Chapter 001 established that machine learning is an approach to artificial intelligence founded on induction. This chapter works out what that formula means when you sit down to design a system.
Both paradigms handle the same three objects — rules, data, answers. They differ only in where each object sits in the production chain.
| Paradigm | Inputs supplied | Output produced | Author of the rules |
|---|---|---|---|
| Deterministic programming | Rules + Data | Answers | A human designer |
| Inductive learning | Data + Answers | Rules | The algorithm, from the observations |
The discriminating point : in the first paradigm the rule is an input to the system; in the second it is the output.
That single displacement drives everything else in this chapter. It determines who is accountable for the rule, how the rule is verified, what happens when the world changes, and what kind of failure the system exhibits when it goes wrong.
Before comparing the two paradigms further, the object they both produce needs a definition that does not depend on how it was obtained.
Rigorous definition
A deterministic mapping from a description space X to a decision space Y, which assigns to every description x in X exactly one decision in Y: formally, a function h : X → Y. In classification, Y is a finite set of categories; in regression, an interval of the real line.
In plain terms
A procedure that, for any situation you can describe, returns one decision and only one.
The essential point
The definition says nothing about the origin of the rule. An IF ... THEN ... clause, a statutory rate schedule, a tree learned on 200,000 case files and
a neural network are all decision rules in this sense. They differ only in how h
was obtained.
Point of caution
A model is no less a rule than a hand-written clause: it is a rule whose statement nobody ever wrote down. That property is the source of both its power and its resistance to audit.
The consequence is worth stating plainly. Choosing between the two paradigms is never a choice between "having a rule" and "not having a rule". Both approaches end with a rule in production. The choice is about who writes it, what evidence justifies it, and what it costs to keep it current.
The two paradigms rest on two distinct modes of inference, identified by logic long before computing existed.
Deduction
The mode of inference in which a conclusion is derived from premises by transformation rules that preserve truth. If the premises are true and the derivation is valid, the conclusion is necessarily true. Illustration : "Every message carrying an executable attachment is blocked. This message carries one. Therefore it is blocked."
Induction
The mode of inference in which a general proposition is formed from a finite set of particular observations. The conclusion exceeds the content of the premises: it is not guaranteed, only more or less probable given what was observed. Illustration : "Across 40,000 observed messages, those combining certain terms were reported as unwanted in 96% of cases. Therefore a new message combining those terms is probably unwanted."
The problem of induction
Induction offers no logical guarantee. Statistical learning theory (Vapnik and Chervonenkis, 1970s) does not dispose of that objection: it bounds it with probabilistic guarantees, conditional on explicit assumptions, the principal one being that future data are drawn from the same distribution as the training data.
Operational consequence
A deterministic system is wrong when its rule is wrong. An inductive system is wrong some fraction of the time by construction, including when every step has been carried out correctly.
| Criterion | Deduction | Induction |
|---|---|---|
| Direction of inference | General to particular | Particular to general |
| Status of the conclusion | Necessary if the premises are true | Probable, revisable |
| Content of the conclusion | Contained in the premises | Exceeds the premises |
| Effect of a new observation | None on the rule | May revise the rule |
| Corresponding paradigm | Deterministic programming | Machine learning |
| Failure mode | Rule badly specified | Insufficient generalization |
The last row deserves emphasis. The two failure modes call for different engineering responses. A badly specified rule is found by reading the specification and reproducing the case; it is fixed once, and the fix is permanent. Insufficient generalization is found only by measurement on data the model never saw, and the fix — more data, different features, a different model family — is itself subject to measurement.
Induction needs a formal object to aim at. Supervised learning postulates one.
Rigorous definition
We postulate the existence of an unknown target function f : X → Y linking the description space X to the space of values to be predicted Y. It is accessible only through a finite sample of observed pairs (x₁, y₁), ..., (xₙ, yₙ). The algorithm selects, from a hypothesis space H fixed by the choice of model family, a hypothesis f-hat that minimizes a measure of error on the sample, in the hope that f-hat approximates f across the whole domain.
In plain terms
There is a real relationship between what you observe and what you want to predict. Nobody knows its statement. You build an imitation of it from the examples you have.
The three gaps between f-hat and f
Point of caution
The existence of f is a working assumption, not an established fact. When the target depends on factors absent from the descriptions x, no function from X to Y is exact, whatever the algorithm. Section 5 develops this point.
