se.

Monitoring Stack

Formula Track Karne Wale Key Metrics

1. Input Drift Detection (KL Divergence) $D_{K L} (P_{train} ∥ P_{prod}) = \sum_{i} P_{train} (x_{i}) lo g \frac{P _{train} ( x _{i} )}{P _{prod} ( x _{i} )}$

KL divergence kyun? Measure karta hai ki production distribution use karke training distribution approximate karne mein kitna information lost hota hai. High divergence = model unfamiliar inputs dekh raha hai.

Derivation: KL divergence reference distribution ke under expected log-ratio hai. Yeh extra bits hain jo $P_{train}$ ko encode karne ke liye chahiye jab code $P_{prod}$ ke liye optimized ho.

Practice mein: Feature distributions ko weekly track karo. Agar KL divergence threshold se zyada ho jaaye (historical data se tune karo), toh retraining trigger karo. 2. Prediction Confidence Entropy $H = - \sum_{i = 1}^{C} p_{i} lo g p_{i}$ jahan $p_{i}$ class $i$ ke liye predicted probability hai. Kyun? Low entropy (confident predictions) acha hai. Rising average entropy = model uncertain ho raha hai = likely OOD data encounter kar raha hai.

Example: Image classifier ka average entropy: Train: 0.3, Week 1: 0.32, Week 4: 0.55 🚨 Kya hua? Seasonal change (winter → spring). Trees without leaves (trained pe) vs with blossoms (novel).

Worked example Monitoring Dashboard (Real-World)

Scenario: Hospital mein deploy hua medical diagnosis assistant. Dashboard panels:

Prediction volume (requests/hour) — Spike detection (bot attacks, viral news)
Average confidence — 24h rolling window pe entropy track karta hai
Feature drift — Training set ke comparison mein top 10 features ke liye KL divergence
Demographic breakdown — Age/gender ke hisaab se predictions (bias monitoring)
Human override rate — % predictions jo doctors ne reject kiye
Latency P50/P95/P99 — Performance degradation pakadna

Alert triggers:

Override rate 6 hours ke liye > 15% → On-call ML engineer ko page karo
Latency P95 > 2 seconds → Infrastructure scale karo
Feature drift KL > 0.5 → Retraining job queue karo

Yeh setup kyun? Har metric alag failure modes pakadta hai. Override rate usefulness ka ground truth hai. Latency usability affect karta hai. Feature drift future accuracy drops predict karta hai.

Circuit Breakers

Analogy: Jaise ghar ka circuit breaker current spike hone pe trip karta hai. AI circuit breaker "trip" karta hai jab predictions unreliable ho jaati hain.

Worked example Circuit Breaker Implementation

class CircuitBreaker:
    def __init__(self, error_threshold=0.1, window=100):
        self.error_threshold = error_threshold
        self.window = window
        self.recent_errors = deque(maxlen=window)
        self.state = "CLOSED"  # CLOSED = normal, OPEN = tripped
    def call_model(self, input_data):
        if self.state == "OPEN":
            return fallback_prediction(input_data)  # Safe default
        
        prediction = model.predict(input_data)
        confidence = prediction.max()
        
        # Track low-confidence as "error"
        is_error = confidence < 0.7
        self.recent_errors.append(is_error)
        
        error_rate = sum(self.recent_errors) / len(self.recent_errors)
        if error_rate > self.error_threshold:
            self.state = "OPEN"
            alert_oncall("Circuit breaker tripped!")
        
        return prediction

Yeh code kyun? Sliding window (last 100 predictions) recent degradation pakadta hai. State machine flapping rokta hai. Fallback ensure karta hai ki users ko koi na koi response mile.

Real scenario: Spam classifier. Attacker 1,000 adversarial emails bhejta hai. 15% emails ke liye confidence 0.7 se neeche girta hai. 100 emails ke baad circuit breaker trip kar jaata hai. Baaki 900 human reviewers ke paas jaate hain. Blast radius: 1,000 ki jagah sirf 100 misclassified emails.

