Module 3: Optimization, loss, and regularization

Module 3: Optimization, loss, and regularization#

AINS6003 — Deep Learning & Neural Networks

90-minute lecture deck

Essential question: How do we train deep networks reliably when loss surfaces are non-convex and data are noisy?

Why This Matters#

A modeling team has a prototype that trains inconsistently and needs a defensible training configuration before scaling experiments

For non-CS graduate students, the goal is not to become a software engineer in one class session. The goal is to learn how to read an AI workflow, ask better questions, and explain what the evidence does or does not support.

90-Minute Teaching Arc#

Time

Segment

Purpose

0-10

Orientation and stakes

Connect the topic to a professional decision.

10-25

Conceptual model

Build intuition before code or formulas.

25-40

Worked example

Translate vocabulary into a small concrete case.

40-55

Evidence and interpretation

Read outputs, metrics, or artifacts carefully.

55-70

Guided student activity

Let students change one variable and observe.

70-82

Risk, limits, and communication

Name what could go wrong and how to explain it.

82-90

Assignment handoff

Clarify deliverable, rubric, and next step.

Learning Outcomes#

  • explain the core technical idea in precise neural-network vocabulary

  • connect architecture and training choices to the shape of the data and task

  • run or interpret the module lab as reproducible evidence

  • identify limitations, failure modes, and next steps for a defensible experiment

By the end, students should be able to explain the idea without hiding behind jargon and should know what evidence would make a recommendation stronger.

Plain-Language Framing#

Ask students to complete this sentence before the technical terms arrive:

This method helps a professional decide whether ______ because it uses ______ as evidence.

Then revisit the sentence at the end of the lecture and improve it with module vocabulary.

Core Vocabulary#

  • loss functions as task-specific objectives: define it in one sentence, then connect it to the scenario.

  • stochastic gradient descent and adaptive optimizers: define it in one sentence, then connect it to the scenario.

  • learning rate and batch effects: define it in one sentence, then connect it to the scenario.

  • weight decay, dropout, and early stopping: define it in one sentence, then connect it to the scenario.

  • train, validation, and test separation: define it in one sentence, then connect it to the scenario.

Instructor note: pause after each term and ask for a student-generated example.

Conceptual Model#

Use a three-part model:

  1. Input: What information is available?

  2. Transformation: What does the AI or analytic method do to the information?

  3. Decision: What human or organizational action could change because of the result?

This keeps the discussion accessible for students new to Python.

Worked Example Setup#

Use the professional scenario as the example case. Ask:

  • Who owns the decision?

  • What evidence would they trust?

  • What would count as a bad recommendation?

  • What would a cautious first experiment look like?

Write the answers on the board before opening the lab.

Method Pattern#

  • Hold architecture and data split fixed while comparing optimizer behavior.

  • Plot or tabulate training evidence rather than relying on a single final score.

  • Tune regularization against validation behavior, not training loss alone.

  • State what the toy experiment can and cannot prove about production behavior.

This is the repeatable professional move students should practice across the program.

Lab Bridge#

Lab notebook: Module 3 Lab: SGD vs Adam on a synthetic surface

Use Colab as the recommended first environment. Have students run all cells first, then change exactly one value, threshold, feature, or assumption. The point is observation and interpretation, not typing a lot of code from scratch.

Reading Lab Outputs#

When students see a number, plot, table, or printed result, ask four questions:

  1. What changed?

  2. Is the change large enough to matter?

  3. What assumption produced the result?

  4. What would we need before using this outside the toy setting?

Guided Activity#

Students work in pairs or small groups:

  • Run the lab unchanged.

  • Change one small input or parameter.

  • Capture the before/after result.

  • Write a two-sentence interpretation for a nontechnical stakeholder.

Share two examples with the room.

Common Failure Modes#

  • Selecting the configuration with the lowest training loss despite overfitting.

  • Changing too many variables at once and losing causal interpretation.

  • Using a validation set repeatedly until it becomes a hidden training set.

  • Ignoring randomness and reporting an unstable run as a conclusion.

A strong lecture names these risks before students over-trust the output.

Discussion Checkpoint#

Use these prompts at the 60-minute mark:

  • What did the method make easier to see?

  • What did the method hide or simplify?

  • Who might be harmed by a confident but wrong interpretation?

  • What evidence would make you more comfortable recommending action?

Assignment Handoff#

Module 3 Assignment: Optimizer and regularization comparison

  • Train the same small model with SGD and Adam on a synthetic dataset

  • Compare at least two regularization settings such as weight decay or dropout

  • Summarize loss curves, final accuracy, and run-to-run sensitivity

  • Recommend the next training gate and justify the tradeoff

Students should leave knowing the artifact they are producing, the evidence they must include, and the limitation they must state.

Rubric Translation#

Translate grading into student language:

  • Correct: terms and results are used accurately.

  • Evidence-based: claims point to notebook output, scenario facts, or documented assumptions.

  • Context-aware: the recommendation fits the stakeholder decision.

  • Honest: limitations and risks are named clearly.

Closing Reflection#

Exit prompt:

In one paragraph, explain what this module helps you decide, what evidence the lab produced, and what you would still need before trusting the result in a real organization.

Collect this verbally, in the LMS, or as the opening paragraph of the assignment.

Instructor Timing Notes#

If time runs short, preserve the lab bridge and assignment handoff. Compress vocabulary rather than skipping interpretation. Students new to AI need repeated practice moving from output to meaning.

If time runs long, add a second student share-out focused on limitations and stakeholder communication.

If Students Are New to Python#

  • Explain that a notebook mixes text, code, and output in one page.

  • Run the notebook once before asking students to change anything.

  • Ask for one small change, not open-ended coding.

  • Grade interpretation, evidence, and limitation statements more than syntax fluency.

  • Keep Colab as the first-choice environment unless the activity truly needs the full repository.

  • Give students permission to describe the result first, then refine the vocabulary after discussion.