Understand Artificial Intelligence - ECE469 Course Overview

ECE469: Artificial Intelligence Course Introduction

The Basics Lectures will be in person in Rm. 502 Wednesdays from 4 pm 5:50 pm Thursdays from 4 pm 4:50 pm If we are forced to be remote for any lectures, we will do those lectures using Teams My Cooper webpage: http://faculty.cooper.edu/sable2 From there, you can find a link to the course webpage I ll post syllabus info, slides, and projects on the course webpage E-mail: carl.sable@cooper.edu Textbook (recommended, not required): "Artificial Intelligence: A Modern Approach", by Stuart Russell and Peter Norvig I would recommend getting the fourth (newest) edition The book has a nice associated website: http://aima.cs.berkeley.edu

Grading Project #1: 33 1/3% You will implement a game-playing program that plays either Checkers or Othello; I consider this a very tough project The project will be evaluated based on how well it plays against me! Project #2: 33 1/3% You will implement a neural network from scratch and create or modify a dataset to train and test it The project will be evaluated based on correctness (not efficiency, as long as it is reasonable) and on the dataset that you create Three in-person quizzes (closed book, closed notes): 33 1/3% The projects can be done using any programming language, as long as I can run it, but for the first project, efficiency matters

What does artificial intelligence artificial intelligence (AI AI) even mean, anyway? There is no agreed upon answer, but let s discuss it.

Course Content The course will be divided into six parts: 1. Introductory Concepts 2. Search and Games 3. Dealing with Uncertainty 4. Machine Learning 5. Natural Language Processing 6. Philosophy and Ethics of AI Each part except the final one consists of multiple topics; you can check out the course schedule for more details

Other AI Topics (not covered in the course) Speech recognition Robotics Knowledge representation Planning Computer art/music Computer vision

Four Approaches to AI The book mentions four possible approaches to AI, classified on two dimensions (acting vs. thinking and human vs. rational) Thus, the goals of the four approaches (which we will expand upon over the next four slides) could be described as: Thinking humanly Thinking rationally Acting humanly Acting rationally

Thinking Humanly The book refers to this as the cognitive modeling approach Goal: Create a sufficiently precise theory of the mind and then express the theory with a computer program We can learn about human thought through introspection, psychological experiments, and brain imaging Example: In the late 1950s, Newell and Simon created the General Problem Solver (GPS) The goal was to build a universal problem solver that could solve formalized symbolic problems GPS could only solve simple problems, not real-world problems GPS later became the basis for SOAR, which is still a major ongoing research effort Example: The Center for Brains, Minds, and Machines (CBMM) involves collaboration between MIT, Harvard, and other institutions Their website says that the CBMM is "dedicated to the study of intelligence - how the brain produces intelligent behavior and how we may be able to replicate intelligence in machines" This Center includes experts in computer science (especially AI and ML), robotics, psychology, and neuroscience, all working together with a common goal

Thinking Rationally The book refers to this as the "laws of thought" approach This dates back at least to Aristotle's syllogisms (involving his attempt to codify correct thinking) These rules initiated the field called logic Two obstacles: Stating a problem formally (especially when uncertainty is involved) Implementing a solution (simple methods can exhaust computational resources, and unsolvable problems might result in infinite loops)

Acting Humanly The book refers to this as the Turing test approach Turing proposed the Turing test in 1950 in a paper titled "Computing Machinery and Intelligence" To pass the Turing test would (presumably) require at least the following capabilities: Natural language processing (NLP) Knowledge representation (KR) Reasoning Machine learning (ML) Turing argued that a machine that passes the Turing test would actually be intelligent; however, this is highly debatable Our textbook updates the notion of the Turing test to also include vision processing and robotics; they call this the total Turing test The CBMM talks about Turing++ questions related to images We will discuss the Turing test and whether it is a valid test of intelligence in greater detail during our final topic of the course

Acting rationally The book refers to this as the rational agent approach An agent is something that acts; examples of agents include machines, computer programs, and people A rational agent acts to achieve the best expected outcome All the skills needed for the Turing test can aid rational agents; for example: knowledge and planning help one make good decisions communication enables one to express thoughts and interact with others learning enables one to improve over time This is the approach favored by our textbook and most AI practitioners Note that humans are not rational agents according to the above definition

Disciplines that have Contributed to AI Philosophy Mathematics Economics Neuroscience Psychology Computer engineering Control theory Linguistics

Early AI Milestones 1943: McCulloch and Pitts proposed a model for artificial neurons 1950: Turing s "Computing Machinery and Intelligence" 1951: Minsky and Edmonds built the first neural network (out of vacuum tubes, motors, and clutches); it simulated 40 neurons 1956: At a workshop at Dartmouth, Newell and Simon introduced the Logic Theorist, a program that could prove mathematical theorems At the same workshop, there was agreement on the term "artificial intelligence" 1958: McCarthy invented LISP and described the hypothetical Advice Taker, a complete AI system embodying the principles of KR and reasoning

AI in the 1960s Successes with Rosenblatt s perceptrons (simple neural networks without hidden layers) Successes with microworlds (problems dealing with very specific, limited domains) Example: Winograd's SHRDLU seems to understand language about the "blocks world" In 1969, Minsky and Papert proved a fundamental limitations of perceptrons; they can only recognize linearly separable data

AI in the 70s Expert systems included a domain-specific knowledge base and a problem- solving engine The knowledge base often relied on hard-coded, domain-specific facts constructed with the aid of human experts The problem-solving engine of an expert system would generally be simple (e.g., a chain or tree of if-then-else rules) Examples included DENDRAL (which predicts molecular structure given information from a mass spectrometer) and MYCIN (a medical diagnosis system) Schank emphasized cased-based reasoning to find solutions to new problems by modifying solutions to similar problems seen in the past Minsky proposed using frames, which assumes facts about particular objects and event types All of this has been out of favor for some time now

The Comeback of Machine Learning In the mid-1980s, several groups "rediscovered" backpropagation for training neural networks with hidden layers Other methods of machine learning (e.g., Bayesian methods, support vector machines, etc.) outperformed neural networks in the 1990s The creation of the Web helped lead to much larger datasets to train machine learning systems After 2000, "big data" led to even better performance

Deep Learning Over the past decade, deep learning has dominated machine learning (and arguably AI in general) Convolutional neural networks led to breakthroughs in computer vision Variations of recurrent neural networks (such as LSTMs) and later transformers led to breakthroughs in natural language processing Deep learning along with reinforcement learning has led to breakthroughs in games

Some Notable AI Accomplishments Games (e.g., Deep Blue beat Kasparov in chess, Checkers has been weakly solved, recent achievements in Go and poker, NPCs inhabit video game worlds) Machine learning applications (e.g., speech recognition, OCR, bioinformatics, data mining, successes involving deep neural networks) Classification (e.g., medical diagnosis, Hit Song Science, object recognition in images); this is really a subcategory of machine learning NLP applications (e.g., spam filters, information retrieval; recent successes involving deep learning and large language models) Google (relies on NLP and information retrieval algorithms, but also uses an advanced PageRank algorithm that considers the web to be a graph) Robotics (e.g., Mars rovers, Roomba vacuum cleaners, self-driving cars) Recommendation engines (e.g., Netflix for movies, Pandora for music)