ab-mapping

A→B Mapping — The Core Mental Model of Supervised Learning

R — Original Text (Reading)

"The most common type of machine learning is a type of AI that can learn a mapping from input to output."

— Andrew Ng, p02

I — Methodological Skeleton (Interpretation)

All supervised learning is structurally the same thing: given input A, predict output B. This seemingly simple framework has astonishing unifying power—spam filters, speech recognition, machine translation, ad click-through rate prediction, and visual quality inspection all share exactly the same underlying structure. The primary reason AI projects fail is not insufficient technology, but failure to correctly frame the problem in the form of "A→B". Once you can clearly answer "what is our input" and "what do we want to predict", the problem turns from a vague business demand into an executable machine learning task. The role of this framework is to force you to formalize the problem: distill the chaos of the real world into two clearly defined variables and the mapping relationship between them.

A1 — Applications in the Book (Past Application)

Case 1: Spam Filter

Problem: How to make an email inbox automatically identify spam
Use of the methodology: Input A = the content of an email, output B = whether it is spam (yes/no)
Conclusion: A clear (email→label) pair, supervised learning applies directly
Result: Became one of the most classic and most successful applications of supervised learning

Case 2: Speech Recognition

Problem: How to make machines understand human speech
Use of the methodology: Input A = a segment of audio waveform, output B = the corresponding text
Conclusion: A mapping from audio to text, with a clear structure
Result: The technical foundation of modern voice assistants

Case 3: House Price Prediction (Two Directions)

Problem: How to use AI for real estate valuation
Use of the methodology: Forward A = house area, B = predicted price; reverse A = budget, B = recommended area
Conclusion: The same business problem can be defined with different A→B mappings, and the choice depends on actual data availability
Result: Shows that the A→B framework needs to be defined flexibly in light of data reality

Case 4: Vehicle Localization in Autonomous Driving

Problem: How an autonomous vehicle knows the positions of surrounding vehicles
Use of the methodology: Input A = images captured by cameras, output B = the position coordinates of surrounding vehicles
Conclusion: A mapping from images to coordinates, with a clear structure
Result: A core module of autonomous driving perception systems

A2 — Trigger Scenarios (Future Trigger) ★

In what situations will users need this skill?

The business team proposes a vague idea of "wanting to use AI" and needs to turn it into a concrete technical problem.
The technical team is assessing whether a requirement falls within the scope of supervised learning.
A product manager is writing an AI requirements document and needs to clearly define the inputs and outputs.
An entrepreneur is conceiving an AI product and needs to verify whether the problem structure holds up.

Language signals (activate when these appear in the user's words)

"Can AI do this?"
"Does this count as a machine learning problem?"
"Is our scenario suitable for using AI?"
"Can AI be used to predict X?"
"How do we turn this problem into an AI problem?"

Distinction from adjacent skills

And one-second-rule difference: the one-second rule judges "whether this thing is technically feasible," while A→B mapping formalizes the problem into an executable structure. Map first, then judge feasibility.
And ml-feasibility difference: the two-factor framework is a complete feasibility assessment (concept + data), while A→B mapping is only the first step—framing the problem as an input-output pair.
And automate-task difference: task decomposition is "breaking a job into tasks," while A→B mapping is "framing a task as input-output."

E — Executable Steps (Execution)

After the skill is activated, the agent should execute the following steps:

Define input A
- Ask the user: "What data do you have? What is the input to this AI system?"
- Completion criterion: The user can describe the input in one sentence (e.g., "an email," "a product image," "a customer conversation")
Define output B
- Ask the user: "What do you want AI to give you? Is it a category, a number, or a piece of text?"
- Completion criterion: The user can describe the expected output in one sentence (e.g., "whether it is spam," "defect type," "predicted price")
Verify the availability of (A, B) labeled pairs
- Ask the user: "How many already labeled (input, output) samples do you have on hand?"
- Completion criterion: The user can give a rough order of magnitude for the data (e.g., "1,000 labeled emails," "no historical labeled data")
- Stop condition: If the user cannot define a clear A or B, then prompt "this problem may not yet be suitable for supervised learning; it is recommended to first clarify the problem"

B — Boundary (Boundary) ★

Do not use this skill in the following situations

The problem has no clear input and output (e.g., "I want to understand customer behavior patterns"—this is data science/exploratory analysis, not supervised learning)
The problem requires complex long-horizon reasoning (e.g., "formulate next year's marketing strategy"—a single output B cannot be defined)
The problem falls within the scope of unsupervised learning (e.g., clustering, dimensionality reduction, anomaly detection—there is no clear B)
The user only wants to understand the data and does not need to make predictions

Failure modes the author warns about in the book

Framing the wrong A or B for the problem, causing all subsequent work to be built on the wrong formalization
Confusing "prediction" with "recommendation"—the A→B definition of a recommendation system is more complex than that of supervised learning

The author's blind spots / limitations of the era

The A→B framework assumes you can clearly define "what the correct output B is"—in reality, many business problems are vague (such as "good customer service" or "creative advertising") and cannot be simply labeled
This framework is mainly applicable to supervised learning, and its applicability to reinforcement learning and generative AI is limited
The book does not fully discuss: when B itself requires subjective human judgment (such as "is it beautiful" or "is it good"), annotation consistency becomes a bottleneck

Easily confused neighboring methodologies

Difference from the "data science process": data science does not necessarily require A→B; it focuses more on discovering insights from data
Difference from "recommendation systems": the input-output structure of recommendation systems is more complex and is not entirely equivalent to a simple A→B mapping

Related skills

ml-feasibility (depends-on): after defining the A→B mapping problem, the two-factor framework is needed to evaluate feasibility
one-second-rule (composes-with): the one-second rule helps determine whether B is "simple" enough to be learned
ml-workflow (depends-on): after defining A and B, proceed to the four-step ML process for execution

Audit information

Verification passed: V1 ✓ / V2 ✓ / V3 ✓
Test pass rate: pending testing
Distillation time: 2026-06-29