Read the diagram from the top left. Only two boxes are ever observed: the sample and f-hat. The target function f is never seen, never printed, never validated directly. Everything a practitioner does — splitting data, measuring error, comparing models — is an indirect attempt to bound the distance between the one object the algorithm produced and the one object nobody can inspect.
This asymmetry explains a habit that outsiders find excessive: the insistence on evaluating a model on data withheld from training. It is not procedural formality. It is the only available evidence about a quantity that cannot be measured head-on.
The following case reconstructs a typical trajectory in email filtering between the mid-1990s and the mid-2000s. The orders of magnitude are representative.
Month 1 — three rules are enough.
IF subject contains "viagra" THEN spam
IF subject contains "lottery" THEN spam
IF body contains "click here" THEN spamThree rules, written in half a day, intercept 82% of unwanted messages at zero maintenance cost. The system is a success by every measure available at the time, and the decision to write rules rather than collect data was the correct one.
Month 3 — obfuscation. Senders start altering the spelling of the filtered
terms: V1AGRA, V-I-A-G-R-A, VÍAGRA, \/IAGRA, Vi@gra, or the same
letterforms in homoglyphic Cyrillic characters. The interception rate falls to
51% in three weeks. The team adds one rule per observed spelling and the corpus
grows from 3 to 47 rules.
Month 6 — false positives. The accumulated rules start blocking legitimate mail: correspondence from a partner pharmacy, a medical newsletter followed by 4,000 customers, internal marketing campaigns that contain the words "click here". Three incidents reach senior management in a single month. The team no longer dares delete rules, since each one intercepts real spam. Instead it adds exceptions: sender allowlists and per-domain waivers. The corpus reaches 210 rules, 60 of which are exceptions to other rules.
Month 12 — new vectors. The text moves into an image, then into an attachment, then gets fragmented by invisible characters. Each vector demands its own family of rules: 610 rules.
Month 24 — unmanageability.
| Indicator | Month 1 | Month 6 | Month 12 | Month 24 |
|---|---|---|---|---|
| Number of rules | 3 | 210 | 610 | 1,800 |
| Interception rate | 82% | 79% | 77% | 76% |
| False positive rate | 0.1% | 0.9% | 1.6% | 2.3% |
| Staff assigned to maintenance | 0 | 1 | 2 | 3 |
| Response time to a new campaign | 1 day | 3 days | 7 days | 11 days |
| Share of rules with an identifiable author | 100% | 70% | 35% | 12% |
Nobody on the team knows the full set of 1,800 rules, and any modification produces unpredictable regressions elsewhere in the corpus. Performance has fallen despite a threefold increase in maintenance effort.
Read the last row before the others. The share of rules whose author can still be identified drops from 100% to 12%. At that point the corpus has stopped being a specification and become an artifact: it encodes decisions nobody can reconstruct or justify, and it is therefore no longer safe to change.
The failure cannot be blamed on the team's competence. It is structural.
| Mechanism | Effect on the rule corpus | Governing quantity |
|---|---|---|
| Combinatorial explosion | The number of cases to cover grows faster than the capacity to write them | Number of variants of a single pattern |
| Coupling between rules | Every new rule alters the behavior of existing rules | Number of rule pairs |
| Adversarial obsolescence | Rules expire on a schedule set by a hostile third party | Gap between attack cycle and correction cycle |
Mechanism 1 — combinatorial explosion. A single six-character term admits a
number of equivalent spellings that grows multiplicatively. Count the plausible
substitutions position by position for viagra: four for v (v, V, \/,
and the Greek nu), four for i (i, 1, l, í), four for each of the two
occurrences of a (a, @, á, 4), two for g (g, 9). That already
gives 4 × 4 × 4 × 4 × 2 = 512 forms. Optionally inserting a separator at each of
the five internal positions multiplies that count by 2⁵ = 32, for a total of
16,384 forms — for one word, before considering case variation or homoglyphs
drawn from other alphabets.
Manual writing progresses linearly while the space to cover grows exponentially. No amount of maintenance effort closes that gap. This is the decisive asymmetry, and it does not depend on how good the team is.
