CSCI 4750/5750 · Spring 2027 · Lecture Schedule

42 class meetings (Mon/Wed/Fri 10:00–10:50), Mon Jan 11 – Mon May 3, plus the registrar’s final-exam slot (Wed May 5, 12:00–1:50 PM). Project presentations run on three days: Fri Apr 30 and Mon May 3 (in class) and Wed May 5 (exam slot). Key dates per the official SLU 2026–27 academic calendar and the registrar’s final-exam table: no classes Mon Jan 18 (MLK Day), Wed Feb 10 (Wellness Day), Mar 8–12 (Spring Break), and Fri Mar 26 + Mon Mar 29 (Easter Break); SLU mid-term exam week Mar 1–5; classes end Mon May 3; Study Day Tue May 4; final exams May 5–11.

Readings: [ISLP n] = James, Witten, Hastie, Tibshirani & Taylor, An Introduction to Statistical Learning with Applications in Python (2023), chapter n; [PML n] = Murphy, Probabilistic Machine Learning: An Introduction (2022); [MML n] = Deisenroth, Faisal & Ong, Mathematics for Machine Learning (2020); [UDL n] = Prince, Understanding Deep Learning (2023); [S&B n] = Sutton & Barto, Reinforcement Learning: An Introduction, 2nd ed. (2018). All five are free online (links in the syllabus). Papers are the primary reading where listed.

Every lecture ends its main content with a “Now” segment: where the day’s classical idea lives in 2026–27 industry or research. Those connections are listed in italics below.

Subject to change. I’ll tune pacing and topics as I learn what works for this class, and the “Now” readings especially are tentative. Canvas always has the most current version of everything; watch announcements there for updates.

Unit 1 — Foundations: Learning from Data

# Date Topic & key ideas Reading Out / Due
1 Mon Jan 11 Welcome: what is machine learning? Programs that learn from examples; supervised, unsupervised and reinforcement learning; the recipe every method follows (data, model, loss, optimizer, evaluation); course map; logistics; background survey via in-class questions. Now: one system everyone has used, taken apart into this course’s lectures. ISLP 1; PML 1
2 Wed Jan 13 The learning problem & nearest neighbors. Features, labels, hypotheses; loss, risk and empirical risk; train/test splits; k-NN classification and regression worked by hand; distance choices; the curse of dimensionality. Now: vector search behind recommendation and retrieval is nearest neighbors at billion scale. ISLP 2; PML 16.1
3 Fri Jan 15 Linear algebra for ML. Data as a matrix; vectors, dot products, norms, angles; matrix–vector products as linear maps; independence, rank, inverses; projections; a first look at eigenvectors. Now: embeddings and cosine similarity. MML 2.1–2.6, 3.1–3.4, 3.8 HW1 out
4 Wed Jan 20 Calculus & optimization for ML. Gradients and the chain rule in vector form; Jacobians and Hessians in brief; convexity; gradient descent and the step size; local vs. global minima. Now: a frontier model’s training run is this loop, repeated about a million times. MML 5.1–5.3, 7.1, 7.3
5 Fri Jan 22 Linear regression I: least squares. Model and squared loss; the normal equations on a three-point example; geometry: projection onto the column space; probability: Gaussian noise makes least squares the maximum-likelihood fit; R². Now: linear probes that read concepts out of large models. ISLP 3.1–3.2; MML 9.1–9.2
6 Mon Jan 25 Linear regression II: gradient descent, SGD & features. Closed form vs. iterative cost; SGD and mini-batches; learning-rate pathologies; feature scaling; polynomial and basis features; encoding categories. Now: batch sizes of millions of tokens. ISLP 3.3, 7.1; PML 8.2, 8.4 Project team sign-up opens (Canvas)
7 Wed Jan 27 Generalization: overfitting, bias–variance & cross-validation. Training vs. test error; the polynomial-degree story; the bias–variance decomposition; validation sets and k-fold CV; learning curves. Now: double descent, and why huge networks do not overfit the way the U-curve predicts. ISLP 2.2, 5.1; UDL 8
8 Fri Jan 29 Regularization: ridge, lasso & priors. Ridge in closed form; lasso and sparsity, drawn; regularization paths; choosing λ by cross-validation; MAP view: Gaussian prior ↔︎ ridge, Laplace prior ↔︎ lasso. Now: weight decay in AdamW, in every LLM training recipe. ISLP 6.2; PML 11.3–11.4

