How A Machine Learning Model Is Built: Data, Training, Parameters And Inference

People often talk about “the AI” as a single thing, but a machine learning model goes through distinct stages, and each one creates different risks. Bad data at the start, a flawed test in the middle or no monitoring at the end can each cause a model to fail in use.

This article follows one simple example, a supervised model that predicts whether an invoice is a duplicate, through its whole life. Other kinds of model, covered in Three Ways Machines Learn, differ in the details, but the same broad stages apply. Along the way it explains four words that come up in every AI conversation: data, training, parameters and inference.

A Model Is A Function With Adjustable Numbers

At its core, a model is a mathematical function. It takes an input, such as the supplier, amount, date and reference number of an invoice, and produces an output, such as a score from 0 to 1 for how likely the invoice is to be a duplicate.

What makes it a learned model is that the function contains adjustable numbers called parameters, such as weights and a bias. In the common approach Google’s course teaches, these start as small random values, so the model’s first answers are no better than guesses, and training nudges them until the answers become useful.¹ The simplest models have only a handful of parameters. Large language models have far more: GPT-3, described in 2020, had 175 billion.²

The Life Of A Model

Data: The Raw Material

The patterns a model learns come from its data. For the invoice example, that means a history of invoices with a label saying which were duplicates. If the history only covers one country, or if past staff missed many duplicates and labelled them wrongly, the model will learn those blind spots.

Before training, the data is divided into three parts. The training set is what the model learns from. The validation set is used during development to compare versions of the model and tune settings. The test set is held back for a final, independent check. Google’s machine learning course warns that the more often the same test set guides decisions, the more the model ends up tuned to that particular set, which weakens what its score says about real data.³

Training: Reducing The Error

For models like this one, training usually works in a loop called gradient descent. The model makes predictions on a batch of training examples, and a loss function measures how far those predictions are from the correct labels. The algorithm then works out which direction to move each parameter to make the loss smaller, moves it a small step that way, and repeats until the loss stops improving.¹ Not every method trains this way; decision trees, for example, are built by repeatedly splitting the data instead.

The developer chooses some settings that training itself does not learn, such as the learning rate (how large each step is), the batch size and the number of passes through the data. These are called hyperparameters, to tell them apart from the parameters the model learns, and the validation set is where they are tuned.⁴

Inference: Using The Trained Model

Once training ends, the parameters are frozen and the model is deployed. Running it on new input is called inference. Each new invoice goes in and a duplicate score comes out. No learning happens at this stage unless the system is specifically built to keep updating.

This split matters for security and privacy. NIST lists data poisoning and the extraction of models or training data through a system’s interfaces among common security concerns for AI.⁵ Training data can be tampered with before the model learns from it, and personal data in it may later be exposed. Inference is where a deployed model meets real users and real attackers. Both kinds of attack are covered in AI Security.

After Deployment: Monitoring And Retraining

The world changes after a model ships. New suppliers appear, invoice formats change, and fraudsters adapt. When live data starts to look different from the training data, accuracy can fall without anyone noticing. This is usually called drift, and NIST lists data, model and concept drift among the reasons AI systems may need more frequent maintenance than ordinary software.⁵

NIST’s AI Risk Management Framework says AI systems should be tested before deployment and regularly while in operation.⁵ In practice that means tracking the model’s accuracy on recent, correctly labelled cases, alerting when it drops, and retraining on fresh data. The lifecycle is a loop, not a straight line.