Neural Networks and Digital Images

Neural Networks and Digital Images#

About the project#

  • Duration: 3-4 hours in class, 1-2 hours preparation at home

  • Prerequisites: Basic linear algebra (vectors, matrices, dot product), basic calculus (derivatives/gradients), basic Python programming

  • Python packages: numpy, matplotlib (optional: scikit-learn, scikit-image, scipy, sympy)

  • Learning objectives: Represent digital images as matrices and vectors, understand and implement the ReLU activation function, compute and interpret gradients, apply the gradient method (gradient descent/ascent) in simple optimization problems, build and visualize shallow ReLU networks and understand why they are piecewise linear, train a neural network to classify handwritten digits using softmax and cross-entropy loss, understand holdout train/test evaluation, design a small ReLU network “by hand” for a geometric classification task*