Loops over data
AI programs often loop over lots of data: for item in data: ... We might count, sum, or check a condition. Here we loop over a list and collect a result—like building a simple 'dataset' of answers.
Appy Says…
AI training is really just a loop. Run predictions, calculate the error, adjust the model slightly to reduce that error, repeat — millions of times. This loop is the engine behind every neural network you've ever heard of.
What is the AI Training Loop?
The training loop is the repetitive process by which a model improves: forward pass → measure error → backward pass (adjust parameters) → repeat.
- •Epoch: one full pass through the entire training dataset
- •Batch: a mini-subset of data processed in each step (not all at once)
- •Loss function: measures how wrong the predictions are (lower = better)
- •Optimiser: algorithm that adjusts model parameters to reduce loss (e.g. Adam)
- •Learning rate: how big each adjustment step is — too high = unstable; too low = slow
- •Gradient descent: the mathematical direction to adjust parameters
Think of it like a Roblox aim trainer
Every shot you miss gives you feedback — too left, too high. You adjust your aim slightly. After thousands of repetitions, your aim improves. The training loop is this cycle automated: predict → measure error → adjust → repeat until error is minimised.
How It Works
- •1. Forward pass: feed input through model → get prediction
- •2. Calculate loss: compare prediction to correct answer using loss function
- •3. Backward pass (backpropagation): calculate how much each parameter contributed to the error
- •4. Update parameters: move each parameter slightly in the direction that reduces loss
- •5. Repeat for all batches in training data = one epoch
- •6. Run for multiple epochs until loss stops decreasing (convergence)
Real-World Examples
- •GPT-4 trained for weeks on thousands of GPUs running this loop trillions of times
- •Image classifier: starts identifying random noise; after 50 epochs, identifies cats accurately
- •Google Translate: trained on billions of sentence pairs, loop ran for months
- •TikTok recommendation: continuous online learning loop — model updates as you watch
Key Facts
- •GPT-3 training cost ~$4.6 million in compute — the loop ran for weeks on thousands of GPUs
- •The Adam optimiser (2014) is the standard training algorithm for most modern neural networks
- •Batch size: 32 or 64 is common — balance between stability and training speed
- •Early stopping: halt training when validation loss stops improving to prevent overfitting
Watch Out!
Learning rate is the most sensitive hyperparameter. Too high: loss spikes and the model diverges (gets worse). Too low: training takes forever and may get stuck in a local minimum. Start with 0.001 (Adam's default) and adjust from there.
Remember
Training loop: forward pass → loss → backward pass → update parameters → repeat. Epochs = full dataset passes. Loss decreasing = model improving.
What You Learned
- •AI training loop: predict → measure loss → adjust parameters → repeat for many epochs
- •Loss function measures error; optimiser (Adam) reduces it; learning rate controls step size
- •Unlocks: understanding how every neural network from image classifiers to LLMs is trained
Key Facts
- →GPT-3 training cost ~$4.6 million in compute — the loop ran for weeks on thousands of GPUs
- →The Adam optimiser (2014) is the standard training algorithm for most modern neural networks
- →Batch size: 32 or 64 is common — balance between stability and training speed
- →Early stopping: halt training when validation loss stops improving to prevent overfitting
Real-World Examples
Remember
Training loop: forward pass → loss → backward pass → update parameters → repeat. Epochs = full dataset passes. Loss decreasing = model improving.
Quick Quiz
Why do we loop over data?