The following terms circulate in the press, in vendor material and on social media as if they were interchangeable. They are not.
Artificial Intelligence · Machine Learning · Deep Learning
Data Science · Big Data · Algorithm · Neural NetworkThe confusion is not merely cosmetic. It has three practical consequences.
It corrupts scoping. A project described as "an AI project" commits nothing. A project described as "a supervised binary classification model trained on eighteen months of labeled transactions" commits to a data requirement, an evaluation protocol and a failure mode. The first phrasing survives a steering committee; the second survives contact with production.
It corrupts expectations. A stakeholder who believes that Machine Learning requires Big Data will refuse a viable project on 4,000 observations. A stakeholder who believes Deep Learning is uniformly superior will fund a six-week neural network for a problem that gradient boosting solves in four minutes.
It corrupts assessment. In a technical interview, the request "define Machine Learning" is not a memory test. It measures whether the candidate distinguishes an objective from an approach, and an approach from a class of models. Candidates who conflate the three rarely recover.
Until the taxonomy is fixed, everything downstream stays blurred. This chapter fixes it once, and the rest of the course relies on it without restating it.
Strict inclusion:
Deep Learning ⊂ Machine Learning ⊂ Artificial Intelligence
Every Deep Learning technique is Machine Learning. Every Machine Learning technique falls under Artificial Intelligence. The converses are false.
Read the inclusion in both directions, because each direction carries a different piece of information.
Downward, inclusion tells you what is inherited. A convolutional network is a Machine Learning method, so everything the course will say about training data, generalization, overfitting and evaluation applies to it without amendment. There is no separate theory of evaluation for deep models.
Upward, inclusion tells you what is not implied. A system that qualifies as Artificial Intelligence need not learn anything at all. This is the single most common error in the field, and section 3.1 is devoted to it.
One caution before going further. The three terms are not on the same conceptual footing, which is why they cannot be swapped:
| Term | What kind of thing it is | The question it answers |
|---|---|---|
| Artificial Intelligence | An objective | What are we trying to get a machine to do? |
| Machine Learning | An approach | How do we obtain the decision rule? |
| Deep Learning | A class of models | What functional form does the model take? |
A statement that mixes these levels is malformed. "We chose AI over Machine Learning" is malformed in the same way as "we chose transportation over the bicycle."
Rigorous definition
The field of computer science concerned with building systems capable of performing tasks that would be described as cognitive if performed by a human being: perception, reasoning, learning, planning, decision making, language understanding.
Operational definition
An AI system is a system that, from inputs received from its environment, produces outputs — predictions, recommendations, decisions, actions — which influence that environment, according to objectives defined explicitly or implicitly by a human.
This second formulation is close to the definition adopted by the OECD and carried into several regulatory frameworks. It is the more useful one in a professional setting because it is verifiable: given a system, you can check each clause against it.
In plain terms
Getting a machine to carry out a task that would require a human to think.
Nature of the concept
AI is an objective, not a technique. It covers heterogeneous approaches, several of which involve no learning whatsoever.
Point of caution
The word "intelligence" invites a comparison with human cognition that the technical content does not support. A system qualifies as AI by what it does, not by any resemblance to how a person does it. An itinerary planner that computes an optimal route is AI; it has no notion of a city.
Dating
The term was coined in 1956 at the Dartmouth conference. The field therefore predates modern applied Machine Learning by several decades — a fact worth holding onto, since it immediately makes "AI equals Machine Learning" untenable.
This is the most frequent misreading in the entire field, and it survives well into professional practice.
| Technique | Principle | Learning |
|---|---|---|
| Rule-based system | A human expert formalizes the rules | No |
| Tree search | Exhaustive or heuristic exploration of the state space | No |
| Formal logic | Inference from axioms and deduction rules | No |
| Combinatorial optimization | Search for an optimum under constraints | No |
| Machine Learning | Induction of rules from observations | Yes |
The discriminating question is not "is the system sophisticated?" nor "does the system make decisions?" It is narrower and entirely mechanical:
Where did the decision rule come from — was it written by a person, or was it derived from observed data?
