2009 AGI Summer School
List of Lecture Topics and
Background Readings
Each session is 1.5
hours. Each lecture is 1 session
unless otherwise specified.
Many lectures specify
readings along with them. Some
readings are explicitly specified as optional, mostly not because theyÕre
peripheral but just because theyÕre long.
The readings list given here is very partial and will be completed as
the Summer School approaches.
Note that a number of
lectures touch on the theory of Probabilistic Logic Networks, and so the book
will be made available to
registered students. But this will
be considered optional reading.
Lectures are grouped here by
the lecturer. See the summer
school schedule for information on the temporal ordering of the lectures.
Dr. Joscha Bach
Readings
TBA
Man
as Machine
The
traditional (western) philosophy of mind has traditional difficulty with
understanding the human mind in the context of natural sciences. Instead, a
subjectivist, hermeneutical perspective is preferred. Here, we will discuss how
the mind can be studied as an information processing system; we will look at
different approaches and focus on the appropriate level of abstraction. We will
see if and how Artificial Intelligence can provide us with a deeper
understanding of human cognition, and we will confront several of the
counter-arguments.
Understanding
motivation, emotion and mental representation using computational models (2
sessions)
Cognition
goes beyond problem solving. We will look at the question of how a cognitive
agent arrives at problems to solve, i.e., we will ask: how does motivation
work? How are motivation and emotion related? And how can mental
representations refer to real-world events and objects? These questions are
fundamental to understanding and realizing an artificial intelligent system,
and we will discuss how we can build computational systems capable of these
feats.
The
MicroPSI architecture
MicroPSI is a cognitive
architecture intended to model the integration of reasoning, perception, memory
and motivation. It is implemented in Java and uses a neurosymbolic
representation for computational agents, that can either be situated in
simulation environments, or that can be used to control robots.
Dr. Allan Combs
Readings
TBA
The
Brain as a Neuroscientist Sees It (2 sessions)
An introduction to how
cognitive neuroscientists think about intelligence and consciousness in the brain,
including an introduction to brain anatomy as well as cellular level
neurodynamics and their implications for the nature of intelligence.
The Nature of
Consciousness
A multidisciplinary
review of the phenomenon of consciousness, covering neuroscience, philosophy,
AI, nonlinear dynamics and other perspectives.
Nonlinear Dynamics
and the Mind
An overview of nonlinear
dynamics and chaos and their relation to brain science and the nature of
intelligence. Discussion of
implications for artificial intelligence.
Dr. Hugo de Garis
Evolvable Neural
Networks
Application of neural net
evolution to learning components of intelligent systems, including object
identification, face recognition, movement learning and other abiitiies.
Humanoid Robotics for
AGI
A Review of the Nao humanoid
robotics platform and its use for AGI
Dr. Nil Geisweiller
Readings:
For probabilistic logic
lectures (materials to be supplied to students in advance):
Introduction
to Probabilistic Logic Networks (with Ben Goertzel)
The
basic principles of the Probabilistic Logic Networks reasoning framework will
be presented. The Òindefinite
probabilitiesÓ framework will be briefly reviewed. The basic PLN inference rules will be described, including
rules for first order and higher order inference.
Probabilistic
Logic: Spatial, Temporal and Intensional Inference
We
will show how to utilize existing spatial and temporal reasoning methods, such
as the Region Connection Calculus and Allen's Interval Algebra, to carry out
spatial and temporal inference in Probabilistic Logic Networks and other
probabilistic logic systems.
We will explain how intensional inheritance (inheritance in terms of
patterns and properties rather than subsets) works in the Probabilistic Logic
Networks framework, and give a few supportive examples.
Readings
Program
Representation for General Intelligence (with Ben Goertzel)
We
will discuss the representation of programs in a manner compatible with
automated program learning, including issues regarding the transformation of
specailized and general programs into hierarchical normal forms in which
syntactic and semantic properties are well-correlated.
Readings
Competent
Program Evolution Using Probabilistic Modeling Based Evolution, (The MOSES
Algorithm)
We
will explain the different parts that compose MOSES and how they work: demes,
normalization, probabilistic modeling and representation building. We will show
different examples and explain them, and possibly conclude by presenting some
potential variations and improvements.