Unit 2 — Classification & Classic Supervised Learning

# Date Topic & key ideas Reading Out / Due
9 Mon Feb 1 Logistic regression. Sigmoid and log-odds; cross-entropy as negative log-likelihood; the gradient (p − y)x; decision boundaries; softmax regression for many classes. Now: the last layer of a language model is softmax regression over its vocabulary. ISLP 4.1–4.3; PML 10.1–10.3
10 Wed Feb 3 Evaluating classifiers. Confusion matrices; accuracy and base rates; precision, recall, F1; ROC curves and AUC worked by hand; PR curves; thresholds and costs; calibration. Now: medical-AI evaluation, and benchmark leaderboards. ISLP 4.4.2; PML 5.1
11 Fri Feb 5 Generative classifiers: Naive Bayes & discriminant analysis. Bayes’ rule for classification; the Bayes-optimal classifier; Gaussian class-conditionals → LDA and QDA; Naive Bayes; generative vs. discriminative. Now: generative models used as classifiers. ISLP 4.4; PML 9 HW1 due; HW2 out
12 Mon Feb 8 MLE, MAP & Bayesian inference. Likelihood; the coin-flip MLE; Beta–Bernoulli prior → MAP and the posterior; posterior predictive; Bayesian linear regression and error bars; conjugacy. Now: Bayesian optimization for hyperparameters, Thompson sampling in A/B tests. PML 4.2, 4.5–4.6, 11.7
— Wed Feb 10 No class — Wellness Day.
13 Fri Feb 12 Support vector machines. Maximum margin; hard and soft margins; hinge loss; the SVM as regularized hinge-loss minimization; support vectors; subgradient training. Now: margins in contrastive and metric learning. ISLP 9.1–9.3; PML 17.3
14 Mon Feb 15 Kernels. Feature maps; the kernel trick; polynomial and RBF kernels; Gram matrices worked; kernel ridge regression; where kernels stop scaling. Now: the neural tangent kernel; attention read as a kernel smoother. ISLP 9.3; PML 17.1
15 Wed Feb 17 Decision trees. Recursive partitioning; entropy, information gain and Gini worked by hand; regression trees; depth and overfitting; pruning; interpretability. Now: trees inside nearly every production fraud and credit model. ISLP 8.1; PML 18.1
16 Fri Feb 19 Ensembles I: bagging & random forests. Why averaging reduces variance (with the numbers); the bootstrap; out-of-bag error; feature subsampling; impurity vs. permutation importance. Now: random forests as the default strong baseline. ISLP 8.2.1–8.2.2; PML 18.2–18.4, 18.6
17 Mon Feb 22 Ensembles II: boosting. AdaBoost weights worked by hand; gradient boosting as gradient descent in function space; XGBoost/LightGBM; why boosted trees still win on tabular data. Now: TabPFN, a pretrained “foundation model” for small tables (Nature, 2025). ISLP 8.2.3; PML 18.5; Grinsztajn et al. 2022 (skim) Project teams declared (Canvas, 11:59 PM)
18 Wed Feb 24 Learning theory: why learning is possible. Generalization gaps; a finite hypothesis class and Hoeffding’s inequality; PAC learning; VC dimension (lines in the plane shatter three points); no free lunch. Now: networks that memorize random labels, and what that did to classical theory. Shalev-Shwartz & Ben-David ch. 2–3 (skim); Zhang et al. 2017 (skim) Midterm problem set out
19 Fri Feb 26 ML in practice: the workflow. Framing a problem; splits done right; data leakage, with real cases; preprocessing pipelines; hyperparameter search (grid, random, Bayesian); baselines; reproducibility. Now: the leakage and reproducibility crisis in ML-based science. scikit-learn “Common pitfalls”; Kapoor & Narayanan 2023 (skim) HW2 due
— Mon Mar 1 Project studio (workshop). Teams (declared Feb 22; stragglers matched on the spot) work through a structured worksheet — question, data & evaluation, risks & plan — with the instructor circulating. The worksheet is handed out in class. —
— Wed Mar 3 MIDTERM EXAM (in class, 50 minutes; SLU mid-term week is Mar 1–5). Closed book; every exam problem comes from the set released Wed Feb 24. Covers Lectures 1–19. —

