DAY 04 · MODULE 4 · THE FINALE

From data to deployed. 🔄

The last module. Build a model end-to-end — data → development → serving — plus MLOps, and a peek at how a neural network actually learns.

Module 4 — 6 Chapters 🗺️

The official AI Development Workflow course. The whole journey, from raw data to a model in production.

🍽️

1 · The ML Workflow

3 stages overview

🥕

2 · Data Preparation

Upload + features

👨‍🍳

3 · Model Development

Train + evaluate

🚀

4 · Model Serving

Deploy + monitor

🔄

5 · MLOps

Automate the workflow

🧠

6 · How ML Learns

Inside a neural net

One analogy ties it together — building an ML model is like running a restaurant. You'll see. Jom.

Chapter 1 · The ML Workflow

Building ML = Running a Restaurant 🍽️

Three stages. You start with raw ingredients and finish by serving the dish.

STAGE 1

🥕 Data Preparation

Prep the raw ingredients

→

STAGE 2

👨‍🍳 Model Development

Experiment with recipes

→

STAGE 3

🍽️ Model Serving

Serve the meal

Why serving mattersA model that never reaches production is just a theoretical model — no use. You have to serve the meal.

It's a Loop, Not a Line 🔁

The workflow is iterative. You go back and forth — e.g. while training you might return to the data for better features; while monitoring you might catch drift and adjust.

🖱️ AutoML

no-code path

Build the whole workflow through the UI
Little ML expertise, no coding

⌨️ Code path

Vertex AI Pipelines

Workbench / Colab + pre-built SDKs
Automate the workflow programmatically

Good news — the iterative steps can be automated with MLOps. More on that in Chapter 5.

Chapter 2 · Data Preparation

Stage 1 — Prep the Ingredients 🥕

Two steps: upload the data, then engineer the features. Data can come from Cloud Storage, BigQuery, or your local machine.

AutoML tabular objectives

📈

Regression

Predict a number

🏷️

Classification

Predict a category

📅

Forecasting

Predict over time

Quality & quantity matter: a model needs a lot of data, and how good & how much you have decides how well it learns.

Feature Engineering 🔪

Before cooking, you peel the carrots and chop the onions. Same with data — you process it before the model trains.

What's a feature?A factor that contributes to the prediction — like an independent variable in statistics, or a column in a table.

Preparing features is challenging & tedious — so Vertex AI gives you a Feature Store to make it easier. Next slide.

Vertex AI Feature Store 🗄️

A central repository to manage, serve & share features — for real-time (online) and batch (offline) serving. Even ready for Gen AI embeddings.

🤝

Shareable

One central source — consistency across the org.

♻️

Reusable

Save time, no duplicated effort.

📈

Scalable

Auto-scales for low-latency serving.

🖱️

Easy to use

Friendly UI.

Remember the four: shareable, reusable, scalable, easy.

Chapter 3 · Model Development

Stage 2 — Cook & Taste 👨‍🍳

Two steps that loop: train the model (cook the recipe), then evaluate it (taste the meal). Repeat.

To set up training you specify…

📂

Method

Your dataset · AutoML or custom

🎯

Objective

The task — e.g. classification

⚙️

Details

Target column, features, budget

AutoML then trains and picks the best model from thousands — thanks to neural architecture search & transfer learning (from Module 3).

Evaluate — The Confusion Matrix 🔢

Did the model "predict cat" correctly? Four outcomes. (✓ = the model was right.)

Actually Cat

Actually Not

Predicted
Cat

✅ True Positivesaid cat · is cat

❌ False Positivesaid cat · isn't · Type I

Predicted
Not

❌ False Negativesaid not · is · Type II

✅ True Negativesaid not · isn't

The four cells are the foundation for almost every classification metric — including precision & recall, next.

Try It · Chapter 3

Tune the Spam Threshold ⚖️

🕹️🔴 real spam, 🟢 real email. Slide the line — everything to its right is flagged as spam. Goal: watch precision & recall trade off. Low = catch all (recall ↑), high = only the obvious (precision ↑). Both extremes → +5⭐.

