AI Glossary / AI and machine learning / Beginner
Training
Definition
Training is the process of adjusting a machine-learning model using data or experience. It determines the model's learned parameters so that its outputs better satisfy a chosen objective on the examples used for learning.
What it means
A model starts with assumptions, a structure and initial parameter values. Training changes the adjustable parts according to evidence and an objective. In supervised learning, that evidence includes examples paired with target answers.
Other learning settings use different signals. A system may learn structure from unlabeled data or learn behavior from rewards. Training therefore does not always mean a person supplies a correct answer for every example.
Training differs from inference, which uses an already configured model to produce an output. A chatbot answering a question is not necessarily changing its model weights during the conversation.
For business readers, the important distinction is between making a model fit its learning objective and proving that it is useful. Training is one stage of development; evaluation and operational testing establish what the resulting model can reliably do.
How it works
The team defines the task, prepares examples and chooses a model and objective. Data quality, target meaning and appropriate separation from evaluation data are established before fitting.
A learning procedure adjusts parameters. In many neural networks, repeated passes calculate predictions, measure loss and use gradients to update weights. Other algorithms fit trees, estimate distributions or solve a different optimization problem.
Settings such as model size, learning rate and stopping criteria influence the process. Development data helps select these settings and detect when further fitting stops improving generalization.
The team saves the trained model together with enough information to identify its data, configuration and software context. A final evaluation and deployment checks follow. Lower training loss alone does not demonstrate better performance on new examples.
Three practical examples
- Document routing: fit a classifier using messages and reviewed department labels.
- Demand prediction: estimate model parameters from historical sales and relevant inputs, then evaluate on later periods.
- Image recognition: adapt a pretrained model using examples of the product categories a company needs to distinguish.
Think of it this way
Think of adjusting a recipe through repeated trials against a defined taste and texture target. Each trial provides information for the next adjustment.
The analogy has a limit. A learning algorithm optimizes the objective it is given, which may capture only part of what people want. Improving that objective can still produce a result that disappoints users if important qualities were omitted.
Worked example
Separating learning progress from useful progress
Suppose an invented model is trained for ten rounds. Its average training error falls from eight units to four units. Error on a separate development collection falls from nine to six units.
After ten more rounds, training error falls to two, but development error rises to seven. The model fits the training examples more closely while performing worse on the development examples.
The team investigates overfitting, model capacity and stopping choices instead of assuming that more training is automatically better. It selects its approach using development evidence and keeps a separate final assessment.
These figures illustrate the interpretation of learning curves. They do not define a universal stopping rule, because noise, metric choice and the task all affect the pattern.
Illustrative case study
A fictional distributor trains a model on product descriptions from one supplier. Training performance is strong, but descriptions from other suppliers use different abbreviations and formatting.
The team expands representative training coverage, evaluates by supplier and documents the versions used. It also compares the learned approach with a simpler matching baseline.
The lesson is that training learns from the evidence provided. Repeating a narrow collection more often cannot reliably substitute for examples of the conditions the deployed model will encounter.
Common misconceptions and mistakes
- Training means storing every example for later lookup. Some systems memorize aspects of data, but learning methods generally adjust a model representation.
- Every interaction retrains the model. Many applications use fixed model parameters during ordinary use.
- A finished training run means the product is ready. Independent evaluation, integration and monitoring are separate requirements.
Business use
Training matters when adapting a model to a particular task or updating it as conditions change. Its business cost includes data collection, annotation, experimentation and ongoing maintenance, not just computing time.
Ask what objective is optimized, which data is used and how improvement is measured outside that data. Record the model version and its intended scope. A faster training process is useful only if the resulting model meets the quality and operational requirements.
People also ask
Must every business train a model from scratch?
No. Existing models, configuration, retrieval or transfer learning may meet the need with less custom training.
Can training be repeated?
Yes. Retraining may incorporate new evidence, but it should be evaluated rather than assumed to improve the existing model.
Is training deterministic?
Not always. Random sampling, initialization and computing details can affect results, so reproducibility practices and repeated runs may matter.
Check your understanding
Training error keeps falling while development error rises. Is continuing for more rounds clearly the best decision?
Answer: No. The pattern may indicate overfitting or another mismatch. The team should investigate and choose using evidence of performance beyond the training examples.
Related terms
Learn next
Read training set and loss curve. Review one experiment's training and development results together, then identify the evidence used to select the final model.
More in AI and machine learning
- Timestep - A timestep is one position in a sequential calculation.
- Totally Random Trees Embedding - Totally random trees embedding is an unsupervised method that represents an input by the leaves it reaches in a collection of randomly constructed trees.
- Tower - A tower is a substantial neural-network component within a larger neural network.
- Transduction - Transduction is learning or inference directed at a particular supplied collection of cases, rather than producing a general prediction rule for arbitrary...
- Tree-based feature selection - Tree-based feature selection uses a fitted decision tree or tree ensemble to help choose which input features a model should retain.
- Trigram - A trigram is a sequence of three consecutive units, usually words or characters.
Sources
Google, Machine Learning Glossary, training: https://developers.google.com/machine-learning/glossary#training. Core terminology consulted. Learning-curve values and business cases are original examples.
Previous: Trained model Next: Training curve