The ML Pipeline, End to End

This chapter pulls every piece from this module together into one connected picture: the full pipeline a model travels through before, during, and after it reaches production.

• -Data collection gathers data from the sources covered earlier: logged, licensed, scraped, or synthetic.
• -Cleaning and labeling fixes quality issues and attaches correct answers, for supervised tasks.
• -Splitting divides the data into train, validation, and test sets before any modeling decisions are made.
• -Training runs the algorithm, gradient descent and backpropagation for neural networks, to fit the training set.
• -Evaluation measures performance on the validation set repeatedly during development, then the test set once at the end.
• -Deployment exposes the trained model behind some interface, an API or an embedded component, so a real product can use its predictions.
• -Monitoring continuously watches the deployed model's real-world performance, since production behavior can quietly diverge from what testing showed.

One detail hides inside the "training" and "evaluation" stages above that deserves its own spotlight. Every algorithm in this course has settings you choose before training even begins, and those settings are not learned by gradient descent the way weights are. K in KNN, the number of trees in a forest, the regularization strength λ from the next module, the learning rate itself. These are called hyperparameters, to distinguish them from a model's parameters, which training does learn.

Choosing good hyperparameters is itself a search problem, and it connects directly back to the bias-variance tradeoff the next module introduces. A hyperparameter like K in KNN or the depth of a tree directly controls model complexity, so tuning it really means searching for the sweet spot between underfitting and overfitting.

• -Grid search tries every combination from a predefined set of values for each hyperparameter. Exhaustive and simple, but the number of combinations multiplies fast: V values across H hyperparameters means Vᴴ combinations to try.
• -Random search samples combinations randomly from the same space instead of trying every one. Counterintuitively, it often finds a good combination faster than grid search, especially when only a few of the hyperparameters actually matter much.

A model that performed beautifully on its test set can still degrade in production, since the real world keeps changing after the test set was collected. Costs shift, user behavior shifts, the underlying population being predicted shifts. This phenomenon, often called model drift, is why monitoring is listed as a genuine pipeline stage rather than an afterthought. A deployed model needs the same kind of ongoing attention a deployed traditional service does, just watching different signals.

This end-to-end view is also exactly the mental model worth carrying into Phase 2. Every concern in this pipeline, data quality, evaluation discipline, hyperparameter tuning, monitoring, drift, reappears in the LLMOps and production modules later in this course, just applied specifically to large language model systems instead of classical ML.