Deep Learning
Neural networks and the architectures that power modern AI.
- Neurons, Layers, and ActivationsBeginner
A neuron is affine (Wx+b) then a nonlinearity. Stack layers; without ReLU/GELU the whole net is still one linear map.
- What are CNNs?Intermediate
CNNs slide shared filters over images so the same detector works everywhere. Pooling downsamples. Still used in 2026 vision — not SOTA for language.
- What are RNNs?Intermediate
An RNN folds a sequence into a hidden state, one step at a time. Sequential by design, historically crucial, limited by vanishing gradients.
- LSTMs and GRUsIntermediate
LSTMs and GRUs add gates so memory survives longer than a vanilla RNN. Useful in some time series; transformers took NLP.
- What are Autoencoders?Intermediate
An autoencoder squeezes input through a bottleneck and reconstructs it. The code is an embedding; denoising and VAEs are the useful cousins.
- What are GANs?Intermediate
A GAN pits a generator against a discriminator. Adversarial training still teaches; in 2026 diffusion and flow models dominate image generation.
- Why Transformers Replaced RNNsIntermediate
Transformers beat RNNs because attention is parallel and any two tokens are one hop apart. RNNs remain in a few sequential niches.