Mechanism 2 — coupling between rules. A corpus of n rules contains n(n−1)/2 pairs that can interact.
| Number of rules | 3 | 47 | 210 | 610 | 1,800 |
|---|---|---|---|---|---|
| Number of pairs | 3 | 1,081 | 21,945 | 185,745 | 1,619,100 |
Confirming that a new rule conflicts with none of the others means examining 1,800 interactions on every addition. That check is not feasible by hand, so it is not performed. The regressions are discovered in production, by users.
Note the shape of the growth. Between month 12 and month 24 the corpus roughly triples, but the number of interacting pairs grows by a factor of nearly nine. The verification burden does not track the size of the system; it tracks the square of it.
Mechanism 3 — obsolescence against an adaptive adversary.
Rigorous definition
An agent whose objective function includes circumventing the decision system, and who receives an observation signal on that system's decisions allowing it to adjust its behavior.
In plain terms
Someone who benefits from fooling the system, who can see what gets through and what gets blocked, and who adapts accordingly.
The critical property: asymmetry of cycles
The sender tests a variant and observes the outcome within hours, at near-zero cost. The defending team must notice the decline, diagnose the vector, write the rule, verify that it breaks nothing, and deploy it — several days. As long as the attacker iterates faster than the defender, performance converges toward that of a system filtering only obsolete campaigns.
Point of caution
This property is not specific to rule-based systems: a model facing an adaptive adversary degrades too. The difference lies in the unit cost of an update — retraining is tooled and repeatable, rewriting 1,800 rules is not.
The three mechanisms compound. Combinatorial explosion sets how many rules must exist; coupling sets how expensive each one is to add safely; adversarial obsolescence sets how quickly the whole corpus expires. The month-24 numbers are the arithmetic consequence, not an accident of execution.
Rather than specifying the patterns characteristic of spam, you assemble a corpus of messages whose nature is known and let the algorithm identify the discriminating regularities.
The loop from user reports back into the corpus is the structural difference. In the rule-based system, every new spam campaign is a cost. In the inductive system, every new spam campaign that a user reports is an input.
| Dimension | Rule-based system | Inductive system |
|---|---|---|
| Response to a new spelling | Write a dedicated rule | Absorbed if the corpus is refreshed |
| Adaptation effort | Proportional to the number of variants | Constant: retrain |
| Source of the signal | The analyst who observes | Reports from users |
| Effect of traffic volume | Rising maintenance load | Additional resource for the model |
| Legibility of a decision | High: the triggered rule is named | Reduced: a weighting of many indicators |
Point of caution : the inductive reformulation does not eliminate the work. It moves the work to maintaining a labeled corpus, a retraining pipeline and performance monitoring. The gain comes from the fact that this work is tooled and repeatable, not from its disappearance.
That caution matters at the point of proposing the change. A team that pitches machine learning as a way to stop maintaining the filter has mis-sold the project. The correct pitch is that the maintenance becomes a fixed, automatable cost instead of one that grows with the adversary's imagination.
The order of the questions is not arbitrary. The first question is never "which algorithm should we use" but "is the rule known". Machine learning becomes relevant only when the answer is no.
Three of the four terminal states in the tree are not machine learning. That proportion is closer to professional reality than the volume of published material on the subject would suggest.
| Condition | Justification | Illustration |
|---|---|---|
| The rule is not formalizable | The decision criteria cannot be put into words, not even by the experts who apply them | Recognizing a face, judging a composite risk |
| Factors and interactions are numerous | Manual writing diverges and coupling between rules becomes unmanageable | Credit scoring on 60 variables |
| The environment changes | Written rules expire; retraining absorbs the change at constant cost | Fraud detection, email filtering |
| Personalization must operate at scale | One rule per individual is impossible to write and to maintain | Recommendation across several million accounts |
| An approximate answer is acceptable | The model produces an estimate carrying uncertainty, never a certainty | Price estimation, demand forecasting |
| Sufficient labeled history exists | Induction requires observations whose target value is known | Transactions with an observed outcome |
These six conditions are conjunctive. A problem that satisfies five and fails the sixth — typically the absence of labeled history — is not a machine learning problem as it stands.
