Gradient descent: what model training changes
3Blue1Brown connects classification errors with changes to a neural network’s weights and biases.
Follow the downhill direction
This 2017 lesson connects a digit classifier's mistakes to a cost function. Grant Sanderson moves from a landscape analogy to gradients, showing how parameter adjustments can lower training cost. The explanation also distinguishes improvement on training examples from checking unfamiliar examples.
Read a loss curve
Google's separate regression lesson follows repeated loss calculations and parameter updates. Its loss curves help you connect each iteration with training progress and convergence. Use this simpler model to understand the loop before returning to the neural network.
Familiarity with weights and biases helps. This guide uses the creator's written companion.
