- Categories of Machine Learning (Supervised/Unsupervised/Reinforcement Learning)
- The Bias / Variance Tradeoff
- Linear Regression and linear classifier (Multi-layer Perceptron)
- Loss function & Optimizers (Gradient Descent and its variants)
- Neural Networks, Convolutional Neural Networks (CNN)
- Why deeper is better than wider?
- Recurrent Neural Networks, GRU, LSTM
- Attention mechanism and Transformer (*)
- What’re the difference between Transformer v.s. CNN architecture? Pros & Cons.
- Vision Transformer (ViT), Detection Transformer (DETR) , Deformable DETR
- CLIP: Contrastive Learning to learn alignment between Language and Image modalities.
- Foundation Models:
- LLM
- VLM
- VLA for robotics/autonomous driving.
我和 ML 的相遇，最初是在新竹家的書桌前，記得應該是大二上結束的寒假，貴人紹齊賢拜分享給我這套最讚的入門影片：台大電機李宏毅教授的『機器學習概論』OCW:
印象很深刻，老師的口條非常好，用寶可夢為例子來教機器學習的概念。
同時間還上了 Andrew Ng 在 Coursera 開設的 Introduction to Neural Networks，以及全球知名的 Stanford CS231n by Fei-fei Li & Justin Johnson。因為 Justin Johnson講話非常快速，很有記憶點，當他在我去密西根大學的那年，開設 Deep Learning for Computer Vision 時我真的驚呆了！覺得怎麼會有這麼巧的事！雖然那年課程全面線上，但能夠上到名師的課還是相當興奮的！
今天要聊的機器學習的種類，以『訓練方式』來分類，可大致分為三大類型：
- Supervised Learning
- Unsupervised Learning
- Reinforcement Learning
- Semi-supervised Learning: First train the model using a small set of training data and ground-truth (supervised learning), then, use the trained model to generate the “pseudo ground-truth” for the large dataset (for those who don’t have GTs). Then, use all training data + (small GT + large pseudo GT) to train the model.
- Self-supervised Learning: GT are from data itself. For example, autoregressive models (next-token prediction): Transformer are trained self-supervised way because the goal is to predict the next word in the sequence.
- Another example is Masked Autoencoder: the goal is to predict the masked pixels, which already present in the original dataset.