If a person wrote it, the system does not learn, however elaborate it may be.
Example of AI without learning — a programmable thermostat
The system perceives its environment and decides. No rule was induced from data: an engineer wrote the thresholds. Feed it ten years of temperature readings and its behavior will be identical to what it was on the first day. That invariance is the signature of a non-learning system.
Example of AI without learning — Deep Blue (1997)
The system that beat Garry Kasparov evaluated roughly 200 million positions per second using an evaluation function designed by engineers working with grandmasters. It did not learn from the games it played. It searched a state space, and it searched it faster than any human can.
The contrast with a modern chess engine trained by self-play is instructive. Both play chess at superhuman level. Both are AI. Only one of them derived its evaluation of a position from data rather than from human specification. The observable output is similar; the epistemic status of the rule is completely different, and so is what happens when the environment changes.
Three forces keep the conflation alive, and knowing them helps you correct it without sounding pedantic.
Commercial vocabulary. "AI" is the term that sells. A vendor has no incentive to say "we ship a rules engine with 340 hand-written conditions" when "AI-powered" is available and, strictly speaking, not false.
Recency bias. The visible successes of the last fifteen years happen to be learning-based, which makes the learning-based subset feel like the whole.
The reverse error, less often noticed. Because "AI" now connotes learning, practitioners sometimes hesitate to call a deterministic rules engine an AI system at all — and then fail to recognize that it is the correct solution to the problem in front of them. Section 4.4 returns to this.
Rigorous definition (Tom Mitchell, 1997)
A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E.
This is the standard academic reference. Its merit is that it is operational rather than descriptive: it forces you to name three components explicitly, and a project that cannot name all three is not a Machine Learning project yet.
Applying the definition to a concrete case
Spam detection:
Alternative definition (Arthur Samuel, 1959)
The field of study that gives computers the ability to learn without being explicitly programmed for the task at hand.
In plain terms
Instead of writing the rules, you supply a large number of examples and the algorithm induces the rules itself.
Nature of the concept
Machine Learning is an approach within AI. Its central concept is induction: moving from observed particular cases to a general rule.
Point of caution
Induction is not deduction. A rule induced from data is supported by the data, never proven by it. Every model in this course produces an estimate carrying an irreducible error, and the whole discipline of evaluation exists to quantify that error rather than to eliminate it.
The definition is quoted constantly and used rarely. Used properly, it is a specification checklist. Take each component in turn.
T — the task. The task is what the program does on a new case, not what the analyst does with the results. "Reduce churn" is a business objective, not a task. "Assign each active subscriber a probability of canceling within 30 days" is a task: it names the input, the output and the horizon.
E — the experience. In supervised learning, experience is a set of past observations for which the correct answer is already known. Two properties matter more than volume. The observations must be representative of the population the model will face, and the correct answers must have been recorded under the same conditions in which the model will operate. A corpus of emails labeled by one administrator in 2019 is experience; whether it is relevant experience for 2026 traffic is a separate and harder question.
P — the performance measure. P is chosen, not given. Choosing it is a business decision disguised as a technical one, because different measures rank the same two models in opposite orders. Chapters 052 to 075 are devoted entirely to this choice, and the recurring question of the course is stated there: do not ask which metric is best, ask which error costs more.
Applying the three-part decomposition to a second, non-textual case makes the pattern visible:
| Component | Spam detection | Residential property valuation |
|---|---|---|
| T | Classify an email as spam or not spam | Estimate the sale price of a property from its characteristics |
| E | A corpus of previously labeled emails | Recorded sales: characteristics plus the price actually paid |
| P | Proportion of emails classified correctly | Mean absolute error in dollars on unseen sales |
| Output type | A category | A number |
| Chapter | Classification, chapter 010 | Regression, chapter 010 |
Notice what the two columns share. In both, the correct answer for past cases was available before training. That shared property is the definition of supervised learning, formalized in chapter 004.