The word "as it stands" carries the practical content. Failing the sixth condition is often a temporary state that instrumentation can fix within two or three quarters. Failing the first — the rule is perfectly formalizable — is permanent, and no amount of data collection changes the verdict.
| Contraindication | Reason | Alternative |
|---|---|---|
| The exact rule is known, stable and verifiable | The model would substitute an approximation for an exact answer, with nothing gained in return | Implement the formula |
| Exactness is a legal or contractual obligation | A statistical approximation is legally inadmissible on an amount owed | Explicit rule, audited, versioned |
| The volume of observations is insufficient | Induction on too small a sample produces a rule that does not generalize | Domain expertise, prior data collection |
| The error is catastrophic and unrecoverable | No model reaches 100%: the residual fraction of errors is structural | Systematic human review |
| No history exists because the process is new | There is nothing to induce from | Prototyping, provisional rules, instrumented collection |
Rigorous definition
The property of a decision system that allows, for any decision produced, the reconstruction of the chain of justification that led to it and of the exact state of the system on that date, reproducibly by a third party. It requires that the applied rule be statable, that the version in force be retained, and that the result be reproducible identically.
In plain terms
Being able to show, months later and to someone outside the team, exactly why this decision came out this way and to obtain the same result again.
Point of caution
A statistical model satisfies reproducibility but makes the statement of the rule difficult: post hoc explanation methods produce an approximate justification of the decision, not its statement. When the obligation bears on the individual motivation of a decision, that approximation may be insufficient.
Situation 1 — the rule is known, exact and stable. Computing a sales tax, converting at a published rate, the installment schedule of a fixed-rate loan, validating an identifier against its check digit. A model trained to reproduce these operations would reach 99.7% accuracy where direct implementation reaches 100%. The substitution is a pure regression.
Situation 2 — no history is available. A process created last week has no observations whose target value is known. The question is not which algorithm to choose but how to instrument the process so that observations get collected and labeled.
Situation 3 — auditability is an obligation. When a decision must be individually justified to the person affected or to a regulator, the ability to state the applied rule becomes a design constraint on the same footing as performance. A model remains relevant as decision support, with the enforceable decision produced by an explicit rule or by an operator.
This section is the most important in the chapter. It separates a technical understanding from an operational one.
No machine learning model reaches 100% accuracy on new data.
The statement follows from the inductive nature of the inference: the rule is extrapolated from a finite sample onto a domain that exceeds it. In practice, an announced accuracy of 100% on new data signals either data leakage, covered in chapter 028, or an evaluation carried out on the training data. The relevant question is therefore never "does the model make mistakes" but:
Who handles the errors, at what cost, and within what delay?
A model reaching 95% accuracy is generally presented as a good model. Set against the volume processed:
Daily decision volume : 100,000
Model accuracy : 95%
Correct decisions : 95,000
Erroneous decisions : 5,000 per day| Handling assumption | Load induced | Conclusion |
|---|---|---|
| No handling, error without consequence | None | Model deployable as is |
| Handled by the end user, 30 seconds each | 42 hours of attention per day, diffuse | Acceptable if the error is visible and correctable |
| Internal human review, 4 minutes per case | 333 hours per day, roughly 42 full-time positions | The recovery mechanism costs more than the model |
| Error not detectable by the user | Zero load, unbounded risk | The most dangerous configuration |
The last row is the most important. A costly but visible error is a sizing problem. A silent error is a design problem.
The arithmetic is worth doing out loud, because the gap between rows two and three is where most deployment decisions actually turn. Thirty seconds of a user's attention on an error they can see and dismiss is an annoyance. Four minutes of an internal operator's time on the same volume is a department. The model is identical in both rows; only the recovery path differs.
The first branch of this tree comes before any question of volume or cost. If the error is not detectable, no amount of accuracy makes the system safe, because nothing in the system will ever report that it is failing.
| Context | Nature of the error | Severity | Reversibility | ML appropriate |
|---|---|---|---|---|
| Content recommendation | Irrelevant suggestion | Low | Immediate: the user ignores it | Yes, without reservation |
| Email filtering | False positive: legitimate mail quarantined | Medium | Good if the quarantine is browsable | Yes, provided the quarantine is accessible |
| Predictive maintenance | Spurious alert, or a failure not anticipated | Medium to high | Partial: cost of intervention or cost of downtime | Yes, with an asymmetric threshold |
| Credit decisions | Unjustified refusal to a solvent applicant | High | Low: harm suffered, litigation possible | Yes, subject to explainability and an appeal path |
| Automated moderation | Removal of lawful content | High | Conditional on an appeal path existing | Yes, with guaranteed human re-review |
| Medical diagnosis | False negative: undetected pathology | Very high | Low to none | No, not as an autonomous decision: decision support |
| Payroll or tax computation | Incorrect amount | High | Correctable, but a breach of an exactness obligation | No: the exact rule exists |
Three readings of this table need to be mastered.
