Explain CNN, Genetic Algorithms, and Backpropagation

Problem Explain Convolutional Neural Networks, Genetic Algorithms, and Backpropagation — with the actual formulas, not just conceptual descriptions.

Be ready to discuss

• CNNs: convolutional filters slide over the input grid to extract local spatial features with shared weights, pooling downsamples for translation invariance and cheaper compute, and dense layers produce the final prediction. State the discrete convolution and compute output dimensions from input size, kernel, stride, and padding.
• Why convolution beats a fully-connected layer on grid inputs: parameter sharing, local receptive fields, and translation equivariance.
• Genetic algorithms: population-based optimization inspired by natural selection — encode candidate solutions, score them with a fitness function, then select, cross over, and mutate across generations. Cover selection schemes, the exploration/exploitation balance, and when you would reach for one at all (non-differentiable or black-box objectives).
• Backpropagation: gradients of the loss with respect to every weight via the chain rule, propagated backward from output to input, then applied by gradient descent. Be able to derive it for a small network and write the per-layer weight update.
• The mechanics underneath: the forward pass caching activations, local gradients per layer, and where vanishing/exploding gradients originate.
• What's evaluated: exact formulas and derivations on the board — the bar is derivation, not vocabulary.