A diagnostic use of the definition. When a stakeholder proposes a project, ask for T, E and P in that order. The proposal usually fails at E — the labeled history does not exist, or exists only for the cases that were acted upon. That failure is worth discovering in a meeting rather than in month three.
| Inputs supplied | Output produced | |
|---|---|---|
| Classical programming | Rules + Data | Answers |
| Machine Learning | Data + Answers | Rules |
The answers are on the input side. This is the entire reversal, and it explains in one line why the availability of labeled history governs whether a project is feasible at all. A team that cannot supply past answers cannot supply the input the learning algorithm consumes.
Two consequences follow immediately, and both are developed in chapter 002.
The produced rules are not readable as specification. A trained model is a set of numeric parameters, not a document a domain expert can review line by line. Interpretability has to be recovered afterwards with dedicated tooling (chapter 080).
The produced rules are valid only within the regime of the training data. Classical code behaves identically on inputs its author never imagined, for better or worse. A model extrapolating beyond its training regime produces confident nonsense, which is why monitoring for distribution drift is a production requirement rather than a refinement (chapter 082).
Task: determine whether an image contains a cat.
A first attempt at formalization by rules:
IF triangular ears AND whiskers AND fur THEN catThe counter-examples arrive immediately:
| Counter-example | How the rule fails |
|---|---|
| A fox | Satisfies all three conditions, is not a cat |
| A cat photographed from behind | Whiskers not visible |
| A sphynx cat | No fur |
| A cat curled into a ball | No identifiable structure |
| A kitten | Different proportions |
Each patch invites its own counter-example. The number of rules required diverges without ever reaching acceptable coverage, and the rule set becomes unmaintainable long before it becomes accurate.
The deeper reason is worth stating precisely, because it is the criterion you will apply to future problems. The task belongs to perception, and the criteria a human uses to recognize a cat are not available to that human for inspection. You recognize a cat in under 200 milliseconds and cannot report which features you used. A rule you cannot articulate is a rule you cannot write down. Machine Learning is the response to precisely this situation: the examples exist even though the rule cannot be stated.
The same task expressed as Machine Learning
Two elements of this diagram deserve attention now, because they recur in every chapter that follows.
The learning algorithm and the model are distinct objects. The algorithm is the procedure; the model is what the procedure produces on this particular dataset. Chapter 008 makes the distinction formal.
The output is a probability, not a verdict. The model returns 0.94, and a human decides what threshold converts 0.94 into an action. That threshold is a business decision, and chapter 062 shows how much it changes system behavior.
No adult transmits the following definition to a three-year-old:
"A cat is a quadruped mammal of the family Felidae, characterized by erect triangular ears, facial vibrissae and vertically slit pupils."
The child is exposed to hundreds of instances, each accompanied by a verbal label: "look, a cat." Eventually the child correctly labels an individual cat it has never seen, of a breed it has never encountered, in a posture it has never observed.
The rule was never transmitted. It was induced from labeled examples. That is exactly the structure of supervised learning: examples, labels, and generalization to unseen cases.
The analogy is also useful where it breaks down. A child needs a handful of examples, transfers effortlessly from photographs to live animals, and asks questions. A model needs thousands of examples, transfers poorly across conditions it never saw, and asks nothing. Use the analogy to explain the mechanism, never to justify an expectation about sample size.