3. Human Oversight Architecture

Sabse critical safety layer: jab AI fail ho, humans jaldi intervene kar sakein.

Human-in-the-Loop (HITL) Patterns

Feedback Loops

Kill Switches and Rollback

Kyun? Jab production mein model harm karta hai, har minute matter karta hai. Bureaucracy safety ko mar deti hai. Data ke sabse paas wale insaan ko act karne ka authority hona chahiye.

4. Documentation and Transparency

Kyun? Downstream users (doosre engineers, product managers, auditors) ko jaanna chahiye ki model kab use nahi karna hai. Transparency trust build karti hai aur informed consent enable karti hai.

Worked example Model Card Contents (Summarized)

Model: Sentiment Analysis for Product Reviews Intended Use: E-commerce platforms ke liye reviews ko Positive/Neutral/Negative classify karna

Out-of-Scope Uses:

❌ Mental health screening (clinical language pe trained nahi)
❌ Employee performance reviews (privacy concerns) Training Data: 500k Amazon reviews (2020-2023), English only Performance:
Overall accuracy: 91%
Demographic ke hisaab se: Age 18-35 (93%), 35-50 (90%), 50+ (87%)
Known bias: Non-native English speakers ke liye kam accuracy (82%) Limitations:
Sarcasm detection: 68% accuracy (vs 91% literal statements)
Niche products pe degrade hota hai (e.g., industrial equipment) Ethical Considerations: Negativity bias amplify ho sakta hai (negative reviews zyada salient lagte hain). Recommend hai ki sirf labels ki jagah raw scores dikhaye jaayein.

Out-of-scope kyun include karein? Misuse rokne ke liye. Mental health life-critical hai; yeh model uske liye validated nahi hai.

Common Deployment Mistakes

Responsible AI Checklist

Production pe deploy karne se pehle verify karo:

Recall 12-Saal-Ke Bachche Ko Explain Karo

Socho tumne ek robot banaya apni Halloween candy sort karne ke liye. Apne room mein (testing) yeh perfectly kaam karta hai — chocolate yahan, gummies wahan. Lekin jab tum isse school le jaate ho 500 bachchon ki candy sort karne ke liye (deployment), problems saamne aati hain:

Kuch bachchon ke paas doosre countries ki candy hai jo robot ne kabhi nahi dekhi (data drift)
Tumhare 10 bags ke comparison mein 500 bags ke saath robot slow ho jaata hai (latency issues)
Yeh accidentally peanut candy ko nut-free pile mein rakh deta hai — allergy wale bachche ko takleef ho sakti hai! (high-stakes error)

Responsible deployment ka matlab hai:

Pehle kuch doston ke saath test karo (staged rollout) poore school se pehle
Robot ko closely dekho (monitoring) taaki galtiyan jaldi pakad sako
Kisi teacher se double-check karwao (human oversight) uncertain decisions mein
Purana manual sorting method ready rakho (rollback) agar robot kharab ho jaaye
Likho ki robot kis cheez mein acha/bura hai (model card) taaki doosron ko pata chale kab use nahi karna Robot helpful hai, lekin perfect nahi. Yeh steps sab ko safe rakhte hain saath mein time bhi bachate hain.

Connections

6.4.1-Adversarial-examples — Deployment ko production mein adversarial inputs se defend karna hoga
6.4.8-Fairness-metrics — Bias monitoring fairness metrics jaise demographic parity use karta hai
6.4.12-Explainability-methods — Human oversight ke liye interpretable predictions chahiye (SHAP, LIME)
6.3.2-MLOps-principles — Deployment practices broader MLOps lifecycle ka hissa hain
5.2.4-A-B-testing — Staged rollouts A/B testing methodology use karte hain
6.4.14-AI-governanceframeworks — Model cards aur audits governance requirements ke saath align karte hain

#flashcards/ai-ml

Staged rollout kya hota hai aur responsible deployment ke liye yeh kyun critical hai? :: Staged rollout progressively model ko bade populations ke saamne expose karta hai (internal → beta → 1% → 10% → 100%) har stage pe explicit pass/fail criteria ke saath. Yeh critical isliye hai kyunki rare edge cases (0.01% occurrence) testing mein nazar nahi aate lekin production mein 1,000 baar/day hote hain. Staged rollouts in failures ko limited blast radius ke saath discover karte hain.

