The companion piece to the agents workshop, aimed one level deeper: a one-hour session where participants train their first neural network.
What participants learn
- What data is: exploring MNIST’s 70,000 real handwritten digits.
- What a neural network is: building a three-layer architecture in TensorFlow.
- What training means: the epochs / loss / accuracy cycle, watched live.
- What a prediction is: classifying digits the model has never seen.
Why MNIST, still
Because it’s the smallest dataset where everything is real: real images, real training dynamics, real mistakes. The point of the hour isn’t the classifier: it’s that “training an AI” stops being magic.