Machine Learning is appropriate when the following conditions hold. They are cumulative in spirit: the more of them a problem satisfies, the stronger the case.
| Condition | Justification | Illustration |
|---|---|---|
| The rule cannot be formalized | The criteria are not verbalizable | Face recognition |
| The number of factors and interactions is large | Writing the rules by hand diverges | Customer risk scoring |
| The environment changes | Hand-written rules go stale | Fraud detection, spam |
| Personalization must be massive | One rule per individual is impossible | Large-scale recommendation |
| An approximate answer is acceptable | The model produces an estimate, never a certainty | Property valuation |
| Sufficient labeled data exists | Induction requires observations | Transaction history |
The symmetrical list matters more, because it is the one that is skipped.
| Situation | Reason | Alternative |
|---|---|---|
| The exact rule is known and stable | The model would introduce an unnecessary approximation | Implement the formula |
| Exactness is a legal obligation | An approximation is not legally defensible | Explicit, audited rule |
| Data volume is insufficient | The induction has no statistical foundation | Domain expertise, collect first |
| An error is catastrophic and irreversible | No model reaches 100% | Mandatory human oversight |
| No historical data exists | There is nothing to induce from | Market study, prototyping |
Two of these deserve a sentence of expansion.
The exact rule is known. Payroll tax is computed by a published schedule. Training a model to approximate that schedule to within a few dollars replaces a correct answer with a wrong one, at greater cost, with worse traceability. Formalization is not a fallback for teams that cannot do Machine Learning; it is the superior solution whenever the rule is available.
No historical data exists. A model cannot be trained on a phenomenon that has never been recorded. The launch of a product with no predecessor, a regulation with no precedent, a market with no history — these are not hard Machine Learning problems, they are non-problems for the method. The correct move is to instrument the process so that the data exists in twelve months.
Guiding principle: Machine Learning is a tool for statistical approximation. Use it only when the approximation is acceptable and direct formalization is impractical. Both halves of that sentence must hold.
Rigorous definition
A subset of Machine Learning based on artificial neural networks comprising several successive layers of non-linear transformation, each layer producing a representation of the data at an increasing level of abstraction, these representations being learned rather than specified by hand.
The discriminating point
What characterizes Deep Learning is not depth as such, but the automatic learning of representations (representation learning). In classical Machine Learning, the engineer designs the explanatory variables. In Deep Learning, the network builds its own hierarchy of intermediate representations.
Origin of the word "deep"
The qualifier refers to the number of layers stacked between input and output — the depth of the computation graph. It carries no qualitative connotation whatsoever. A deep network is not a profound one.
In plain terms
A many-stage network that builds, stage by stage, the features of the data that matter for the task.
Point of caution
Depth is a necessary condition, not a sufficient one, and it is not the interesting one. A three-layer network that learns its own representations is doing representation learning; a hundred-layer architecture applied to features an engineer designed by hand is a deep model that has been denied the property that makes deep models valuable.
The essential point: none of these layers was specified by hand. The labeling "layer 1 = edges" is an a posteriori interpretation of what the network turned out to learn, obtained by inspecting activations after training. Nobody wrote an edge detector.
This single property accounts for both the power and the opacity of Deep Learning, and the two are the same fact seen from two sides.
It is powerful because it removes the bottleneck that limits classical methods on unstructured data. Nobody has to invent, for each new task, the right descriptors of an image or a waveform.
It is opaque because the intermediate quantities carry no name and no unit. In a logistic regression, a coefficient attaches to "monthly charge, in dollars." In a convolutional network, unit 1,847 of layer 6 attaches to nothing anyone specified, and explaining a decision requires dedicated tooling (chapter 080).
| Criterion | Classical ML | Deep Learning |
|---|---|---|
| Data volume required | 10² to 10⁵ observations | 10⁴ to 10⁷ and beyond |
| Compute resources | Standard CPU | GPU or dedicated accelerators |
| Data types it suits | Structured tabular data | Unstructured data: image, audio, text, video |
| Feature engineering | The engineer's responsibility | Learned by the network |
| Interpretability | High to moderate | Low without dedicated tooling |
| Time to a working model | Minutes to hours | Hours to weeks |
| Representative algorithms | Regression, trees, random forests, gradient boosting | CNNs, RNNs, Transformers |
The volume figures are orders of magnitude, not thresholds. The requirement is driven by the number of parameters to estimate and by the complexity of the phenomenon, not by a number that can be quoted out of context.