Shadow mode deployment kya hota hai?

Production mein naya model purane model ke saath saath chalana, jahan users sirf purane model ka output dekhte hain jabki engineers dono models ke predictions offline compare karte hain. Zero user risk ke saath production-scale validation milta hai.

Failure rate p ko confidence 1-α ke saath detect karne ke liye needed samples ka formula derive karo :: Hum chahte hain P(n trials mein zero failures) ≤ α. P(zero failures) = (1-p)^n. Toh (1-p)^n ≤ α → n ln(1-p) ≤ ln(α) → n ≥ ln(α)/ln(1-p) (inequality flip hoti hai kyunki ln(1-p) < 0). Example: p=0.001 ko 95% confidence ke saath detect karne ke liye: n ≥ ln(0.05)/ln(0.999) ≈ 2,996 samples.

KL divergence kya hai aur deployment ke dauran isse kyun track karte hain?

KL divergence D_KL(P_train || P_prod) = Σ P_train(x) log[P_train(x)/P_prod(x)] measure karta hai ki production distribution use karke training distribution approximate karne mein kitna information lost hota hai. High divergence matlab model unfamiliar inputs dekh raha hai (data drift), jo future accuracy drops predict karta hai aur retraining trigger karta hai.

AI circuit breaker kya hota hai?

Ek automated failsafe jo safe fallback (simpler model, human handoff, ya error message) pe revert karta hai jab monitoring anomalies jaise high error rate ya low confidence detect kare. Cascading failures rokta hai "trip" karke jab predictions unreliable ho jaati hain.

Human-in-the-loop systems mein confidence-based routing explain karo :: Predictions confidence scores ke basis pe route hoti hain: high confidence (>0.9) → automated, medium confidence (0.7-0.9) → model recommendation dikhate hue human review, low confidence (<0.7) → senior review jahan anchoring bias se bachne ke liye model hidden hota hai. Automation efficiency aur error prevention ka balance karta hai.

Previous model pe automatic rollback kya trigger karna chahiye?

Error rate 10+ minutes ke liye >5%, user complaints baseline se >3σ spike karein, bias metrics thresholds violate karein, ya koi bhi ML engineer unacceptable behavior observe kare chahe metrics green hon. Rollback one-click hona chahiye aur bina approval ke <5 minutes mein executable hona chahiye.

Model card kya hota hai aur isme kya include hona chahiye?

Ek structured document jo disclose karta hai: intended use, out-of-scope uses, training data, demographics mein performance, known limitations, aur ethical considerations. User-facing AI ke liye required hai taaki yeh informed decisions enable ho sakein ki model kab use nahi karna hai aur transparency ke zariye trust build ho.

Sirf accuracy monitor karna kyun insufficient hai?

Accuracy high rehte hue bhi bias bad sakta hai. Example: 95% accurate content moderation jo political speech ko disproportionately flag karne lagti hai — accuracy high rehti hai (zyaadatar content political nahi) lekin harmful ban jaati hai. Disparate impact, fairness metrics, user satisfaction, aur override rates bhi monitor karne chahiye.

Concept drift kya hai aur continuous monitoring kyun chahiye?

Concept drift tab hota hai jab inputs aur outputs ke beech relationship time ke saath badal jaata hai kyunki duniya badal jaati hai (seasonality, user behavior shifts, new adversarial patterns). January mein deploy kiya model July tak subtly obsolete ho jaata hai. Regular retraining (e.g., quarterly) aur drift monitoring chahiye jo auto-trigger kare retraining jab KL divergence threshold se zyada ho.