Instructor highlights: course foundational aim, spanning classical ML (non-neural) early lectures and a later shift to DL-focused topics (GANs, reinforcement learning, transformers).
Structure: first ~7–8 lectures on classical ML and math foundations; last ~5 lectures introduce deep learning, GANs, reinforcement learning, and transformers; a separate DL/NLP course is available next semester for deeper coverage.
Emphasis: math and matrix operations underpin AI/ML; practical hands-on Python throughout the course.
Tools emphasized: Python is essential; Google Colab recommended for simplicity and GPU access; Colab provides GPU power for practice; Colab is used for assignments to ensure consistent execution.
Practical workflow: language/tools are aids, not substitutes for understanding; use tools to assist coding but learn the language and libraries to prompt tools effectively; avoid relying on GenAI for actual assignments; no AI tools during exams.
Progression and outcomes: by course end, students should be proficient in classical ML techniques and familiar with reinforcement learning and GANs; DL details are reserved for the separate DL course.
Syllabus, office hours, and accessibility
Attendance: mandatory for live sessions; if travel or other reasons, request recording.
Office hours: updated to Wednesdays and Fridays, 07:30 PM Eastern; Zoom link provided on syllabus; topic tag for emails should be 8505 to aid prioritization.
Contact: instructor emails with the 8505 tag; response typically within 24–48 hours.
Syllabus edits: schedule times adjusted to fit students; changes posted on slides/Blackboard.
Blackboard resources: slides, electronic reserves, tests, assignments, and exams housed on Blackboard with quick access to a Zoom-based office-hours link.
Python, Colab, and environment choices
Python necessity: heavily Python-based course; a quick proficiency survey shows ~80% familiarity vs ~20% needing catch-up.
Colab vs local: Colab used for uniformity and GPU access; local setups may have variability; Colab mitigates environment issues during assignments.
Alternative environments: you can use other IDEs (e.g., VS Code) if you prefer, but Colab remains the recommended platform for coursework due to compute access and consistency.
AI tooling caveat: tools like Copilot can assist coding, but learning the language and libraries is essential to prompt effectively and avoid over-reliance.
Hardware costs: GPU usage via cloud platforms incurs cost; manage usage to avoid unnecessary expenses.
Textbooks, references, and readings
Textbooks: treated as reference material; core content is what’s covered in class and assignments.
Access to textbooks: department resources provide access; both free and paid options may be available.
Reading guidance: course will indicate precursor chapters/pages; reading ahead is encouraged but not required; focus remains on covered lectures and assignments.
Additional resources: recommended videos such as StatQuest for concepts like decision trees; instructor may provide additional links.
Course topics and structure: classical ML first, DL later
Classical ML focus (pre-DL):
Core algorithms: SVMs, regression, decision trees, ensembles, dimensionality reduction, etc.
Emphasis on math foundations: matrix manipulations, linear algebra.
Practical math needs: concepts like learning rate, optimization basics, and regression/classification frameworks.
Transition to DL topics (later lectures):
Introduction to deep learning, GANs, reinforcement learning, and transformers.
Deep learning emphasizes end-to-end learning from data with neural networks; the DL course covers these topics in greater depth.
Real-world emphasis: ML as a driver for practical solutions; ML lifecycle and data-centric thinking are central.
What is AI, ML, DL? foundational concepts explained
AI: umbrella term for systems that mimic human-like intelligence.
ML: subset of AI where models learn from data to improve performance with experience; not explicitly programmed for every rule.
DL: subset of ML using neural networks; can learn representations (features) from data without explicit feature engineering.
Data science vs data engineering:
Data science focuses on extracting actionable insights from data.
Data engineering focuses on data collection, quality, and provisioning for data scientists.
Practical takeaway: not all problems require DL; classical ML can be sufficient with proper data and feature engineering.
Data, data quality, and the ML pipeline
Data quality is critical: garbage in, garbage out; data must be representative of the problem domain and properly cleaned.
Data pipeline steps (high level):
Data collection and ingestion
Data cleaning and cleansing
Feature extraction/engineering
Splitting data into training, validation, and test sets
Model training and hyperparameter tuning
Evaluation and deployment
Monitoring and maintenance (drift handling and retraining)
Training vs validation vs test sets:
Training data: used to fit the model.
Validation data: used to tune hyperparameters (hyperparameter tuning).
Test data: withheld to assess final model performance.
Data split practice: splits are typically randomized by libraries; ensure the split maintains representative distribution and avoids data leakage.
Overfitting vs underfitting:
Overfitting: model learns noise and training data too well; poor generalization.
Underfitting: model too simple to capture underlying patterns.
Techniques: regularization, early stopping, model complexity control, cross-validation, and proper data splitting.
Normalization/standardization:
Feature scaling via standardization: x^=σx−μ to bring features onto similar scales.
Good practice to prevent certain features from dominating model learning.
Feature engineering and discretization: transform continuous features, one-hot encode categorical features, log-transform skewed features, etc.
Hyperparameters and search methods:
Hyperparameters are set by the programmer and not learned from data.
Tuning approaches: grid search, random search, Bayesian optimization.
