Frequently Asked Questions About Ng Machine Learning Orientation Framework
22 answers covering everything from basics to advanced usage.
// Basics
What is the difference between machine learning and explicit programming?
Explicit programming requires a human to write every rule and decision path by hand. Machine learning replaces this by having the system infer patterns from data. If you can write every rule exhaustively, explicit programming works. When rules are too complex, too numerous, or change over time — like in fraud detection or image recognition — machine learning is the better approach because the system adapts from examples rather than static code.
What does Andrew Ng mean by learning without being explicitly programmed?
Andrew Ng defines machine learning as the science of getting computers to learn without being explicitly programmed. This means instead of a developer writing if-then rules for every scenario, the machine is given data — examples of inputs and correct outputs — and discovers the patterns itself. The phrase comes from Arthur Samuel's original 1959 definition and is the foundational idea that separates ML from traditional software development.
What are the known ML application patterns in the framework?
The framework identifies six core patterns: ranking and retrieval (like web search), classification and labeling (like photo tagging), recommendation (like streaming services), speech and language understanding (like voice-to-text), anomaly and spam detection (like email filtering), and optimization in industrial settings (like wind turbine power generation). Mapping your problem to one of these patterns narrows down the ML approach and methods you will need.
What is the consumer-to-industrial ML spectrum?
It is the framework's classification of ML problems along a continuum. Consumer applications — like search ranking, photo tagging, content recommendation — typically involve massive datasets, tolerate some errors, and serve individual users. Industrial applications — like medical diagnostics, manufacturing defect detection, and energy optimization — often involve smaller specialized datasets, demand high accuracy, carry regulatory requirements, and have higher consequences for errors. The spectrum affects every downstream decision.
What are examples of ML already being used without people realizing it?
Web search ranking, social media photo tagging, Netflix and Spotify content recommendations, smartphone voice-to-text, and email spam filtering are all machine learning systems most people use daily without awareness. The framework's second principle — Ubiquitous Invisible ML — uses these examples to train your eye. Once you recognize ML in everyday tools, you can more easily spot where similar patterns apply to your own domain and problems.
// How To
How do I write a good plain-language problem statement for ML?
Write one sentence describing what decision, prediction, or action needs to happen, avoiding all technical jargon. Focus on what you want the outcome to be. For example, 'I need to identify defective items coming off an assembly line' or 'I need to predict which customers will cancel their subscription next month.' A clear problem statement prevents the common mistake of jumping to solutions before understanding the actual need.
How do I check if my problem has a valid learning signal?
Ask two questions: What are the inputs the system would receive? And what is the correct output it should learn to produce? Then verify that you have — or can collect — examples of these input-output pairs. For fraud detection, that means labeled transactions. For medical imaging, that means scans with confirmed diagnoses. If you cannot point to real data with real labels or feedback, ML cannot be applied yet and you need to address the data gap first.
How do I map my business problem to a known ML pattern?
Compare your problem's structure to the six core patterns. If you are sorting or ranking items, it maps to ranking/retrieval. If you are assigning categories or labels, it is classification. If you are suggesting items to users, it is recommendation. If you are processing speech or text, it is language understanding. If you are finding outliers, it is anomaly detection. If you are improving a measurable process, it is optimization. Pick the closest match.
How do I justify ML adoption to non-technical stakeholders?
Use step six of the framework: state in one or two sentences why ML is needed over explicit programming. Anchor it to the specific problem. For example: 'Fraud patterns change faster than we can write rules, so the system needs to learn from new data automatically.' Reference everyday ML they already use — spam filtering, search rankings, voice assistants — to make the concept tangible. Avoid jargon and focus on business outcomes.
How long does it take to go through the full framework?
For a single well-understood problem, the six steps take 15 to 30 minutes. If you need to research data availability or consult domain experts, it may stretch to a few hours spread across sessions. The framework is intentionally lightweight — it is a decision tool, not a project plan. The output is a concise problem statement, pattern mapping, data assessment, and one-to-two-sentence ML justification. Speed is a feature, not a limitation.
// Troubleshooting
What if my problem could be solved by both rules and ML?
If explicit rules fully and reliably solve the problem, ML adds unnecessary complexity. The framework explicitly warns against this. However, if rules work now but the problem is scaling, evolving, or you expect edge cases to multiply, ML may be warranted as a future-proof investment. The key question is whether the rules will break down as complexity grows. If they will, start building the ML pipeline early.
What if I do not have labeled data for my ML problem?
