LSTM, Clearly Explained ↗
Before or after H1: explain the roles of sigmoid, tanh, retained memory, and exposed state.
Creator’s lesson / source codeNCKU · Fall 2026 · Monday 14:00–17:00 · English-taught
Machine Learning
in Engineering Science
Understand the mechanism. Run the experiment. Reproduce the result. Explain what the evidence supports.
Instructor: Prof. Chi-Hua Jonny Yu
Focus: ML foundations, reproducibility, interpretability, and bounded agent workflows.
01 · LEARN
Work through definitions, mechanisms, equations, and numerical examples.
02 · REPRODUCE
Keep data identity, split boundaries, seeds, versions, and actual measured outputs.
03 · EXPLAIN
Connect results to assumptions; test interventions and state limits.
The complete semester
One connected learning path.
Monday meetings run from 14:00 to 17:00. ML foundations lead into sequence modeling, reliable evaluation, and agent workflows. The 19 October class follows the three-part sequence requested for this week.
Read · predict · run · modify · explain
Your weekly workbench.
Every teaching week has a complete worked practice notebook with Markdown explanations, expected checkpoints, troubleshooting, and an evidence export. The same .ipynb works in Colab and local Jupyter.
Weekly notebooks are in-class practice and optional extensions. Graded Homework occurs only at announced milestones. Midterm and final assessments each include a written examination and an oral examination.
W01 · 07 SepCourse introduction, data, and evidence
Describe a row, a unit, a target, and an evidence chain.
Week 01 guided practice
Describe a row, a unit, a target, and an evidence chain.
Open course-introduction slides
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W02 · 14 SepFeatures, targets, and leakage-safe splits
Build a preprocessing pipeline and inspect held-out classification errors.
Week 02 guided practice
Build a preprocessing pipeline and inspect held-out classification errors.
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W03 · 21 SepSupervised learning: regression and classification
Choose a baseline and distinguish continuous from categorical targets.
Week 03 guided practice
Choose a baseline and distinguish continuous from categorical targets.
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W04 · 28 SepNo class — Teachers’ Day holiday
Follow the current semester calendar; no new in-class practice.
No new notebook is assigned on this calendar holiday. You may revisit the previous teaching module.
W05 · 05 OctEvaluation, validation, and overfitting
Explain precision, recall, specificity, thresholds, and validation. Homework milestone: ML Problem Brief.
Week 05 guided practice
Explain precision, recall, specificity, thresholds, and validation. Homework milestone: ML Problem Brief.
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W06 · 12 OctUnsupervised learning and representation
Interpret PCA and clusters while separating geometry from physical meaning.
Week 06 guided practice
Interpret PCA and clusters while separating geometry from physical meaning.
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W07 · 19 OctSequences → LSTM → Attention → Transformer
Trace LSTM memory, compute attention, and train/test a tiny causal Transformer across three one-hour modules.
H1 · 14:00–15:00 · 60 MIN · 30 SLIDES
Sequential Data and LSTM
From ordered windows and honest splits to RNN state, LSTM gates, trained forecasts, and baseline comparisons.
H2 · 15:00–16:00 · 60 MIN · 30 SLIDES
Why Attention, and How It Works
Resolve the context bottleneck; compute QK scores, softmax and weighted retrieval; inspect masks and interpretation limits.
H3 · 16:00–17:00 · 60 MIN · 30 SLIDES
From Attention to a Transformer
Add positions, heads, residuals and feed-forward layers; train a small decoder and test future-token isolation.
How the three hours work: about 90 minutes of explanation and worked reasoning, with 90 minutes of prediction, guided execution, experiments, and explain-back distributed across the three modules. Every slide has a two-minute facilitation cue; notebook operations provide the practice within that time.
Completion evidence: an LSTM baseline comparison, a verified attention calculation, a Transformer causality test, and one saved evidence record per notebook.
Download the full W07 LaTeX lecture chapter · Read the nine-page W07 PDF · Download all three slide decks and notebooks
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W08 · 26 OctNo class — observed Retrocession Day holiday
Consolidate the sequence notebooks; no new in-class practice.
No new notebook is assigned on this calendar holiday. You may revisit the previous teaching module.
