Machine Learning
支持向量机 (Support Vector Machine)
反向传播算法 (backward propagation, BP)
Backpropagation, Gradient Descent, - Chain Rule - Forward pass - Backward pass
ChatGPT Principle
ChatGPT 的原理,就是文字接龙(语言模型),通过输入,输出一个机率分布(Probability distribution),然后随机 取样(sample)一个字,加到输入后面,然后再输入到模型中。
GPT 使用预训练(Pre-train), 或者叫自监督学习(Self-supervised learning), 生成的模型叫做基石模型(Foundation model)。 预训练作用:多语言习得,在一种语言上训练,其他语言自动学习。
微调(Finetune),在基石模型(Foundation model)上继续学习。
监督学习(Supervised Learning),就是提供成对的学习资料。
强化学习(Reinforcement Learning (RL)): 我们在某些任务上不知道如何标注数据,但是我们可以提供判断结果的好坏。(It is challenging to label data in some tasks. We can know the results are good or not.)
Professional Words
- Machine Learning ~= Looking for Function
- Deep Learning
- Neural Network
- Downstream tasks
- Generative Adversarial Network
- Anomaly Detection
- Explainable AI
- Model Attack
- Domain Adaptation
- Network Compression
- Life-long Learning
- Meta Learning = Learn to Learn
- Few-shot learning