MNIST Data

Build a neural-network model from scratch.

← Calculus projects

Learn Mathematics Calculus

Question

How can gradients train a neural network to recognize handwritten digits?

Training adjusts the network weights so a pixel pattern becomes a probability for each digit.
Training adjusts the network weights so a pixel pattern becomes a probability for each digit.

Goals

  • Build the essential parts of a neural network without relying on a machine-learning framework.
  • Use a loss function and gradient-based updates to classify MNIST images.

Method

Treat each image as a vector of pixel values and learn a sequence of transformations.

  • Normalize the MNIST data and implement a feed-forward network.
  • Calculate the loss and propagate its derivatives backward through the network.
  • Train with gradient descent, then examine accuracy and common classification errors.