The framework's fourth step surfaces this gap intentionally. Without a learning signal — labeled examples or feedback data — ML cannot be applied effectively. Your options include: collecting and labeling data manually, using semi-supervised or unsupervised approaches if applicable, leveraging pre-trained models or transfer learning, or generating labels through domain-expert annotation. Address the data gap before committing to an ML project.
What happens if I misclassify my problem on the consumer-to-industrial spectrum?
Misclassification leads to mismatched expectations for data volume, error tolerance, and deployment complexity. Consumer ML (recommendations, content) tolerates some errors and relies on massive user data. Industrial ML (medical diagnosis, defect detection) demands high accuracy with potentially smaller datasets and has regulatory or safety implications. Getting the spectrum wrong means underestimating risk, over- or under-investing in data quality, and building the wrong infrastructure.
My team keeps describing problems using ML jargon — how do I fix that?
The framework's first step and one of its explicit pitfall warnings address this. Require every ML discussion to start with a plain-language problem statement — no algorithms, no model names, no acronyms. If someone cannot describe the problem in a sentence a non-technical person would understand, the problem is not yet well-defined. This discipline prevents teams from jumping to solutions before confirming ML is even appropriate.
What if my problem does not fit any of the six known ML patterns?
Most real-world problems do map to one of the six patterns when stated clearly. If yours genuinely does not, it may be a novel research problem, a multi-pattern hybrid (like combining classification with optimization), or it may not actually be an ML problem. Revisit step one to ensure your problem statement is specific enough. If after clarification it still does not fit, consult ML literature for emerging application categories or consider whether traditional methods suffice.
// Comparisons
How does this framework compare to CRISP-DM or other ML methodologies?
CRISP-DM is a full lifecycle methodology covering business understanding through deployment and monitoring. The Ng ML Orientation Framework operates upstream — it helps you decide whether ML is the right approach before you enter any lifecycle model. Think of it as step zero. Once the framework confirms ML applicability and you have a clear problem statement, pattern mapping, and learning signal, you can then feed that into CRISP-DM, TDSP, or any other ML project methodology.
How is this different from just asking ChatGPT if my problem needs ML?
Asking an AI assistant gives you a one-time answer without structured reasoning. The Ng ML Orientation Framework gives you a repeatable, six-step process that forces you to examine your problem from multiple angles — explicit programming sufficiency, pattern mapping, data availability, and consumer-vs-industrial classification. It builds your own judgment rather than outsourcing it. The framework also produces documentation — a problem statement and justification — that you can share with stakeholders.
// Advanced
Is this framework only for beginners or can experienced ML engineers use it too?
Experienced ML engineers benefit from the framework when scoping new projects, onboarding cross-functional teammates, or writing project proposals. The structured output — plain-language problem statement, pattern mapping, learning signal identification, and ML justification — is valuable for communicating with product managers, executives, and domain experts who do not have ML backgrounds. It also prevents the experienced-engineer pitfall of assuming ML is needed for every problem.
Can I use this framework for deep learning and generative AI problems too?
Yes. Deep learning and generative AI are subsets of machine learning. The framework's pattern mapping step extends naturally: generative tasks (text generation, image synthesis) map to language understanding and computer vision patterns. The learning signal check still applies — generative models require massive training data. The consumer-to-industrial classification is especially important for generative AI since consumer chatbots and industrial document automation have vastly different risk profiles and data needs.
How do I adapt this framework for a team-based ML evaluation process?
Run the six steps as a structured workshop. Have each team member independently write a plain-language problem statement, then converge. Assign the explicit-programming check to the engineering lead. Have a domain expert verify the learning signal. Use the pattern mapping as a group exercise. The consumer-to-industrial classification involves product and compliance stakeholders. The final justification becomes a shared document. This turns individual framework use into a collaborative decision artifact.
Can this framework tell me which ML algorithm to use?
No, and that is by design. The framework operates upstream of algorithm selection. Its purpose is to confirm ML applicability, identify the problem pattern, and verify data readiness. Once you have completed the six steps and confirmed ML is appropriate, you move into model selection using the pattern mapping as a guide — classification problems suggest different algorithms than recommendation or anomaly detection problems. The framework sets the stage; algorithm selection follows.
How does this framework handle problems where ML might cause harm?
The consumer-to-industrial classification in step five implicitly raises stakes and risk tolerance. Industrial and healthcare applications demand higher accuracy and have ethical and regulatory implications. However, the framework does not include an explicit ethics or bias audit step. For high-stakes domains, supplement the framework with a fairness and impact assessment after confirming ML applicability. The framework gets you to the right question — ethics review answers the next one.