Introduction to Self-Supervised Learning (SSL)

Technical terms

Supervised Learning: using data with fine-grained human-annotated labels to train AI models.
Semi-supervised Learning: using a small amount of labeled data and a large amount of unlabeled data.

面向对象编程

面向对象主要有四大特性：抽象、封装、继承和多态。
OO开发范式大致为：划分对象→抽象类→将类组织成为层次化结构(继承和合成) →用类与实例进行设计和实现几个阶段。

Introduction

Simply increasing the number of parameters may have a negative impact on the performance.
Question is: how to scale the model effectively?
eg. Switch-Transformer(1.6T parameters)
The core framework is Mixture of Experts.

Sequential decision problems

Deals with fully observable and partially observable cases.
Markovian assumption: Probability depends only on current state $s$ not the history of earlier states.
Markov decision process (MDP): A sequential decision problem for a fully observable stochastic environment with Markovian transition and additive rewards.

This Student Performance dataset includes scores from three exams and a variety of personal, social, and economic factors that have interaction effects upon them.
Data source: http://roycekimmons.com/tools/generated_data/exams

Decision Networks

Essentially, decision networks are extensions of Bayesian networks.
Bayesian networks decrease the number of variables leading to the result.
Decision Networks (aka. influence diagrams) is to combine BN with action and utility nodes to figure out how to get a specific result.

Uncertainty

The real environment is full of uncertainty.
Partial observability(traffic) + non-determinism(car break down)

Utility theory

“How agent can make a rational decision in a random environment?”
Question for machine is what a rational decision is and how to make rational decisions.

Why do we need neural architecture search?

To get higher accuracy, the neural nets are getting deeper and more complicated.
It requires a lot of expert knowledge and takes lots of time and eventually will go beyond human capacity.

Neural Architecture Search (NAS) is a kind of Automated Machine Learning (AutoML). Basic idea is to generate some candidate children networks, train them, optimize their accuracy until successfully select the best architecture for a given task.

Planning Graphs

Search trees grow exponentially large very fast. Depth at which the goal first appears means the minimum number of steps required to achieve the goal.

Forward search (Progression search)

Forward search is not useful without heuristic.
From the current state, we try to apply actions and see what states we can get in this domain until we reach the goal.