Tier 4: Deep Learning

Tier này dạy gì?

MLP & backprop deep: optimizers (SGD/momentum/Adam), batchnorm, weight init, vanishing/exploding.
CNN: convolution, pooling, receptive field; LeNet → AlexNet → ResNet.
RNN/LSTM: sequence modeling, vanishing gradient, gates; seq2seq.
Transformer: self-attention (Q/K/V), multi-head, positional encoding.

Lesson 01

MLP, forward + backprop (chain rule); SGD/momentum/Adam; batch norm; weight init (Xavier, He); vanishing gradient.

Lesson 02

Convolution, pooling, padding, stride; receptive field; classic architectures LeNet → AlexNet → ResNet.

Lesson 03

RNN unrolling, vanishing gradient; LSTM/GRU gates (forget, input, output); sequence-to-sequence.

Lesson 04

Self-attention Q/K/V, scaled dot-product, multi-head; positional encoding; encoder-decoder; mở rộng CLIP (T5-L04).