- ✓The quality of an AI model's training data is often the single most important factor in its performance: models trained on poor-quality, biased or unrepresentative data will produce poor-quality, biased or unreliable outputs regardless of architectural sophistication.
- ✓Overfitting occurs when a model learns the training data too closely, performing well on training examples but failing to generalise to new data: it is one of the most common and consequential problems in machine learning practice.
- ✓Hyperparameter tuning, the process of selecting the optimal configuration settings for a machine learning model, requires a systematic approach such as grid search or random search rather than arbitrary trial and error.
- ✓Cross-validation, the practice of evaluating a model's performance on multiple different splits of the available data, provides a more reliable estimate of how it will perform on genuinely new data than a single train-test split.
- ✓Iterative improvement of an AI system is a structured process that requires clear metrics, controlled experiments and careful documentation: treating it as an engineering discipline rather than an art produces more reliable and reproducible results.
Listen to the full episode inside the course. Enrol to access all 80 episodes, plus assignments, tutor support and Student Finance funding.
Start learning →Alex: Welcome back. Today we're looking at something that's central to working with AI systems professionally: how do you improve them? This is about the iterative work of training, tuning and optimising. Sam, why is this a distinct skill set from building an AI system in the first place?
Sam: Because building a basic AI system, even one that works reasonably well, has become relatively accessible thanks to modern frameworks and pre-trained models. The harder and more valuable skill is knowing how to systematically improve a model's performance, diagnose why it's failing and make the right interventions. That requires a combination of statistical understanding, practical debugging skill and domain knowledge.
Alex: Let's start with training data, because you've mentioned several times that it's often the most important factor.
Sam: The phrase 'garbage in, garbage out' is a cliche but it's true. A model trained on poor-quality data will produce poor-quality outputs regardless of how sophisticated the model architecture is. Data quality issues include missing values, incorrect labels, data that is not representative of the real-world distribution the model will encounter in production, and bias in the data that reflects historical discrimination or unequal representation. Addressing these issues before training, not after, is the most cost-effective approach to improving model quality.
Alex: What about overfitting? I know this is a very common problem.
Sam: Overfitting happens when a model learns the training data too well, capturing noise and specific examples rather than the general underlying patterns. An overfitted model performs excellently on the training data but poorly on new, unseen data. You can detect it by evaluating the model on a held-out validation set that was not used in training: if performance is much better on training data than on validation data, the model is probably overfitting. Remedies include using more training data, applying regularisation techniques that penalise overly complex models, and using dropout in neural networks.
Alex: What are hyperparameters and why does tuning them matter?
Sam: Hyperparameters are configuration settings that control the learning process rather than being learned from data: things like the learning rate, the number of layers in a neural network, the depth of a decision tree. Choosing good hyperparameters can make the difference between a model that barely works and one that is significantly more accurate. Grid search systematically tries every combination of hyperparameter values, random search samples combinations randomly which is often more efficient, and more sophisticated approaches like Bayesian optimisation model the relationship between hyperparameter values and performance to guide the search more intelligently.
Alex: And cross-validation is the standard approach to evaluating model performance honestly?
Sam: Yes. Cross-validation involves splitting your data into multiple folds, training on some folds and evaluating on others, and rotating through so every data point is used for evaluation exactly once. This gives a much more reliable estimate of how the model will perform on genuinely new data than a single train-test split, and it helps you detect whether your performance estimate is sensitive to how the data was split.
Alex: Really technical and practical. Thanks, Sam. We'll look at AI ethics in the next lesson.