← less spammymore spammy →

🎯 Recall

—

🔍 Precision

—

Recall = of all real spam, how many caught (TP÷(TP+FN)). Precision = of all flagged, how many right (TP÷(TP+FP)). Push one up, the other drops — a trade-off.

Try It · Chapter 3

Sort the Predictions 🔢

🕹️The model is detecting cats. Drag each outcome into its cell. Goal: read a confusion matrix — “true” = the model was right, “positive” = it said yes. Correct → +5⭐.

Said "cat" — and it IS a cat

Said "not cat" — and it ISN'T

Said "cat" — but it isn't (Type I)

Said "not cat" — but it is (Type II)

✅ True Positive

✅ True Negative

❌ False Positive

❌ False Negative

Try it yourself first, ya. "True" = the model was right; "Positive" = it said yes.

Feature Importance & Explainable AI 📊

Vertex AI shows a bar chart of feature importance — how much each feature contributes to a prediction. Longer bar = more important.

Example — predicting customer churn

Contract length

0.94

Monthly charges

0.71

Tenure

0.55

Tech support

0.33

Customer gender

0.05

Contract length drives the prediction most; gender barely matters — a feature you could drop. This is part of Explainable AI — tools to understand & interpret what a model is doing.

Chapter 4 · Model Serving

Stage 3 — Serve the Meal 🍽️

Two steps: deploy the model (serve the dish), then monitor it (check it's running well). Model management runs across the whole workflow.

⚡

Online

Deploy to an endpoint · real-time, low latency · instant recommendations

📦

Batch

No endpoint needed · many rows at once · bi-weekly campaigns

📱

Edge

Off-cloud, on a device · low latency, privacy, offline · factory camera

Online needs an endpoint; batch doesn't. Monitoring is automated by Vertex AI Pipelines — it warns you when something crosses a threshold.

Try It · Chapter 4

Online, Batch, or Edge? ⚡📦📱

🕹️Drag each use case onto its serving mode. Goal: match by need — real-time (online), scheduled bulk (batch), on-device (edge). Correct → +5⭐.

Instant product recommendation as a user browses

Marketing campaign sent every two weeks

Object detection on a factory camera, no cloud

Real-time fraud check on a card swipe

Monthly churn scoring of all customers

An app that must work offline on a device

⚡ Online

📦 Batch

📱 Edge

Real-time → online. Scheduled bulk → batch. On-device / offline → edge. Try it first, ya.

Chapter 5 · MLOps & Automation

MLOps — DevOps for ML 🔄

MLOps applies DevOps principles to ML. It solves the production pain: in ML, both data and code keep changing.

The goal: CI · CT · CDContinuous Integration, continuous Training, continuous Delivery — by automating & monitoring every step.

The backbone: Vertex AI Pipelines — supports Kubeflow Pipelines (KFP) and TensorFlow Extended (TFX). Already use TensorFlow on big data? Use TFX. Otherwise, KFP.

Pipelines & Components 🧩

A pipeline is built from components — each a self-contained piece of code that does one task. Think of a component like a function.

🧱 Pre-built components

reuse Google's

e.g. create dataset, train AutoML, create endpoint, deploy model
Check these first before writing your own

✍️ Custom components

your own logic

For special tasks — e.g. a deploy threshold check
Each component: one single responsibility

Build the pipeline → compile it → run it. Vertex AI even has templates so you don't start from scratch.

3 Phases of MLOps Maturity 📈

Phase 2 — Integrate connect components · achieve CI / CT / CD
Phase 1 — Build components start automating with Pipelines SDKs
Phase 0 — Manual GUI / AutoML · no MLOps yet

Phase 0 is not a failureIt's essential — you build the workflow manually first, so you understand it before you automate it.

Chapter 6 · How ML Learns · optional

Inside a Neural Network 🧠

Optional deep-dive. Every network — DNN, CNN, RNN, even LLMs — stems from the basic Artificial Neural Network (ANN). Three layers:

Output layer the prediction ŷ
Hidden layer neurons · the "thinking"
Input layer your features go in

Each neuron is a node; the lines between them carry weights — the information the network learns during training.