First reading : severity depends on the context of use, not on the error rate. The same 95% model is excellent for recommendation and unacceptable for payroll.
Second reading : reversibility is the decisive operational criterion. A severe but recoverable error is managed by design — a cautious threshold, a verification queue, an appeal path. An irreversible error cannot be managed that way.
Third reading : the two error types of a classifier do not carry the same cost. That asymmetry is a design parameter, adjustable through the decision threshold, and is developed in chapter 062.
A production line does not pursue zero defects at any price. It sizes an acceptable defect rate, then designs the detection and rework capacity to match. That rate is not a property of the machine: it follows from the unit cost of a defect reaching the customer, the cost of inspection, and the volume produced. A manager who demands zero defects without sizing the inspection does not get zero defects. He gets undetected defects.
The acceptable error rate of a model is, in the same way, an engineering decision — never a property of the algorithm.
The four limits below are not implementation defects. They follow from the inductive nature of the approach, and no algorithm, no architecture and no volume of data removes them.
Statement
A learning algorithm extracts regularities present in the descriptions supplied. It cannot base a decision on a factor that appears in no variable, either directly or through correlation.
Illustration
A property valuation model fed only floor area and bedroom count cannot price the view, the orientation, or the condition of the roof, which nonetheless determine a substantial share of the price. The model does not report the deficiency: it produces an estimate whose residual error will be attributed to "noise" when it in fact reflects missing information.
Practical consequence
The performance ceiling is set by the information content of the variables, not by the choice of algorithm. This limit explains a frequent disappointment: moving from a regression to a sophisticated ensemble method buys a few points at most when the necessary information is absent from the dataset.
Statement
A predictive model identifies statistical associations between explanatory variables and the target variable. An association is not a cause-and-effect relationship: a high-performing model may rest on associations whose causal interpretation is wrong, or even reversed.
The documented asthma and pneumonia case
A study from the 1990s on patients admitted for pneumonia set out to predict mortality risk, so that low-risk patients could be directed toward outpatient care. The model learned a rule that was exact on the data and dangerous in practice:
Pneumonia patients with a history of asthma have a lower risk of death.
The association is real in the data. Its explanation: those patients were systematically admitted to intensive care because of their history, received immediate and aggressive treatment there, and therefore survived at a higher rate. The induced rule was a consequence of the care protocol, not a clinical property. Applied as written, it would have sent home precisely the patients whom the existing protocol was protecting. The model was not faulty — it reproduced the data faithfully — and only a review by clinicians surfaced the anomaly.
The case is reported notably by Cooper and coauthors (1997), then taken up and analyzed by Caruana and coauthors (2015) in their work on intelligible models in healthcare.
The formulation to retain
A model predicts what happens under the conditions in which the data were produced. It does not predict what would happen if those conditions were acted upon. Any intervention decision based on a purely predictive model falls outside that model's domain of validity.
Statement
The statistical framework of supervised learning assumes that production data are drawn from the same distribution as the training data. When that assumption stops holding, the performance guarantees fall with it.
Two forms of breakdown
Practical consequence
Usual causes: changes in customer behavior, modification of the product or the business process, regulatory change, adaptation by an adversary, replacement of a sensor, external shock. A model in production therefore does not have constant performance: it has a shelf life, which must be monitored rather than assumed. Putting a model into production without a monitoring mechanism is an incomplete operational decision. Drift, its detection and retraining strategies are covered in chapter 082.
Statement
The quality of the induced rule is bounded by the quality and representativeness of the observations supplied. Data that are erroneous, biased or mislabeled produce a rule that is erroneous, biased or badly calibrated.
The aggravation specific to machine learning
A deterministic program fed invalid data fails visibly: an exception, an aberrant value, a rejected check. A learning algorithm does not fail. It converges, it produces a model, and that model produces predictions that look normal. The failure is silent.
| Data defect | Effect on the model | How it shows up |
|---|---|---|
| Partially erroneous labels | The induced rule reproduces the labeling error | Apparent performance fine, decisions wrong |
| Unrepresentative population | The rule holds only on a subgroup | Performance gap between segments |
| Variable unavailable at decision time | Excellent performance in test, collapse in production | Data leakage, chapter 028 |
| Strongly imbalanced classes | The model favors the majority class | High accuracy, zero detection on the rare class, chapter 050 |
Point of caution
The absence of a runtime error is never evidence of validity for an inductive system. Validation requires inspecting the data, analyzing errors by segment, and confronting the learned regularities with domain knowledge.
