Skip to content


Dave Touretzky edited this page May 20, 2019 · 17 revisions
  • Agent: any artificial intelligence program that interacts with the world via a sense-deliberate-act cycle (sometimes called "sense-think-act"). Agents may be physical devices such as robots, or exist purely as software, such as automated stock-trading programs.

  • Artificial intelligence (AI): a body of techniques that allow computers to do things that, when humans do them, are considered evidence of intelligence.

  • Backpropagation: a supervised learning algorithm for training neural networks. For each training example, the network's actual output is compared with the desired output; the difference constitutes an error signal. This signal is propagated backwards from the output layer to earlier layers and used to adjust their weights so as to move the actual output slightly closer to the desired output.

  • Chatbot: a conversational agent that communicates with humans via a text or voice interface. The term is derived from "chatter" + "robot". Chatbots are a promising technology for online customer service applications where they help people find information about products or services. The simplest chatbots just search for keywords and return scripted replies, but more sophisticated chatbots may fall in the category of intelligent assistants.

  • Classifier: an algorithm that examines the features of an input pattern and assigns that input to a category (or "class"). Classifiers have many practical applications, such as spam filtering (using the classes "spam" and "not spam"), or sorting loan applications, where the classes might be "low", "medium", or "high" risk. Classifiers can be programmed by hand, but it is more common to train them using machine learning techniques and examples of each class.

  • Machine learning: A subfield of Artificial Intelligence focusing on techniques that improve a computer's performance on tasks by analyzing training data. Three important types of machine learning are supervised learning, unsupervised learning, and reinforcement learning. Modern AI applications, such as speech recognition systems, are often constructed using machine learning techniques applied to huge training sets.

  • Natural language: the language people use to communicate with each other in everyday life. This is distinct from computer languages used for programming or for giving commands to a computer. Computer languages have a rigid vocabulary and syntax, while natural language is fluid and ambiguous, and includes phenomena such as imagery, humor, sarcasm, and alliteration. Natural language understanding is difficult for computers.

  • Neural network: an approach to computing in which many simple processing units are organized into a network to collectively solve a complex problem. Neural networks draw inspiration from theories about how the brain might work, but their components and organizational structure are far simpler than real neural tissue. Neural networks have contributed much to recent progress in AI due to their incorporation of powerful learning algorithms such as backpropagation.

  • Perception: The extraction of meaning from sensory signals. A microphone senses sound and converts it to an electrical signal. Understanding the sound, e.g., recognizing the words being spoken or the music being played, constitutes perception.

  • Reinforcement learning: a type of machine learning where the training data is labeled with a "reward" signal telling the computer how good its outputs have been. This differs from supervised learning where each training example's label tells the computer exactly what output it should have produced. Reinforcement learning is typically applied to sequential tasks where the reward signal comes only at the end. An example is game playing, where the reward signal comes after the final move and tells the computer whether it won or lost the game. Using reinforcement learning, computers have become experts in domains such as backgammon by playing against themselves for thousands of games.

  • Supervised learning: a type of machine learning where the training data consists of input examples plus the desired output for each input. The desired outputs are called "labels", and the data is called "labeled data". For example, the training data could be pictures, some of which contain cats, and the labels could be 1 for "cat" or 0 for "no cat". The learning algorithm goes through the training set repeatedly, and slowly adjusts the model's parameters to make it more likely to produce the correct output for each input.

  • Turing test: a thought experiment proposed by the mathematician Alan Turing to justify the claim that a machine could be said to possess "intelligence". A human interrogator engages in dialog via teletype machine with both another human being and a computer posing as a human being. Turing proposed that if the interrogator cannot reliably distinguish the machine from the human, the machine can be said to be intelligent. (Wikipedia article)

  • Unsupervised learning: a machine learning technique that finds structure in unlabeled data. One example is clustering, where the computer examines a collection of examples and groups them into categories based on perceived similarity. In a sense, the computer "discovers" the categories; they are not given to it. (If each example came labeled with its correct category, this would be supervised learning.)

To be added: bias, fairness. classification, reasoning, representation, transparency, machine translation, feature vector, logic, search.

You can’t perform that action at this time.