Established by industrial practice and by the comparative literature: on tabular data, tree-based ensemble methods (gradient boosting, random forests) remain highly competitive against deep architectures, while offering training times lower by several orders of magnitude and substantially better interpretability.
The honest version of this statement includes its boundary. It is not a claim that deep models cannot be made to work on tables — with enough tuning they often match tree ensembles. It is a claim about cost. The tree ensemble reaches comparable accuracy in minutes, with default hyperparameters, on a laptop, and tells you which variables drove the result. On tabular problems, that is the whole comparison.
Deep Learning retains a decisive advantage on unstructured data: computer vision, audio signal processing, natural language processing, video. On these, the gap is not one of cost but of capability, and no tree ensemble closes it.
This course covers supervised Machine Learning applied to tabular data, which accounts for the majority of enterprise use cases: churn, credit risk, fraud, demand forecasting, pricing, predictive maintenance, medical screening.
A hammer and an excavator are both valid tools. The choice between them does not depend on the intrinsic power of the tool but on its fit to the task.
Driving a nail with an excavator is technically possible and operationally absurd: disproportionate mobilization cost, insufficient precision, collateral damage.
Demolishing a building with a hammer is impossible regardless of effort.
Professional competence does not consist of mastering the most powerful tool. It consists of selecting the tool that fits the problem — and of being able to justify that choice to someone who has read that the excavator is state of the art.
The question arises in every cohort, and the taxonomy answers it without amendment.
A large language model is a deep neural network, so it is Deep Learning, so it is Machine Learning, so it is Artificial Intelligence. It occupies no new position in the diagram of section 2. What is distinctive is the experience E in Mitchell's sense: the model is trained on very large text corpora with a self-supervised objective, in which the labels are constructed from the text itself rather than annotated by hand. Chapter 003 places self-supervision relative to the three learning paradigms.
Two consequences for this course. First, nothing you learn here is made obsolete by these models: overfitting, leakage, class imbalance and metric selection apply to them exactly as they apply to a logistic regression. Second, they are not the subject of this course, and they are not the appropriate tool for predicting churn from a table of subscriber records.
Rigorous definition
An interdisciplinary field combining statistics, computer science and domain knowledge, whose object is the extraction of knowledge and decision support from data, across the entire chain: acquisition, preparation, exploration, modeling, validation, communication.
Scope
Data Science encompasses Machine Learning as one of its components, alongside data preparation, exploratory analysis, statistical inference and reporting to decision makers.
Empirical distribution of effort
Professional surveys generally place between 50% and 80% of a practitioner's time on acquisition, cleaning and preparation of data. Modeling proper is a minority share of the activity.
Point of caution
That distribution is not a sign of a badly run project. It is the shape of the work. Chapters 013 to 025 of this course reflect it: preparation receives more pages than algorithms, because in practice it receives more hours and it determines more of the outcome.
Rigorous definition
The body of methods used to describe, summarize and interpret a set of data in order to characterize observed phenomena.
The central distinction
Data analysis is principally concerned with explaining the past and characterizing the present. Predictive modeling is concerned with estimating values that have not been observed, generally future ones.
In plain terms
Data analysis answers "what is going on in this data." Predictive modeling answers "what will this new case do."
Point of caution
The distinction is one of purpose, not of technique. A linear regression fitted to explain which factors are associated with an outcome, and reported with confidence intervals, is data analysis. The same linear regression fitted to produce a number for a case not yet observed, and evaluated on held-out data, is predictive modeling. The estimation procedure is identical; the validation protocol is not.
Rigorous definition
A body of data whose volume, velocity or variety exceeds the processing capacity of conventional data management systems, requiring distributed architectures.
The characteristic dimensions
The classical model retains three dimensions, frequently extended:
Read as a test, the three dimensions are disjunctive. A stream of 40,000 sensor readings per second is Big Data on velocity alone, whatever its total size. An archive mixing scanned PDFs, audio recordings and relational tables is Big Data on variety alone.