Taken together, the four limits describe what a supervised model is and is not. It is a compressed record of the regularities present in one finite sample, described by one chosen set of variables, produced under one set of conditions, at one moment in time. Every limit above is a restatement of one of those four qualifiers.
The objective is to estimate the sale price of a house. Both approaches are carried out side by side on the same properties.
A domain expert proposes a formula grounded in experience:
price = area_m2 x 3,500 + bedrooms x 10,000| Property | Area | Bedrooms | Computed price | Actual price | Error | Relative error |
|---|---|---|---|---|---|---|
| A | 90 m² | 3 | $345,000 | $352,000 | −$7,000 | 2.0% |
| B | 120 m² | 4 | $460,000 | $448,000 | +$12,000 | 2.7% |
| C | 75 m² | 2 | $282,500 | $291,000 | −$8,500 | 2.9% |
| D | 140 m² | 4 | $530,000 | $690,000 | −$160,000 | 23.2% |
On the first three properties the formula is acceptable. On the fourth it misses by $160,000. A gap of that size disqualifies the tool: no professional bases a negotiation on an estimate whose possible error reaches a quarter of the value.
Note what the first three rows would have done to a team that stopped there. Three consecutive errors under 3% look like a validated formula. The failure appears only on the property that differs structurally from the others, and the formula gives no indication that it is out of its depth on that property.
| Factor | Situation of property D | Effect on price |
|---|---|---|
| Neighborhood | Sought-after school catchment, transit station 300 meters away | Strong, multiplicative |
| Renovation | Kitchen and bathroom redone 18 months before the sale | Additive, substantial |
| Garage | Double garage, uncommon in the area | Additive, moderate |
The correction looks simple: add the missing terms.
price = area_m2 x A + bedrooms x B + garage x C
+ recently_renovated x D + neighborhood_index x E + FThe structure is plausible. The problem lies elsewhere:
What are the values of A, B, C, D, E and F?
Nobody knows. They appear in no reference table, and they vary by city, by market segment and by period. An expert can propose an order of magnitude for A. He cannot adjudicate between $34,000 and $41,000 for the value of a renovation, nor quantify a neighborhood index on a scale coherent with the other terms. On top of that come interactions — the value of a renovation depends on the neighborhood — plus non-linearities, and instability: the coefficients will be out of date in eighteen months.
Conclusion : the structure of the problem is identifiable by an expert; its parameters are not. That is exactly the configuration that calls for an inductive approach.
This distinction between structure and parameters recurs throughout the course. Domain expertise is nearly always right about which variables matter and how they enter the model. It is nearly always unable to supply the numbers.
You assemble the completed sales, described by the same variables, with the observed price.
| Area | Bedrooms | Garage | Renovated | Neighborhood index | Observed price |
|---|---|---|---|---|---|
| 90 m² | 3 | 0 | 0 | 3 | $352,000 |
| 120 m² | 4 | 1 | 0 | 3 | $448,000 |
| 75 m² | 2 | 0 | 0 | 3 | $291,000 |
| 140 m² | 4 | 1 | 1 | 7 | $690,000 |
| 105 m² | 3 | 1 | 0 | 5 | $468,000 |
| 160 m² | 5 | 1 | 1 | 6 | $712,000 |
| 82 m² | 2 | 0 | 1 | 4 | $371,000 |
The real dataset holds several thousand rows of this shape, and the learning procedure fits in three statements:
X = sales[["area_m2", "bedrooms", "garage", "renovated", "neighborhood_index"]]
y = sales["price"]
model.fit(X, y)The call model.fit(X, y) is the act of induction: the algorithm searches for the
coefficients that minimize the gap between computed prices and observed prices
across all the sales.
| Term | Estimated coefficient | Business reading |
|---|---|---|
| Area | $3,120 per m² | Marginal value of a square meter |
| Bedrooms | $6,400 per bedroom | Own effect, holding area constant |
| Garage | $14,900 | Premium for a garage |
| Recent renovation | $38,200 | Premium for a recent renovation |
| Neighborhood index | $21,500 per point | Effect of the area |
| Intercept | $18,700 | Value at the origin |
None of these values was supplied. They were induced from the observed sales. The area coefficient differs from the expert's intuition — 3,120 rather than 3,500 — because the initial estimate was implicitly absorbing effects that the other terms now capture.
