Machine Learning, Deep Learning & GenAI – Comprehensive Study Notes
Distinguishing Key Terms
Artificial Intelligence (AI)
Broad field aiming to create systems that exhibit behaviors normally requiring human intelligence.
Encompasses rule-based systems, search, optimization, and all data-driven approaches.
Machine Learning (ML)
Subset of AI that learns patterns from historical data to make predictions about future, unseen cases.
Central idea: instead of hand-coding complex rules, an algorithm infers them from data.
Accepts that results are probabilistic; errors are expected but manageable.
Deep Learning (DL)
Family of ML techniques using artificial neural networks with many hidden layers.
Employs backpropagation (first formalized in the by Hinton et al.) to adjust millions of parameters.
Excels at handling high-dimensional, unstructured inputs (images, audio, raw text) while still producing relatively simple outputs (e.g.
"cat" vs "dog", sentiment label, fraud vs non-fraud).
Generative AI (GenAI)
Newly prominent branch of DL focused on self-supervised learning from massive corpora.
Produces complex, creative outputs (paragraphs, code, images) in response to open-ended prompts.
Trained on vast, unlabeled internet-scale text/image datasets; learns distribution of natural language or pixels and samples from it.
Illustrative Business Scenario: Churn Prediction
Problem: In any subscription service (e.g.
a mobile carrier), losing customers (“churn”) is costly; acquiring new ones costs more than retaining existing ones.Traditional (rule-based) approach
Engineers attempt to hand-encode rules: if user made support calls AND has monthly minutes THEN high risk, etc.
Quickly becomes unscalable as variable count explodes; maintenance is brittle.
ML approach
Feed historical labeled data: each row = customer profile + flag "churned" or "stayed".
Algorithm induces a predictive model capturing subtle, multidimensional patterns.
Output: for any current customer, probability of churn; marketing team offers incentives only to the most at-risk, optimizing cost.
Key take-away: ML’s strength is future prediction leveraging past data, not simply describing history.
Why Now? – Hardware & Cloud Catalysts
Early AI research in the – produced foundational ideas but lacked computational power.
Neural-network renaissance owed to:
Cheap, massively parallel GPUs enabling trillions of multiply-accumulate operations.
Elastic cloud platforms providing on-demand clusters; teams sidestep capital expense.
Result: Algorithms like backpropagation become practical on large datasets.
Choosing ML vs. Deterministic Software
Use conventional code when:
Problem can be expressed as unambiguous rules; outputs are fully determined by inputs (e.g.
plotting last week’s network usage).
Consider ML when:
Variables are too numerous/interactive for manual rule writing.
Personalization at scale is required (recommendations for every visitor within milliseconds).
Data streams in real-time and quick classification/anomaly flagging is critical.
Environment drifts and model must adapt (concept drift monitoring still needed).
Detailed Comparison of Output Complexity
Classic ML and DL models: complex input → simple output
Fraud detection: transaction vector → {fraud, non-fraud}.
Image classification: pixels → {cat, dog, bird}.
Sentiment analysis: sentence → {positive, neutral, negative}.
GenAI: complex input → complex output
Natural-language prompt → multiple paragraphs, code snippets, or visual art.
Real-World ML Use Cases Highlighted
E-commerce recommendation & personalization
"Customers who bought X also bought …"; must return within ms for seamless UX.
Anomaly Detection
AWS GuardDuty baseline-driven; flags unusual open ports or traffic patterns.
General recipe: model normal distribution, surface deviations.
Computer Vision
Object tracking across frames: first appearance → disappearance timestamps.
Geospatial / Overhead imagery
Detecting structures, vehicles, or environmental changes.
Log Analytics
Extracting “interesting” events amid terabytes of machine logs.
Time-Series & Sensor Streams
Identifying anomalies in IoT data, radar returns, biomedical signals.
Text Processing & PII Scrubbing
Automatically masking emails, dates of birth, or other sensitive tokens—even when they hide in headers or embedded metadata.
Especially vital under strict regulations (e.g.
GDPR in Europe).
Example: PII Removal Workflow Challenge
Manual regex-based scrubbing fails whenever new entity types appear.
Pipeline pain-points:
Each newly discovered PII pattern forces code change, redeployment.
Headers vs.
body mismatch led to unexpected leaks in a real project.
ML advantage: train a model on examples of sensitive vs.
non-sensitive tokens; model generalizes beyond enumerated rules.
Error Tolerance & Model Evaluation
No ML system is accurate; organizations must define acceptable error trade-offs (false positives vs.
false negatives).Hyperparameter tuning, threshold adjustment, and cost-sensitive metrics help align with business tolerance.
From Proof of Concept (PoC) to Production
Many PoCs fail when:
Data-access scripts used in experiments are not production-grade.
No plan for automated retraining, monitoring, or model drift detection.
Best Practices:
Define data pipelines, versioning, and governance before model training.
Use CI/CD for models; treat them like software artifacts.
Integrate logging and A/B testing to verify real-world impact.
Ethical & Practical Implications
Customer trust: Misclassifying loyal users as churn risks damaging relationships.
Bias risk: Historical data can encode unfair patterns; fairness audits are mandatory.
Privacy: Large language models ingest public data—must respect copyright and personal data constraints.
Quick Recap / Checklist for ML Applicability
High-dimensional data? ✅
Need rapid, individualized predictions? ✅
Environment or patterns evolving? ✅
Deterministic rules infeasible? ✅ → ML likely beneficial.
Session Logistics
Content derived from an introductory module on machine learning; presenter offered a -minute break afterward.