Point of caution
Big Data and Machine Learning are orthogonal notions. Big Data describes a scale and an architecture, not an analytical purpose. You can exploit Big Data with no predictive model at all — regulatory log retention is the standard example — and you can build strong models on a few thousand observations.
Additional caution
The definition is relative to available technology, so it moves. A volume that required a cluster in 2012 fits in the memory of a single machine today. Treat "Big Data" as a statement about the mismatch between a dataset and the systems at hand, never as a fixed number of terabytes.
Illustrated on a single case: a telecom operator observes erosion of its subscriber base. The same business situation produces four different deliverables depending on the level.
| Level | Question | Deliverable | Tooling |
|---|---|---|---|
| Descriptive | What happened? | "Quarterly churn stands at 8.2%" | Querying, dashboards |
| Diagnostic | Why? | "Churn is concentrated among month-to-month plans with more than two support contacts" | Segmentation, statistical tests |
| Predictive | What will happen? | "1,240 subscribers have a probability above 0.70 of canceling within 30 days" | Supervised learning |
| Prescriptive | What should we do? | "Target 300 subscribers with a 15% discount; expected net gain $84,000" | Constrained optimization |
The ordering is a dependency, not a ranking of sophistication. Each level consumes the previous one.
The predictive model of level 3 is built on variables that level 2 identified as associated with churn. The optimization of level 4 consumes the probabilities produced at level 3, and requires two further inputs the model does not provide: the cost of the retention offer and the estimated value of a retained subscriber. A team that skips levels 1 and 2 typically arrives at level 3 with no usable variables, and a team that stops at level 3 delivers a list of 1,240 names that nobody acts on.
Supervised learning, the subject of this course, operates at the predictive level. Recognizing which level a request belongs to is the first thing to do when a stakeholder asks for a model, because roughly half of such requests are descriptive or diagnostic questions in predictive clothing.
The branching criterion at the Machine Learning node is a single question: what supervision signal is available during training? A known correct answer for each observation gives supervised learning; no answer gives unsupervised learning; a delayed reward following an action gives reinforcement learning. Chapter 003 develops the three paradigms and their boundary cases.
The branching criterion below supervised learning is equally narrow: the nature of the target variable. Categorical target, classification. Numeric target, regression. Chapter 010 treats the cases where the answer is not obvious.
Recurring point of caution: Deep Learning is not a fourth family alongside supervised, unsupervised and reinforcement learning. It is a class of models usable within all three paradigms. A network trained on labeled images is supervised; an autoencoder is unsupervised; a policy network optimized by reward is reinforcement learning. The paradigm is determined by the supervision signal, the model class by the functional form. They vary independently, and a diagram that places Deep Learning as a sibling of "supervised learning" is drawing a category error.
| Term | Nature of the concept | Condensed definition | Illustration |
|---|---|---|---|
| Artificial Intelligence | Objective | Automate cognitive tasks | Itinerary planner |
| Machine Learning | Approach | Induce rules from data | Email filtering |
| Deep Learning | Class of models | Multilayer networks that learn their own representations | Image recognition |
| Supervised learning | Paradigm | Induction from labeled examples | Price estimation |
| Unsupervised learning | Paradigm | Discovery of structure without labels | Customer segmentation |
| Reinforcement learning | Paradigm | Optimization of a policy through reward | Robotic control |
| Data Science | Discipline | The full chain of turning data into value | A business function |
| Data analysis | Activity | Description and interpretation of what exists | Dashboard |
| Big Data | Scale and architecture | Volumes exceeding conventional systems | Large-scale logging |
| Algorithm | Procedure | A finite sequence of operations producing a result | Random Forest |
| Model | Artifact | The parameterized result of applying an algorithm to data | A serialized file |
The last two rows are the ones most often collapsed in everyday speech. "Random Forest" names an algorithm; the object saved to disk after training on your data is a model. One is a recipe, the other is the dish. Chapter 008 formalizes the distinction and shows why it matters for versioning, reproducibility and deployment.
