IAI | HSLU | Jana Köhler

Algorithm processes example input–output pairs and learns a function that maps from input to output, e.g., decision trees and neural networks

Algorithm learns from a series of reinforcements (rewards or punishments)

returned from the environment when the agent executes actions, e.g., AlphaZero

•  Neuron fires when a linear combination of its inputs exceeds predefined threshold

• The bias weight (multiplied with input 1.0) is a constant to shift the result of the activation function towards the positive or negative side (offset the result)

Feed-forward networks are directed acyclic graphs, they have no internal state other than the weights

– Single-layer perceptron 

– Multi-layer perceptron

Consider a perceptron with a step function
– Can represent AND, OR, NOT, majority, etc., but not XOR
– Represents a linear separator in the input space
– Only a small fraction of all Boolean functions is linearly separable

• Input and output layers are usually fully connected via at least one hidden layer 

• Network architecture (number of hidden layers/edges typically chosen by hand)

  Description GPT:

• A Multi-Layer Perceptron (MLP) or Feed-forward Network is a widely used neural network model.
• It consists of at least three layers: an input layer, at least one hidden layer, and an output layer.
• These layers are typically fully connected, with each unit in one layer connected to every unit in the next layer.
• Data flows in one direction, from input to output, without feedback loops.

A Feed-Forward Network is a type of artificial neural network where data flows in one direction, from input to output, without feedback loops. To represent any Boolean function, at least one hidden layer is needed in the network.

A single, sufficiently large hidden layer in a Feed-Forward Network can accurately represent any continuous function of the inputs with arbitrary accuracy. It acts as a universal approximator for continuous functions.

A Feed-Forward Network with two hidden layers can represent discontinuous functions. The added complexity and depth of the network make it capable of modeling and approximating functions with discontinuities.

Bild:

In Reinforcement Learning, the "environment" describes the physical world in which the agent operates. It is the system that the agent interacts with and seeks to learn from.

"State" in Reinforcement Learning describes the current situation or configuration of the agent within the environment. It provides information about the context in which the agent is making decisions.

 In Reinforcement Learning, "reward" provides feedback from the environment to the agent. It serves as a numerical signal that indicates how good or bad the agent's actions are in a given state, guiding the learning process.

 An "action" in Reinforcement Learning is a decision or choice made by the agent that leads to a transition to the next state within the environment. It drives the agent's interaction with the world.

The primary goal of Reinforcement Learning is to learn a policy. A policy is a method or strategy that maps states to actions in a way that maximizes the future reward of the agent. The agent aims to make decisions that lead to the highest possible cumulative reward over time.

• The attempt to make computers more "intelligent"?
• The attempt to better understand human intelligence?
• Four possible orientations of AI research:

– Is it about thinking or acting?
– Using a human model with all its defects or a normative model using rational guidelines?

• What cognitive capabilities are necessary to produce intelligent performance? 
• Not important: Being able to solve problems correctly
• Important: Being able to solve problems like a human would

 Cognitive science and cognitive psychology: 

-Introspection (observing ourselves)

-Psychological experiments

-Brain research

"The Turing Test" is a test of a machine's ability to exhibit human-like intelligence. It is described in terms of a game known as the "imitation game." In this game, three participants are involved: a man (A), a woman (B), and an interrogator (C). The interrogator stays in a separate room from the other two participants. The goal of the interrogator is to determine which of the other two participants is the man and which is the woman.

• It is A's objective to try and deceive the interrogator into making the wrong identification.
• B's objective is to assist the interrogator.
• The game raises the question of what happens when a machine takes the role of participant A, attempting to convince the interrogator that it is human, effectively replacing the original question of "Can machines think?"

This test, proposed by Alan Turing, seeks to distinguish between the physical and intellectual capacities of a machine by examining its ability to mimic human behavior.

The question that motivated research in artificial intelligence in the 1950s was whether it was possible to create machines that could exhibit human-like intelligence. In other words, "Can machines think?"

A symbolic representation refers to representing information using symbols, such as words or symbols in a language. Subsymbolic representations involve representing information at a lower level, often involving patterns and statistical relationships.

The orientations of AI research involve whether AI is focused on thinking or acting and whether it uses human-like models or normative, rational guidelines for decision-making. These orientations shape the goals and approaches in AI research.

The adequacy of the Turing Test as a test of intelligence is a subject of ongoing debate. While it's a significant milestone in AI history, it has limitations, as it primarily evaluates a machine's ability to mimic human behavior and not necessarily its underlying intelligence.

“Big data” is high-volume, -velocity and -variety information assets that demand

->cost-effective, innovative forms of information processing 

->for enhanced insight and decision making