What Machine Learning Actually Learns from Data

A common mental model of machine learning is that the model "memorizes the data." That picture is misleading for one important family — fitted parametric models — but it is not wrong for every method. This article focuses on a fitted parametric linear model, the kind you get from ordinary least-squares or logistic regression, to explain what is actually stored and what that implies for your project.

Parameters, not records — for a fitted parametric model

For a fitted parametric linear model, the learned result is a set of numbers: one weight per feature plus an intercept. Training adjusts those parameters so that a function of the inputs produces predictions close to the labels. The official scikit-learn linear model guide describes this fitting process (scikit-learn: linear models). It is tempting to say the model "does not keep a copy of your rows," and for this parametric family that is broadly true after fitting. But it is not a universal property of machine learning. Instance-based methods such as k-nearest neighbours retain the training examples themselves and consult them at prediction time — the data is the model. And even parametric models can memorize information: a sufficiently flexible model can effectively encode individual training rows in its parameters, so parameters do not guarantee privacy. Likewise, a model's size need not be smaller than the dataset; a large tree or network can hold more stored structure than the rows it was trained on.

A worked example: a tiny linear model

Suppose you fit a linear model to predict a track's 14-day streams from two features: tempo (in BPM) and the artist's prior streams. Suppose ordinary least-squares returns the illustrative coefficients: predicted streams = 12 × tempo + 0.001 × prior streams + 200. The parameters are 12, 0.001, and 200 — explicitly illustrative, not measured. For a new track with tempo 120 and prior streams 50,000, the hand calculation is: 12 × 120 = 1,440; 0.001 × 50,000 = 50; plus 200 gives 1,690 predicted streams. You can then run the fitted model on the same input and compare: if the model is wired correctly, it returns the same 1,690. That agreement is the point — the prediction is a deterministic function of the parameters and the input, and you can reproduce it by hand.

Linear does not guarantee generalization

A roughly linear relationship between features and target does not, by itself, guarantee that the model generalizes. Generalization depends on whether the training data is representative of the conditions you will predict on, the level of noise, the sampling process, and honest validation on held-out data. A linear model can fit a genuinely linear relationship well and still fail if the new data comes from a different distribution. Conversely, a nonlinear relationship does not automatically justify a more flexible model: a nonlinear pattern with little data and high noise may be better served by a simple, regularized model than by a flexible one that overfits. The choice of flexibility should follow evidence from validation error, not the assumption that nonlinearity demands complexity.

Do two models on the same data differ?

Whether two fits on the same data differ depends on the algorithm and solver. Ordinary deterministic least-squares has a closed-form solution: given the same data, it returns the same parameters regardless of any random seed. Randomness enters only for algorithms that use stochastic optimization, sampling, or random initialization — some solvers, neural networks, tree methods. For those, fixing seeds and reporting variance across runs matters. A single run's score is one sample from a distribution of possible outcomes only when the algorithm is stochastic.

Data retention follows requirements, not a blanket rule

Whether you may delete training data after fitting is not a technical property of the model. It follows project requirements, consent and data-governance obligations, and reproducibility needs. You may need the data to retrain, audit decisions, validate on fresh splits, or satisfy a retention policy; you may be required to delete it under privacy obligations. Treat retention as a governance decision, not as something the model's smallness permits.

Common mistakes

One mistake is assuming every model discards its training data — instance-based methods do not. Another is treating model parameters as privacy-safe by default; a flexible model can encode sensitive rows. A third is assuming a linear fit generalizes without checking representativeness and validation. A fourth is reaching for a more complex model solely because the relationship looks nonlinear.

An exercise

Fit a small linear model on a toy dataset and print the learned coefficients. Take one new input, compute the prediction by hand using those coefficients, and compare it to the model's output on the same input — they should match. Then check whether your algorithm is stochastic: retrain with a different seed and see if the coefficients change. For the broader learning path, see the Applied AI program previews or the data analysis program.