Unit 3 — Unsupervised Learning & Probabilistic Models

# Date Topic & key ideas Reading Out / Due
20 Fri Mar 5 Clustering: k-means & friends. Lloyd’s algorithm worked by hand; why the objective never increases; k-means++; choosing k (elbow, silhouette); hierarchical clustering; DBSCAN in brief. Now: vector quantization inside image and audio tokenizers. ISLP 12.4; PML 21.1–21.3
— Mar 8–12 No class — Spring Break.
21 Mon Mar 15 Gaussian mixtures & EM. Soft clustering; the mixture model; E-step responsibilities on a 1-D example; M-step; EM never decreases the likelihood; k-means as a limit; latent variables. Now: latent-variable thinking behind VAEs and diffusion. MML 11; PML 8.7, 21.4 HW3 out
22 Wed Mar 17 PCA & the SVD. Variance maximization; covariance and its eigenvectors; the SVD; explained variance; reconstruction; “eigen-clothes” on Fashion-MNIST. Now: PCA as the first look at any embedding space. ISLP 12.2; MML 10; PML 20.1
23 Fri Mar 19 Embeddings & nonlinear dimensionality reduction. Manifolds; t-SNE and how it misleads; UMAP; learned embeddings; reading an embedding map honestly. Now: embedding maps of what a model has learned. PML 20.4 (skim); Wattenberg et al. 2016 (Distill) Project proposal due (11:59 PM)
24 Mon Mar 22 Recommender systems & matrix factorization. The user–item matrix; collaborative filtering; low-rank factorization; alternating least squares; implicit feedback; cold start; feedback loops. Now: two-tower retrieval and sequence recommenders at streaming services. PML 22; Koren, Bell & Volinsky 2009
25 Wed Mar 24 Probabilistic graphical models & causal thinking. Bayesian networks; factorizing a joint distribution; conditional independence; Naive Bayes as a network; inference by enumeration worked; correlation vs. causation; Simpson’s paradox; the do-operator in one slide. Now: causal ML in A/B testing and medicine. PML 3.6
— Fri Mar 26, Mon Mar 29 No class — Easter Break (Mar 25–29).

