Frequently Asked Questions About Ng Machine Learning Orientation Framework
21 answers covering everything from basics to advanced usage.
// Basics
What is machine learning in the simplest terms?
Machine learning is the science of getting computers to learn without being explicitly programmed. Instead of a human writing every rule by hand, you feed the system examples of inputs and correct outputs, and it infers the underlying pattern. This lets it handle new cases it has never seen, which hand-coded rules struggle to do.
What is the difference between machine learning and traditional programming?
In traditional programming, a developer encodes every rule and decision path by hand—explicit programming. In machine learning, the system learns patterns from data instead. You use traditional programming when rules are known and finite, and ML when the rules are too complex, unknown, or the pattern lives in data rather than human knowledge.
Why is 'learning without explicit programming' the key idea in ML?
Because it defines when ML earns its place. Many problems—image recognition, fraud detection, speech—involve patterns no human can fully write out as rules. ML shines exactly where explicit programming breaks down. Keeping this definition front of mind stops you from reaching for ML on problems deterministic logic already solves cleanly.
What are the main categories of ML applications?
The core patterns are ranking/retrieval (web search), classification/labeling (photo tagging), recommendation (streaming), speech and language understanding (voice-to-text), anomaly/spam detection (email filtering), and industrial optimization (wind turbine power output). Mapping your problem to one of these categories narrows which method and data you'll need.
// How To
How do I apply the framework to a fraud detection problem?
State it plainly: detect fraudulent transactions in real time. Check rules: fraudsters adapt, so hand-written rules can't keep up. Map the pattern: this is anomaly/spam detection, like email filtering. Find the signal: historical transactions labeled fraudulent or legitimate. Classify: enterprise/industrial ML. Justify: fraud patterns are too dynamic and numerous for hand-written rules.
How do I identify the learning signal for my problem?
Ask what the inputs are and what correct output the system should learn to produce, then check whether you actually have examples of those input-output pairs or feedback data. For medical triage, inputs are scans and outputs are confirmed diagnoses. Surface this early—if no labeled or feedback data exists, ML cannot proceed regardless of the algorithm.
How do I justify ML over explicit programming to my team?
State one or two sentences anchored to your specific problem: the rules are too complex to write by hand, OR the pattern lives in data not human knowledge, OR the system must generalize to unseen cases. Tie this directly to concrete pain—like 'fraud patterns change weekly, so static rules go stale.' Specificity makes the justification credible.
How do I state my problem in plain language before analyzing it?
Write one sentence describing the decision, prediction, or action that needs to happen, avoiding all technical jargon. For example: 'I need to identify defective items coming off an assembly line.' Plain phrasing keeps you focused on the actual need and prevents ML jargon from obscuring whether ML is truly required.
// Troubleshooting
What if I'm not sure whether my problem has enough data?
Treat data availability as a gap to surface, not assume. Ask whether you already have historical examples with known outcomes, or whether you can realistically collect them. If neither exists, ML is blocked until you build a data pipeline. It's far cheaper to discover this at the orientation stage than after weeks of modeling.
What if my problem seems to fit multiple ML patterns?
That's common—many real problems blend patterns. Medical image triage is classification but also prioritization (ranking). Pick the primary decision the system must make and map to that pattern first; secondary patterns can layer in later. The goal at orientation is to identify the dominant method, not to design the full system.
Why did my ML project fail even though the problem seemed like a good fit?
The most common cause is a missing or weak learning signal—teams assume data exists without verifying labeled or feedback data is actually available. Other causes: the problem was better solved by explicit rules, or a consumer-ML mindset was applied to a high-stakes industrial problem with different error tolerance. Re-run the framework to spot which step was skipped.
What if the rules for my problem are complex but still writable by hand?
If you can genuinely enumerate every rule, even if there are many, explicit programming may still be the better choice—it's deterministic, debuggable, and needs no data. ML becomes preferable when rules are not just numerous but unknown, constantly changing, or hidden in data. The signal for ML is when hand-coding truly breaks down, not merely gets tedious.
// Comparisons
How does this framework compare to a generic 'build an MVP model' approach?
A generic build-first approach risks committing engineering effort before confirming ML even fits. This framework validates upfront: it catches problems solvable by rules, flags missing data, and clarifies risk level. It's a cheap filter that saves expensive missteps, whereas building an MVP first discovers those same problems only after you've invested weeks.
How does the Ng orientation approach differ from a data-science checklist?
Most data-science checklists assume you've already decided to use ML and focus on modeling steps. This framework operates one level earlier—it decides whether ML is the right approach at all. It centers on the definition 'learning without explicit programming' and the learning-signal check, making it a problem-framing tool rather than a modeling workflow.
How does consumer ML differ from industrial ML in practice?
Consumer ML (search, tagging, recommendation) tolerates occasional errors and works at massive scale with abundant behavioral data. Industrial ML (diagnostics, defect detection, energy optimization) demands high accuracy, faces regulatory and safety stakes, and often has scarcer labeled data. Applying a consumer mindset to an industrial problem underestimates the accuracy and data rigor required.
Is machine learning better than rule-based systems?
Neither is universally better—they solve different problems. Rule-based systems win when logic is known, finite, and needs to be transparent and deterministic. ML wins when patterns are too complex to hand-code, change over time, or must generalize to unseen cases. The framework's core purpose is to help you choose correctly rather than defaulting to either.
// Advanced
How does the consumer-to-industrial progression affect my ML strategy?
ML value has moved from consumer applications toward big-company and industrial ones—climate and energy optimization, healthcare diagnostics, manufacturing inspection. Placing your problem on this spectrum tells you what scale, data volume, and error tolerance to plan for. Industrial problems typically need more accuracy and carry higher deployment risk than consumer ones.
How do I evaluate whether an industrial ML problem is worth pursuing?
Beyond confirming a learning signal, assess error tolerance and stakes. Industrial problems like medical diagnosis or defect detection require high accuracy and often regulatory scrutiny, so weigh whether achievable model performance clears the safety bar. Also confirm data volume matches the complexity—industrial patterns often demand more and cleaner labeled examples than consumer applications.
Can I use this framework to spot ML opportunities competitors are missing?
Yes. The 'ubiquitous invisible ML' principle trains you to recognize where ML already operates unseen, then transfer that pattern to new contexts in your domain. Scan your workflows for tasks where rules keep breaking, data on outcomes exists, and the pattern lives in data—those are candidate opportunities others may not have framed as ML yet.
What's the biggest misconception about machine learning this framework corrects?
That ML is magic. The definition 'learning without being explicitly programmed' still requires data, a learning signal, and iteration—not just an algorithm. The framework grounds you in what ML actually needs, so you plan for data collection and realistic performance rather than expecting an algorithm to solve a problem you can't supply examples for.
How does computer vision fit into this framework?
Computer vision is an ML subfield that lets machines interpret visual data, used in factory inspection, photo tagging, and medical imaging. When you map a problem involving images to a pattern—like classifying defective parts or flagging abnormal scans—you're identifying a computer vision task. The framework still applies: confirm rules can't be hand-coded and that labeled image data exists.