Prediction on a property never seen:
Property E : 110 m², 3 bedrooms, garage, not renovated, neighborhood index 4
110 x 3,120 = 343,200
3 x 6,400 = 19,200
garage = 14,900
renovation = 0
4 x 21,500 = 86,000
intercept = 18,700
-----------------------
Estimate = $482,000In professional use, this point estimate is reported together with an interval and an average error, both established during evaluation. An estimate delivered without either is not a professional deliverable, whatever the model behind it.
| Criterion | Deterministic approach | Inductive approach |
|---|---|---|
| Origin of the coefficients | Postulated by an expert, unverifiable | Estimated from observed sales |
| Taking a new factor into account | Rewrite the formula, set the coefficient by hand | Add a column and retrain |
| Adapting to a different market | Redo the expert elicitation entirely | Retrain on local data |
| Measuring quality | Gap noticed after the fact, with no protocol | Error measured on sales not used for learning |
| Auditability of a decision | High: the formula is statable | Medium to low, depending on the model family |
Reading the table : the inductive approach wins on four criteria and loses on auditability. That trade-off is representative, and it is settled by the context of use. An indicative estimate for a real estate agent tolerates medium auditability. An appraisal value enforceable in court does not.
Note also the fourth row, which is easy to skim past. The deterministic formula has no protocol for measuring its own quality. Its errors surface one property at a time, in the field, with no aggregate anyone ever computes. The inductive approach does not merely produce better coefficients: it comes with a measurement discipline that the hand-written formula never had.
Press vocabulary — "the algorithm decides", "the machine learns by itself" — sustains the idea that the decision chain is automated end to end. Examining the actual cycle of a project contradicts it.
| Stage of the cycle | Content of the decision | Decider |
|---|---|---|
| 1. Frame the business problem | Determine which question is worth asking | Human |
| 2. Rule on whether ML is warranted | Test the problem against the conditions in section 3 | Human |
| 3. Define the target variable | Choose what gets predicted and how it is measured | Human |
| 4. Choose and collect the data | Determine sources, period, population | Human |
| 5. Define the performance measure and threshold | Fix what counts as success, arbitrate between error types | Human |
| 6. Prepare the data and build the variables | Clean, encode, aggregate, derive | Human, tooled |
| 7. Choose the algorithm family | Trade off performance, interpretability, cost, constraints | Human |
| 8. Fit the model parameters | Estimate the coefficients minimizing error on the sample | Machine |
| 9. Evaluate and interpret the results | Analyze errors, confront domain knowledge, detect anomalies | Human |
| 10. Decide on production deployment | Rule on release and its safeguards | Human |
| 11. Monitor and arbitrate retraining | Detect drift, decide to restart the cycle | Human |
Out of eleven stages, the machine covers one, and that one is entirely conditioned by the human decisions that precede it: the data from stage 4, the target from stage 3, the error measure from stage 5 and the model family from stage 7 fully determine what stage 8 is able to produce.
This is not a rhetorical deflation of machine learning. Stage 8 does something no human can do: it searches a parameter space of arbitrary dimension for the configuration minimizing a stated criterion, across millions of observations, in seconds. The point is that the value of what it produces is bounded by the ten decisions surrounding it.
The dark blocks are human decisions. The light block is the only stage executed by the machine.
A spectrometer produces measurements no operator could obtain by eye. It is irreplaceable on that segment of the chain. It does not choose the research question, does not draw the sample, does not define the protocol, does not interpret the results and does not decide to publish. A learning algorithm occupies exactly the same position.
Rigorous definition
The position of an automated system on the scale running from the mere production of information to a decision executed without human intervention, characterized by the role assigned to the operator inside the decision loop.
In plain terms
How much of the decision the system actually makes, and what the person is there to do.