W09 · 02 NovML reliability, shift, and error analysis
Stress-test model assumptions and report subgroup failures.
Week 09 guided practice
Stress-test model assumptions and report subgroup failures.
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W10 · 09 NovReproducible ML workflow studio
Rebuild a baseline from frozen data and split IDs. Homework milestone: ML Workflow.
Week 10 guided practice
Rebuild a baseline from frozen data and split IDs. Homework milestone: ML Workflow.
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W11 · 16 NovMidterm written and oral examinations
Explain project data, baseline, evaluation, limitations, and reproduction.
Week 11 guided practice
Explain project data, baseline, evaluation, limitations, and reproduction.
Optional exam rehearsal
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W12 · 23 NovLLMs, bounded agents, and vibe coding
Separate next-token generation from validated action execution.
Week 12 guided practice
Separate next-token generation from validated action execution.
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W13 · 30 NovRetrieval, source evidence, and structured knowledge
Build a lexical retrieval baseline with source IDs and abstention.
Week 13 guided practice
Build a lexical retrieval baseline with source IDs and abstention.
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W14 · 07 DecAgent testing, failure cases, and oversight
Enforce permissions, step budgets, and human review. Homework milestone: Final Agent Plan.
Week 14 guided practice
Enforce permissions, step budgets, and human review. Homework milestone: Final Agent Plan.
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W15 · 14 DecAgent Studio I: build a bounded workflow
Produce traceable descriptive results with a validated input schema.
Week 15 guided practice
Produce traceable descriptive results with a validated input schema.
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W16 · 21 DecAgent Studio II: evaluation and demo rehearsal
Compare candidates using reference cases and recorded failures. Homework milestone: Demo Readiness.
Week 16 guided practice
Compare candidates using reference cases and recorded failures. Homework milestone: Demo Readiness.
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W17 · 28 DecProject integration and oral-exam preparation
Package data identity, code, result cards, limitations, and AI-use disclosure.
Week 17 guided practice
Package data identity, code, result cards, limitations, and AI-use disclosure.
Classroom exercise
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
W18 · 04 JanFinal written and oral examinations
Defend the final project with a reproducible evidence package.
Week 18 guided practice
Defend the final project with a reproducible evidence package.
Optional exam rehearsal
Predict, run, change one factor, and explain the evidence. This is not an additional weekly graded assignment.
Homework · Forecasting and explanation
Build a stock forecasting system.
Compare persistence, Ridge regression, and LSTM using a reproducible chronological evaluation. Explain your system and evidence in a five-minute English video. This is a separately announced HW, not an extra weekly exercise.
Stock Market Prediction System
Includes the full English brief, grading rubric, data-leakage rules, submission checklist, and a guided Colab/Jupyter starter. Free CPU workflow; no paid API or deployment required. Due date: to be announced.
Supplementary explanations
Watch with a question in mind.
The lecture and notebooks are self-contained. These creator-hosted videos provide alternative explanations and optional deeper study; the three-hour class does not depend on watching full videos.
Attention in transformers, step by step ↗
After H2: pause before softmax and compute a query-to-key ranking yourself.
Creator’s lesson / source codeLet’s build GPT, from scratch ↗
A longer optional implementation study; the full video is not required during the three-hour class.
Creator’s lesson / source codeTransformers United: lecture collection ↗
Choose seminars after completing the core notebooks; these research talks are not assumed prerequisites.
Creator’s lesson / source codeTextbook and primary sources
Read the mechanism in detail.
The Week 07 LaTeX chapter follows the same order as the three decks and notebooks: chronological windows, recurrent memory, attention arithmetic, and a causal Transformer.
Colab or local Jupyter
Start with a clean runtime.
- Click Open in Colab and save a copy to your own Drive.
- Use a CPU runtime, then run every cell from top to bottom.
- Read the Markdown explanation and predict the result before each operation.
- Download your evidence JSON and model artifacts before the runtime ends.
For local work: download the .ipynb, install Python and Jupyter, then run the setup cell. The setup installs missing free packages; all core data are generated locally or included with scikit-learn.
Course source and notebook repository ↗
No API keys, paid model calls, GPU allocation, or private datasets are required for the core exercises.