Ensembles:
Combine multiple models (e.g., Random Forest, SVM) and use voting/mixture for improved performance.
Evaluation metrics (beyond accuracy):
Confusion matrix foundations, precision, recall, F1 score, ROC-AUC.
Practical emphasis on data: the more data and the higher its quality, the better the learning; DL generally requires far more data than classical ML.
Data ethics and responsibility: ensure data represents the problem domain and respects privacy; monitor models for drift after deployment.
The MNIST example and classification metrics
MNIST dataset overview: 60,000 training images and 10,000 testing images, 28x28 grayscale for digits 0–9.
Confusion matrix: diagonal elements represent correct classifications; off-diagonal elements represent misclassifications.
Why accuracy alone can be misleading: imbalanced classes (e.g., rare digit) can yield high accuracy even with poor performance on minority class.
ROC curve and AUC: area under ROC curve used to compare classifiers across different thresholds; higher AUC indicates better discrimination.
False positives (Type I error) vs false negatives (Type II error): the relative cost depends on the use case (e.g., disease screening may prioritize minimizing false negatives).
Precision-recall trade-offs and F1 score as a balanced metric when class distribution is imbalanced.
ROC curves: concept of true positive rate (TPR) vs false positive rate (FPR); ideal classifier approaches the top-left corner (high TPR, low FPR).
Transformers, attention, and deeper DL topics (high-level)
Transformers and attention: memory of the 2017 paper Attention Is All You Need; parallelizable computation and attention mechanisms to model word relationships.
Use cases: NLP dominates transformer usage, but transformer architectures also adapt to images and other domains.
Why transformers became dominant: scalable attention computations, unlocked by large-scale data and compute power; integration with GPUs/TPUs accelerates training.
Note: detailed transformer internals will be covered in the DL/NLP course; this course provides a conceptual introduction.
Gartner hype cycle and technology context
Gartner framework: technology evolution over time with phases like Innovation, Peak of Inflated Expectations, Trough of Disillusionment, Slope of Enlightenment, and Plateau of Productivity.
For trends like agentic AI, students should be aware of where technologies sit on the curve to anticipate robustness and ROI discussions.
Practical tips for exam preparation and course workflow
Exams: midterms and final are multiple-choice; timing is typically 2 hours; cheat sheet allowed on one sheet of paper; exams will be conducted in-person in the course room, not online.
Practice materials: a practice test is available in Blackboard; use it to verify camera setup and practice environment before the real exam.
Homework vs exams: homework contributes a significant portion of the grade (e.g., ~30%); high-quality homework helps overall performance.
Assignment workflow and timing:
Assignments open at 4:00 PM Eastern; students should work on a Python notebook provided in electronic reserves.
Students download the notebook/script, run it in Colab or their environment, and answer multiple-choice questions in the assignment.
Submissions are due before the next class meeting; scores are released after evaluation.
Cheat sheet strategy: during exams, jot down the core formulas and concepts you’re most likely to forget; use it as a quick reference rather than re-deriving concepts.
Collaboration and integrity: assignments are to be completed individually; network with peers for understanding but avoid sharing solutions.
Common questions and clarifications covered in the session
Environment preferences: Colab is recommended for consistency and GPU access; other environments are acceptable if you can ensure reproducible results; practice tests should be run on your chosen setup to avoid surprises.
The role of math in the course: while practical and applied, math (especially linear algebra and optimization) is central to understanding ML methods; the course emphasizes practical understanding and the math as needed to support intuition.
Data considerations and sampling: ensure your train/validation/test splits are representative and random; visualize data distributions; guard against data leakage; consider dataset drift and model updates post-deployment.
Teaching philosophy: focus on business problems driving the ML lifecycle rather than chasing novelty; use the data and problem statement to guide the model choice and evaluation.
Break policy and pacing: a short break is scheduled during a long lecture; follow the instructor’s timing to maintain pace.
Quick study-ready summaries
Key definitions:
AI, ML, DL, Data Science, Data Engineering
Supervised vs Unsupervised vs Semi-Supervised vs Reinforcement Learning
Core formulas:
Accuracy=TP+TN+FP+FNTP+TN
Precision=TP+FPTP
Recall=TP+FNTP
F1=2⋅Precision+RecallPrecision⋅Recall
Feature scaling: x^=σx−μ
Important practice:
Always keep the test set separate from training data to assess generalization.
Use validation data for hyperparameter tuning; avoid leaking test information.
Balance model complexity and data quantity to avoid underfitting and overfitting; consider regularization and early stopping.
Exam readiness tips:
Review lecture notes and assignments thoroughly.
Practice with the Colab notebooks and be comfortable interpreting confusion matrices and related metrics.
Understand when to use DL vs classical ML given data size and problem complexity.
Closing reminders
This course lays a strong foundation in classical AI/ML and introduces DL concepts with a focus on practical, data-driven problem solving.
Active participation, timely submission of homeworks, and thoughtful preparation for midterms/finals are essential for success in this accelerated program.
Maintain a strong connection with the cohort and instructor via the official channels (email with 8505 tag, Zoom office hours, Blackboard).