
Already know some of this? A 2-minute test skips the lessons you know.
Take the placement testA free 11-lesson course built from Zachary Huang's 1h 16m of YouTube video. 1 day at 90 minutes a day, with study notes and a quiz in every lesson. A 2-minute placement test skips what you already know.
Taught by Zachary Huang on YouTube. Videos play in YouTube's own player, so the creator keeps every view.
Explain how a single perceptron learns a linear decision boundary using weighted sums and a step activation, and why it cannot solve non-linearly separable problems like XOR.
Explain how convolutional layers use small, shared filters to detect patterns (edges, textures) anywhere in an image with far fewer parameters than an MLP, and how ReLU and pooling build hierarchical feature maps.
Explain why stacking more CNN layers hurts performance (degradation/vanishing gradients) and how ResNet's residual blocks with identity skip connections make deep networks easy to train.
Explain how a Recurrent Neural Network processes tokens one by one while updating a fixed-size hidden state, allowing it to handle variable-length sequences but suffering from vanishing gradients over long distances.
Contrast LSTM's separate long-term/short-term memory with GRU's merged gated memory, and explain how both preserve gradients across many time steps to capture long-range context.
Describe the encoder-decoder architecture for sequence-to-sequence tasks, the bottleneck of a single context vector, and how attention lets the decoder weigh all encoder states at each step.
Explain how self-attention lets every token attend to all others in parallel, why positional encodings are required to preserve order, and how the Transformer block stacks attention, MLP, skips, and LayerNorm.
Describe BERT's encoder-only Transformer, masked language model pre-training, and how the [CLS] token enables fine-tuning for classification, QA, and NER with minimal labeled data.
Contrast GPT's causal (look-back only) self-attention with BERT's bidirectional attention, and explain how next-token prediction on massive text enables zero-shot/few-shot task generalization via prompting.
Explain how Mixture of Experts replaces the dense feed-forward MLP with multiple expert MLPs and a router that activates only a subset per token, increasing capacity while keeping compute constant.
Describe how Mamba (State Space Model) uses a selection mechanism and convolution-like recurrence to maintain a fixed-size memory, enabling linear-time scaling with sequence length and parallel training.