ml-workflow

Source Metadata

Original cangjie-skill frontmatter from the distillation run:

id: f10
title: "ML项目流程 — 收集→训练→部署→迭代"
book: "AI for Everyone"
author: "Andrew Ng（吴恩达）"
source: p11, p16
tags: [machine-learning, workflow, project-management]
related_skills:
  - slug: ab-mapping
    relation: depends-on
  - slug: train-test-split
    relation: depends-on
  - slug: iterate-not-perfect
    relation: depends-on
  - slug: ml-vs-ds
    relation: contrasts-with

R — Reading (original quote)

"The key steps of a machine learning project are to collect data, train a model, establish an A-to-B mapping relationship, and then deploy the model" — p11

I — Interpretation (restate in your own words)

A machine learning project follows a four-stage workflow, the core of which is alooprather than a linear process:

1. Collect labeled data: Prepare (A, B) paired data—A is the input (such as speech or images), and B is the desired output (such as the text "Alexa" or an object label).
2. Train the model: Let the algorithm learn the A→B mapping relationship from the data.Expect multiple iterations, and the first attempt almost always fails.
3. Deploy the model: Push the trained model to real users and run it in a real environment.
4. Iterate and optimize: After deployment, collect feedback data (especially failure cases), add it to the training data, and retrain.

Key insight: This is aclosed loop—deployment generates new data, new data makes the model better, and a better model generates more usage and data.

A1 — Past Application (past application cases)

• Alexa wake word detection: Collect a large number of "Alexa" recordings → train a wake detection model → deploy to devices → discover that British accents fail to be recognized → collect British accent data → retrain → accuracy improves.
• Autonomous vehicle detection: Collect road images → train a vehicle detection model → deploy to test vehicles → discover that golf carts are not recognized → collect golf cart images → retrain → the model becomes more robust.

A2 — Future Trigger (future trigger conditions)

Trigger scenarios：

• The team has just started an ML project and does not know where to begin
• After training the model, they do not know what to do next
• After deployment, they find the results are poor and want to know how to improve

Language signals：

• "How do you do an ML project?"
• "What are the steps of a machine learning project?"
• "How do you advance an AI project?"
• "What is the project workflow?"
• "What should be done next?"
• "What should be done after deployment?"

Distinguishing point: Use this framework when you are facing a project that requires "building and running a system that can continuously make predictions."

E — Execution (execution steps)

1. Collect labeled training data: Prepare a large number of (A, B) paired samples and ensure data quality.
2. Train the model and iterate: Train the model with the training set, evaluate it on the test set, and repeatedly adjust until it reaches a predetermined metric (such as 95% accuracy).
3. Deploy to real users: Integrate the model into a product/service so that real users can use it.
4. Monitor and collect failure cases: Record the mistakes the model makes in real-world scenarios, and add these failure cases to the training data.
5. Retrain: Retrain the model with the expanded data to continuously improve performance.

Completion criteria: The model has been deployed to real users, and a continuous iteration mechanism of "collect failure cases → retrain" has been established.

B — Boundary (usage boundaries)

Not applicable scenarios：

• Data science projects (the output is a report rather than software, and the process is different)
• One-off analysis tasks (no deployment or iteration needed)
• Situations with no deployment channel (the model cannot reach users)

Failure modes：

• Focusing excessively on the amount of training data while ignoring data quality
• Failing to establish a feedback loop after deployment, so the model cannot continuously improve
• Expecting success with a single training run and giving up when setbacks occur

Author's blind spots：

• Andrew Ng underestimated the difficulty of the "last mile"—integrating a model into an existing system is often harder than training the model itself. The engineering workload in the deployment phase (API design, latency optimization, A/B testing, canary releases) may account for more than 50% of the entire project's workload, but the book says little about this.

Related skills

• ab-mapping (depends-on): The prerequisite for the ML workflow is that clear inputs and outputs have already been defined through the A→B mapping
• train-test-split (depends-on): The training phase requires using a training/test set split to evaluate the model's generalization ability
• iterate-not-perfect (depends-on): After deployment, an iteration mechanism of "collect failure cases → retrain" needs to be established
• ml-vs-ds (contrasts-with): The ML workflow is suitable for "building systems" projects, while DS projects produce reports rather than software