AI names an objective; Machine Learning names one approach among several for reaching it. A rule-based expert system is AI and is not Machine Learning.
The error is not merely terminological. It leads teams to reject a rules engine as "not real AI" for problems where the rule is known, stable and auditable — precisely the problems where a rules engine is the correct answer.
Correct formulation: "Machine Learning is an approach to artificial intelligence based on induction from data."
Relative performance depends on the nature of the data. On tabular data, tree-based ensemble methods remain highly competitive, at markedly lower computational cost and markedly lower opacity.
Correct formulation: "The choice between Deep Learning and classical methods is determined by the nature of the data, the volume available, interpretability constraints and the computational budget."
The volume required depends on the complexity of the phenomenon being modeled and on the number of explanatory variables, not on any absolute threshold. Many production models rest on a few thousand observations, and a large volume of unrepresentative data is worth less than a small representative sample.
Correct formulation: "Data quality and representativeness take precedence over raw volume."
Deep Learning is a class of models. A neural network trained on labeled data is supervised learning; an autoencoder is unsupervised learning; a network optimized by reward is reinforcement learning. The paradigm is fixed by the supervision signal, not by the model's architecture.
Correct formulation: "Deep Learning is a class of models that can be deployed within any of the three learning paradigms."
A statistical model identifies regularities in a numeric representation. It holds no semantic representation of the domain. It does not know what a customer is, what a dollar is, or what canceling a subscription means to the person canceling it.
Operational consequence: a model's failure modes do not coincide with a human operator's. A model can fail on cases a human finds trivial while succeeding on cases a human finds hard. That asymmetry has to be anticipated when designing safeguards: reviewing only the cases a human would find difficult will miss exactly the errors a model makes.
Correct formulation: "The model identifies statistical regularities; any interpretation of those regularities as understanding is supplied by the reader, not by the system."
Behavior that varies over time is not evidence of learning. A rules engine whose thresholds are updated quarterly by an analyst changes behavior; a system that draws on a live data feed changes behavior; a program with a random component changes behavior. None of them induced anything.
The test is Mitchell's: performance on tasks in T, measured by P, improves with experience E. If no E is being consumed and no P is improving, the system is not learning, whatever its output does.
Correct formulation: "The system's behavior changes because its inputs or its hand-set parameters changed, not because it learned from experience."
TAXONOMY
Deep Learning ⊂ Machine Learning ⊂ Artificial Intelligence
NATURE OF EACH CONCEPT
Artificial Intelligence : an objective
Machine Learning : an approach based on induction
Deep Learning : a class of models
DISCRIMINATING CRITERION FOR MACHINE LEARNING
The rules are induced from observations,
not specified by a designer.
DISCRIMINATING CRITERION FOR DEEP LEARNING
The intermediate representations are learned,
not designed by hand.
ANALYTICAL LEVELS
Descriptive -> Diagnostic -> PREDICTIVE -> Prescriptive
Supervised learning operates at the predictive level.
REFERENCE DEFINITION (Mitchell, 1997)
A program learns from experience E with respect to a task T
and a performance measure P if its performance at T, as
measured by P, improves with E.
WHAT THE TAXONOMY IS FOR
AI without learning is possible: Deep Blue, a thermostat.
Machine Learning without Big Data is possible and common.
Deep Learning is not a fourth learning paradigm.Summary statement
Machine Learning consists of inducing a decision rule from observations rather than specifying it explicitly; supervised learning is the paradigm in which those observations come with the target value to be predicted.
Associated quizzes : 001.1-quiz-taxonomy.md,
001.2-quiz-artificial-intelligence.md, 001.3-quiz-machine-learning.md,
001.4-quiz-deep-learning.md, 001.5-quiz-data-science-big-data.md,
001.6-quiz-analytical-maturity.md,
001.7-quiz-positioning-supervised-learning.md
Next chapter : 002.0-classical-programming-vs-machine-learning.md