Unit 4 — Neural Networks & Modern ML

# Date Topic & key ideas Reading Out / Due
26 Wed Mar 31 Neural networks I: from perceptrons to MLPs. The perceptron; XOR and the limit of linear models; hidden layers; activations (sigmoid, ReLU, GELU); universal approximation; a 2-2-1 network forward by hand. Now: the same block, stacked a hundred deep. UDL 3–4; ISLP 10.1–10.2
27 Fri Apr 2 Neural networks II: backpropagation & autodiff. Computation graphs; the chain rule on a graph with numbers; backprop through an MLP; vectorized gradients; gradient checking; reverse-mode autodiff and PyTorch autograd. Now: autodiff as the engine of every ML framework. UDL 7 HW3 due; HW4 out
28 Mon Apr 5 Training deep networks. Loss landscapes; momentum; Adam and AdamW; learning-rate schedules and warmup; initialization; batch and layer normalization; dropout; early stopping; the overfit-one-batch ritual. Now: the published training recipes of open-weight models. UDL 6, 9
29 Wed Apr 7 Convolutional neural networks. Convolution worked by hand; filters, stride, padding; pooling; parameter sharing, counted; LeNet → AlexNet (2012) → ResNet; transfer learning. Now: vision transformers, and where CNNs still win. UDL 10–11; ISLP 10.3
30 Fri Apr 9 Sequences, attention & Transformers. Sequence data; RNNs in brief; attention as a soft dictionary lookup worked with numbers; self-attention; the Transformer block; next-token prediction as classification. Now: one architecture for text, images, audio and proteins. UDL 12
31 Mon Apr 12 Self-supervised learning & foundation models. Labels are expensive; pretext tasks; contrastive learning (SimCLR, CLIP) with the InfoNCE loss worked; pretrain → fine-tune → zero-shot; scaling laws in brief. Now: foundation models as the default starting point. PML 19.2; Radford et al. 2021 (CLIP, skim) Milestone check-ins begin
32 Wed Apr 14 Generative models I: autoencoders, VAEs & GANs. Autoencoders; latent spaces; the VAE objective and the reparameterization trick; GANs as a two-player game; judging generated samples. Now: the latent spaces inside image generators. UDL 14, 15, 17 Check-ins
33 Fri Apr 16 Generative models II: diffusion models. Forward noising with numbers; the denoising objective; sampling; guidance; text-to-image and video; flow matching in brief. Now: diffusion beyond pictures: weather, molecules, robot actions. UDL 18; Ho et al. 2020 (DDPM, skim) Check-ins
34 Mon Apr 19 Reinforcement learning I: MDPs & Q-learning. Agent–environment loop; MDPs, returns and discounting; the Bellman equation; value iteration on a gridworld; Q-learning and ε-greedy exploration; DQN on Atari. Now: RL in recommendation, chip design, and data-center cooling. S&B 3, 4.4, 6.5; UDL 19 Check-ins (last block)
35 Wed Apr 21 Reinforcement learning II: policy gradients & learning from feedback. REINFORCE and the log-derivative trick; baselines; actor–critic; AlphaGo; RLHF (reward models + PPO); RL with verifiable rewards for reasoning models. Now: how chat assistants and reasoning models are post-trained. S&B 13; InstructGPT (skim); DeepSeek-R1 (skim)
36 Fri Apr 23 Responsible ML: fairness & interpretability. How bias enters data; fairness criteria (demographic parity, equalized odds, calibration) and why they conflict, worked; the COMPAS debate; feature attributions (permutation, SHAP, LIME); model cards. Now: regulation such as the EU AI Act, and audits. Barocas, Hardt & Narayanan ch. 3; Mitchell et al. 2019 (Model Cards) HW4 due
37 Mon Apr 26 ML in the real world: distribution shift & deployment. Covariate and label shift; train/serve skew; monitoring; A/B tests; feedback loops; data-centric AI; the hidden technical debt of ML systems; cost and latency. Now: what an ML engineer’s week actually looks like. Sculley et al. 2015; Koh et al. 2021 (WILDS, skim)
38 Wed Apr 28 Frontier & wrap-up. Reasoning models and test-time compute; AI for science (protein structure, weather forecasting); robotics and world models; open problems; the full arc from k-NN to foundation models; careers and next courses. selected 2026–27 readings Presentation slides due 11:59 AM Thu Apr 29 (day-1 teams)
— Fri Apr 30 Project presentations, day 1 (in class; 7 min + 2 min Q&A each; day assignments posted Fri Apr 23). — Slides due 11:59 AM Sun May 2 (day-2 teams)
— Mon May 3 Project presentations, day 2 (last class meeting). — Slides due 11:59 AM Tue May 4 (day-3 teams)
— Wed May 5, 12:00–1:50 PM (registrar’s exam slot) Project presentations, day 3 (remaining teams). No written final exam. — Final report due Tue May 11, 11:59 PM

Recurring threads (called out across lectures)

Topics deliberately compressed

Gaussian processes, survival analysis, time-series forecasting, graph neural networks, online learning and bandits, and causal inference beyond a first look get pointers rather than lectures (PML and ISLP cover them for interested students). Language models get one lecture here; CSCI 4845/5845 (NLP) is the course that goes deep.