Readings
Imitation
and Reinforcement Learning in Virtually Embodied Agents Using Program Evolution
The
use of hillclimbing and MOSES to learn programs controlling virtual agents in
online virtual worlds is described.
Specific examples are given involving the learning of programs
controlling pets in the Multiverse and OpenSim virtual worlds.
Readings
Dr. Ben Goertzel
AGI versus Narrow AI
A review of the history of
the AI field, and the foundational theory of intelligence, culminating in the
clarification of the distinction between Artificial General Intelligence and
task-focused Ònarrow AI.Ó Overview
of current software systems aimed at AGI, including OpenCogPrime, NARS, SOAR,
LIDA and ACT-R.
Readings
Review of Past and Present
AGI Research, at http://www.agi-08.org/conference/
(optional)
Mathematics
of Universal and General Intelligence (1/2 session)
This
lecture briefly reviews universal machine learning theories, including
Solomonoff's Algorithimic Probability Theory, Hutter's AIXI algorithm and
Schimdhuber's Goedel Machine.
Relevance of these theories to pragmatic general intelligence is also
covered.
Readings
The OpenCog Prime
Design for AGI (1/2 session)
Overview of OpenCogPrime, a
specific architecture for general intelligence, incorporating PLN probabilistic
inference, MOSES evolutionary program learning, ECAN economic attention
allocation and other aspects, and designed for implementation within the
integrative OpenCog framework.
Readings:
Natural Language
Processing for AGI
A review of current methods
in computational natural language understanding and generation, how they relate
to human language processing, and current ideas regarding how they might be
modified in order to achieve generally-intelligent language processing. Hands-on examples will be given using
the RelEx and link-parser NLP systems.
Readings:
Customizing Virtual
Worlds for AGI (1/2 session)
Review of current virtual
world technology, with a focus on the improvements that must be made to it in
order to make it truly adequate for AGI development, including integration with
robot simulators and expansion of physics engines to include bead physics.
Readings:
Dr. Matthew Ikle
Managing Uncertainty
with Indefinite Probabilities
A review of methods for managing uncertainty in AI
systems, including fuzzy set theory and logic, possibility theory, traditional
probability theory, and (the main focus) imprecise probabilities and indefinite
probabilities.
Readings
Dr. Joel Pitt
The
OpenCog Software Framework (with Ben Goertzel) (2 sessions)
OpenCog
aims to provide research scientists and software developers with a common
platform to build and share artificial intelligence programs. The long-term
goal of OpenCog is acceleration of the development of beneficial AGI, a goal
which includes developing tools and protocols for AGI safety. This lecture will describe the OpenCog
software framework from a theoretical perspective, and also provide hands-on
guidance to working with the code.
Readings
Attractor
Neural Networks and Economic Attention Networks
Attractor
neural networks, such as the Hopfield net, are able to store memories as the
attractor of neuron activation patterns. The theory of the Hopfield net,
including the several variations that allow continuous learning and keyed
retrieval, will be briefly described
The
remainder of the lecture will introduce the idea of economic attention
networks, exploring how the focus of attention for an intelligent system can be
controlled using conserved quantities that are subjected to economic ideas such
as tax, rent, and wages. How ECAN is implemented within OpenCog as a number of
cooperating MindAgents will be explained. An example of ECAN within OpenCog,
that emulates the behaviour of Hopfield net, will also be introduced along
discussion about ÒglocalÓ memories.
Readings
Dr. Pei Wang
Readings:
Approaches
to Defining and Evaluating General Intelligence (1 session)
This
one-day talk will introduce the major approaches in building general-purpose AI
systems, compare them with human intelligence, analyze their theoretical
assumptions, and evaluate their potential and limitation.
Readings
A
Logical Model of Intelligence (3 sessions)
NARS
(Non-Axiomatic Reasoning System) is designed to serve as the core of
general-purpose intelligent systems. It is built according to the belief that
"intelligence" is the capability for a system to adapt to its
environment while working with insufficient knowledge and resources. This
four-day talk will introduce the major components of NARS, and discuss their
properties.
Readings