The three usual regimes
| Regime | Role of the human | Context of use |
|---|---|---|
| Decision support | The human decides, the model informs | Severe errors, motivated decisions, indispensable expertise |
| Human in the loop on exception | The model decides, the human arbitrates uncertain cases | High volume, recoverable errors, identifiable uncertainty |
| Full automation | The model decides alone, the human monitors aggregates | Low-severity errors, massive volume, immediate reversibility |
Point of caution
The autonomy regime is a design decision, independent of the model's performance. A very strong model may be deployed as decision support because the context requires it. A mediocre model may be fully automated because the error carries no consequence there.
Symptom : framing the choice as a modernization, or contrasting an "outdated" method with a "current" one.
Why it is wrong : the two approaches answer two distinct configurations. Machine learning contributes nothing when the rule is known, exact and stable; it introduces an approximation there with nothing gained in return.
Correct formulation : "Deterministic programming and inductive learning address two different configurations. The criterion for choosing is the availability of an exact and stable rule, not the age of the method."
Symptom : describing the project as "we hand the data to the algorithm and it figures everything out".
Why it is wrong : the algorithm executes one stage out of eleven. The choice of target, of data, of performance measure and of model family is human, and it fully conditions the result obtained.
Correct formulation : "The algorithm estimates the parameters. Framing the problem, choosing the data, defining the performance measure and validating the result remain human decisions."
Symptom : deciding to act on a variable because the model assigns it a large weight — "the model shows that a history of asthma lowers the risk, so those patients can go home".
Why it is wrong : a predictive model reproduces the regularities of the process that generated the data, protocols and policies included. It says nothing about what would happen if that process were changed.
Correct formulation : "The model establishes an association valid under the conditions in which the data were produced. Any decision to intervene on a variable requires a separate causal analysis, with a protocol designed for it."
Symptom : concluding in favor of deployment on the strength of a single headline number, without examining the volume or the nature of the remaining errors.
Why it is wrong : 95% accuracy on 100,000 daily decisions is 5,000 errors per day. Whether deployment is warranted depends on their severity, their detectability, their reversibility and the cost of handling them.
Correct formulation : "The accuracy figure is not sufficient. You must establish the expected error volume, the severity of each error type, who detects them, and at what cost they are put right."
Symptom : shipping a model with no monitoring, treating the project as finished at release.
Why it is wrong : the guarantees of supervised learning assume that production data follow the same distribution as the training data. That assumption stops holding as soon as the users, the product, the regulation or an adversary changes.
Correct formulation : "A model has a shelf life. Going to production includes performance monitoring and a retraining procedure." The subject is covered in chapter 082.
THE REVERSAL
Deterministic programming : Rules + Data -> Answers
Inductive learning : Data + Answers -> RULES
The rule is an INPUT in the first case, an OUTPUT in the second.
MODE OF INFERENCE
Deterministic : deduction, conclusion necessary
Inductive : induction, conclusion probable and revisable
FORMAL FRAME
f target function, unknown, never observed
f^ hypothesis selected by the algorithm, an approximation of f
Gaps : approximation (model), estimation (data), drift (time)
CRITERIA FOR CHOOSING, IN ORDER
1. Is the exact rule known and stable? -> deterministic
2. Is there sufficient labeled history? -> if not, collect it
3. Is an approximate answer acceptable? -> if not, deterministic
4. Is the cost of errors controllable? -> if not, decision support
5. Otherwise -> inductive approach
THE CALCULATION TO RUN EVERY TIME
95% accuracy on 100,000 decisions = 5,000 errors per day
Question : who handles them, at what cost, within what delay?
Most dangerous case : the error that cannot be detected.
THE FOUR STRUCTURAL LIMITS
1. No information is created if it is absent from the data
2. A learned association is not an established cause
3. Stationarity of the relationship is an assumption, not a fact
4. Defective data -> defective model, with no error signal at all
DIVISION OF LABOR
11 stages in the cycle. The machine covers 1:
fitting the parameters. The other 10 are human decisions.Summary statement
Deterministic programming and inductive learning are distinguished by the position of the decision rule: specified as an input in the first case, induced as an output in the second. The inductive approach is warranted only when the exact rule is out of reach, a labeled history exists, an approximate answer is acceptable, and the cost of the residual errors is controlled; within it the machine takes charge of one stage only, fitting the parameters, every other stage being a human decision.
Quizzes and interview questions : 002.1-quiz-paradigm-reversal.md to
002.7-quiz-human-machine-roles.md
Next chapter : 003.0-three-learning-paradigms.md