How It Learns — The Loop 🔁

One pass of this whole loop = one epoch. Repeat until the error stops shrinking.

1–2

Weighted Sum + Activation

Combine inputs × weights, add non-linearity

→

3–4

Predict ŷ

Output a guess

→

5

Loss

Measure ŷ vs actual y

6

Backpropagation

Find which weights to blame

→

6

Gradient Descent

Nudge weights downhill

↩

7

Repeat (Epoch)

Until loss bottoms out

Same shape as Module 1's "shrink the loss" — now you've seen the machinery underneath.

Activation Functions — Tap 🔀

🕹️Tap each card to open it — ReLU, Sigmoid, Tanh, Softmax. Goal: know which fires where (binary → sigmoid, multi-class → softmax). They add the non-linearity that makes deep layers worth stacking. (Explore — no points.)

📈 ReLU

tap to expand

Negative → 0, positive → unchanged. The common workhorse for hidden layers.

〰️ Sigmoid

tap to expand

Squashes to 0–1. Used for binary classification & logistic regression.

🌊 Tanh

tap to expand

A shifted sigmoid — outputs between −1 and +1.

🎯 Softmax

tap to expand

For multi-class classification — outputs a probability distribution that adds up to 1.

Binary → sigmoid. Multi-class → softmax. You can mix — e.g. ReLU on hidden layers, softmax on the output.

Loss, Gradient Descent & the Knobs 🎚️

Loss vs cost: loss = error for one example; cost = error over the whole training set. MSE for regression, cross-entropy for classification.

🕹️🎮 Set a learning rate, then tap Step — the ball rolls downhill toward lowest loss. Goal: find a rate that converges. Too small → it crawls; too big → it overshoots & diverges. Reach the bottom → +5⭐.

Learning rate 0.30

—

🔧 Parameters

learned by the machine

Weights & biases
You only set the starting values

🎛️ Hyperparameters

set by a human

Layers, neurons, activation, learning rate, epochs
AutoML can pick these for you

Knowledge Check · Module 4

Lock It In 🧪

Q1The three ML workflow stages, in order:

Data prep → Model development → Model serving

Serving → Training → Data

Model dev → Data prep → Serving

Deploy → Monitor → Train

Q2To automate the workflow with code, you use:

A spreadsheet

Gmail

A pre-trained API

Vertex AI Pipelines

Q3Data for ML can come from:

Only BigQuery

Only local files

Only Cloud Storage

Cloud Storage, BigQuery, or local

Q4To catch as MANY spam emails as possible, optimise for:

Recall

Precision

Latency

Budget

Q5The two steps of model serving are:

Train & evaluate

Deploy & monitor

Upload & engineer

Compile & run

Q6The backbone of MLOps on Vertex AI is:

BigQuery

A pre-trained API

Vertex AI Pipelines

A spreadsheet

Q7Why do we need activation functions?

To add non-linearity

To store data

To make it slower

To delete weights

Q8One full pass of the learning loop is called:

An epoch

A neuron

A weight

A bias

Module 4 — You Can Now Explain… ✅

3-Stage Workflow Iterative Feature Store Confusion Matrix

Precision vs Recall Explainable AI Online / Batch / Edge MLOps · CI/CT/CD

Vertex AI Pipelines Components · 3 Phases Activation Functions Gradient Descent

Finish Module 4 in 2 stepsDo the hands-on lab first — then take the graded quiz.

🧪 Step 1 · Lab — Build an ML Model with AutoML on Vertex AI

Walk the three stages end-to-end; tune precision & recall, read feature importance. Open the lab →

📝 Step 2 · Official Module 4 Quiz →

After the lab, take the graded quiz on Skills Boost. skills.google › course 593 › quiz 617916

Idea → data → model → production. You can build it. 🔄

DAY 4 COMPLETE 🎉

You built the whole thing. 🚀

Foundations → Generative AI → Development Options → the Workflow. Four modules, the full picture of AI & ML on Google Cloud. Rest tonight, hydrate — you've done the hard part.

🧱

M1 + M2

Foundations & Gen AI

🛤️

M3 + M4

Options & Workflow

🎉

Wrap-Up →

Recap & next steps