If the universe is constructed with patterns and humans can interpret this, then so may machines. If machines become intelligent over humans, humans need not apply.
Machine learning is the study of algorithms that involve training a computer to learn patterns and relationships in data without being explicitly programmed. It can be divided into three types:
- Supervised Learning: The algorithm, or model, is trained on labeled data, where the goal is to map each input to the correct output.
- Unsupervised Learning: The algorithm, or model, is trained on unlabeled data, where the goal is to identify patterns or structure in the data without being told what to look for.
- Reinforcement Learning: The algorithm, or agent, is trained on the interactive worlds, where the goal is to learn a policy that maximizes rewards and minimize penalties.
The problems in ML can be largely divided by four types:
- Computer Vision: The problems require machines to interpret and understand visual data, i.e. images and videos. They includes image classification, object localization, video analysis, image captioning and visual question answering.
- Natural Language Processing: The problems require machines to interpret and understand human language, i.e. words and documents. They includes text classification, translation, summarization, speech analysis, question answering and dialogue system.
- Graph Analysis: The problems require machines to interpret and understand graph data, i.e. nodes and edges in graphs. They includes graph classification, link prediction, knowledge graph, social network analysis, recommendation system, combinatorial optimization and path finding.
- Decision Making: The problems require machines to interact with the world and choose an appropriate action based on data and objectives. They include game playing, autonomous vehicles and robotics.
Decision
Reinforcement learning is a subfield of machine learning, popular in decision making problems. The algorithms are generally grouped by:
| Value-based | Policy-based | |||
| Model-based | Model-free | |||
| Classic |
Policy Iteration,
Value Iteration |
Monte Carlo,
SARSA,
Q Learning |
Policy Gradient , Deterministic Policy Gradient | |
| Deep |
Deep Q Network, Double Deep Q Network,
Duel Deep Q Network, Rainbow |
Vanilla Policy Gradient, Natural Policy Gradient,
Trust Region Policy Optimization, Proximal Policy Optimization |
||
| Classic | Actor Critic | |||
| Deep |
Deep Deterministic Policy Gradient, Twin Delayed Deep Deterministic Policy Gradient,
Soft Q Learning, Soft Actor Critic |
|||