8. Learning from nature

Natural computation

• Algorithms derived from observations of nature phenomena

• Simulations
  • to learn more about these phenomena
  • to learn new ways to solve computational problems

Natural computation and Machine learning

What is learning?

• The ability to improve over time, based on experience

• Why? : Solutions to problems are not always programmable

• Examples
  • Handwritten character recognition
  • Adaptive control of production processes
  • Learning to walk by trial and error

Three forms of learning

Techniques

1. Artificial neural networks (ANNs)

   → Inspired by biological nervous systems

2. Reinforcement Learning

   → Inspired by psychology, ethology and behaviourism

   ex : Menace, Q-Learning

3. Evolutionary Computing

   → Inspired by genetics, natural selection and evolution

   ex : Genetic algorithms, Genetic programming

4. Swarm Intelligence

   → Inspired by social animals (bird flocks, ants, etc)

   ex : Particle Swarm Optimization, Ant colony optimization, Cellular automata

Artificial neural networks

• Invented in the 1940’s

• Many simple processing elements, operating in parallel and communicating through weighted connections

• Based on very simple models of biological neurons and synaptic connections

• store information in the weights, not in the nodes

• can be concurrent

• fault tolerant

Deep Learning

Any neural network with more than one hidden layer. Nowadays, usually much deeper, and with different types of layers. For example CNNs

More layers → more levels of abstraction → automatic feature selection

💡

CNN 논문에서 진행한 실험 생각해주면 된다. 각 layer마다 담당하는 층위가 나뉜다.

Reinforcement learning

• Reward : an evaluation of the environmental state

• Goal : To make decisions that maximize the long term reward received by the agent

• The agnet must be allowed to explore (i.e. sometimes do actions that at the time seem sub-optimal)

Evolutionary computing

• Used for learning problems where the task is to maximize some measure of success

• Essentially the same family of problems as in reinforcement learning, but the methods are different

• Methods inspired by genetics, natural selection and evolution

Swarm Intelligence

• Bird flocks and fish schools move in a coordinated way, but there is no coordinator

• Ants and termites quickly find the shortest path between the nest and a food source

• Distributed systems without central control

• Useful not only to simulate but also to solve optimization problems

Bird flocks and fish schools

• Local interaction

• No leader

• Simple local rules - a weighted combination of several goals

• Sufficient to make very realistic simulations
  • used in movies and computer graphics
  • remove the match-velocity result : insect swarm
  • remove collision rule : cultural interaction

Particle swarm optimization

• Originally intended to simulate bird flocks and to model social interaction

  → But stands on its own as an optimization tool