Summary Session

Acknowledgments

James Hendler

This meeting was supported in a large part by a grant from NSF IIS

• Fellows Forum aka Festschrift for AI
• Lunches at AAAI/IAAI
• Panel at AAAI/IAAI reporting back from here
• Web site and archiving of materials from this meeting
• Grant report (sooner than the one for the '56 meeting)

• Festschrift would be a "book" - now it's a web presence (a snapshot of AI in 2006)
• Fellows are still invited to submit position papers (send to hendler [at] cs [dot] umd [dot] edu)
• Site will be opened to public near future
• Non-fellows will be invited to submit (mechanism TBD)

Material for site

• Send slides or text or whatever of presentations
• Send photos (or links to photo sites)
• Send pointers to any other materials
• (wiki)

Can We Design an Architecture for Human-Level Intelligence?

Stuart Russell

* Purposes

• Architectures as templates for agents f: percept -> action
• Every 5 years consider where a current architecture with state-of-the-art components would break
  -> research for the next 5 years
• Understand the profusion of architecture proposals, develop the next one
• It won't be trivial, but complex...
• Can't build cathedrals without it

* Profusion

• RoboSOAR, ACT-R, Theo, Prodigy, ICARUS
• Subsumption, Society Of Mind
• NIST Reference Architecture, model-predictive control
• Modern AI
• Multilayer robotic architectures
• Machine Learning

* Progress

• My boxes-and-arrows are better than your boxes-and-arrows
• Mine explain my brain better than yours do
• Mine do better on this testbed than yours
• Some people thought that no realization can be of this box-and-arrow diagram

* Infrastructure

• Tools and components -- some are becoming available, much more needed
• Representation substrates
• Need an agent-building language providing composition and decomposition of architectural components (might enable self-extending architectures)
• Architectures don't need to be the same, but have common components, which can then be built

* Theory

• Well-designed/evolved architectures solves what optimization problems? What forces drive design choices?
• Each by itself leads to less interesting solutions
• Bounded-optimal solutions have interesting architectures and can be found for some nontrivial cases

How Has AI Computational Modeling Contributed to the Study of Other Domains?

Kevin Ashley

(Attendees: David Smith, Edwina Rissland, Peter Friedland, Bruce Buchanan, Margaret Boden, Kevin Ashley)

* Charge to breakout session

• AI computation modeling has long been urged as a tool for empirical investigating issues of interest to non AI domain experts in various fields
• What contributions have been achieved and how did AI contribute?
• How can we foster?

* How did AI contribute AI model as:

• Explanation of phenomena, suggesting things scientists should look for and helping them interpret their experiments
• Enabling systematic search (eg. pathways to synthesize organic chemicals)
• Enabling systematic manipulation of plausible models for investigation
• Encouraging reflection in knowledge acquisition
• Teaching tools for demystifying domain skills
• Providing ideas/concepts for conceptualizing questions and experiments in other domain
• Tools enabling professionals to delegate tasks to computers
• Focusing on knowledge-sharing, organizational and corporate memory
• Facilitating constraint-based design and use of genetic algorithms
• Making film animation...

* How to Facilitate Contributions to other fields

• Making allies in other disciplines
• Making computer science respect collaborative alliances
• Publishing ideas:
  • Helping authors write domain text books using AI models and ideas, vocabulary
  • New edition of AI handbook with AI tools others can use
  • Showcase of 18 years worth of IAAI papers
  • AI texts with examples embedded in disciplines

Can We Design a Never-Ending Learner to Solve The Natural Language Understanding Problem?

Tom Mitchell

Motivated personally by a lobster dinner bit
Having computers reading the factual content on the web
It was an AI complete problem

Challenge discussed: pulling together good ideas
Language learning systems, everyday better than yesterday
Maybe understanding was the wrong word
-> look at any webpage -> extracted fact should be the same

Why become an inflection point, why is it a good time for this challenge now:

1. we have the web; we have a lot of text data, labeled data, semantic web data
2. progress in NL, resources shared by the community (parsers, wordnet, speech taggers, etc.)
3. self-supervision natural language learning algorithms -> a list of cities, and a list of the text around where the city occurs
4. massive redundancy on the web

Key research questions/hypotheses that have to be address scientifically:

• KnowItAll extracts millions of facts from the web without initial ontology and then learns ontology as it goes
• whether one can learn to interpret natural language without separate linguistic understanding of the world -- no unanimous agreement
• extract things from text (still in the level of tokens)

How do you know we are successful? no consensus

What Can We Learn from Linguistic Semantics about KR & Reasoning?

Lenhart Schubert

* Semantic devices in all human languages, beyond {predication, and/or/not, for all}

• We can use logic constructs from semantics to knowledge
• Generalized quantifiers and generic sentences
• Modification
• Modality/intensionality
• Reification
  ...
  => may broaden our ability to handle real lang represent and reason with commonsense knowledge, mind knowledge...

* Some reactions and ideas from the group

• expressiveness of language in knowledge base...
• provide insight into how human conceptualize the world
• working on better approaches
• tractability
• motivation
• gluing sentences together
  ...

FOL is too weak to be interesting. Some of us are working with richer systems, and tractability can be damned; exceptions used to be frowned upon...

* Suggestions

Danny suggested that let's formulate examples if inference seems to depend on richer representation.

* Examples

• Inference
  • Most dogs are friendly. Snoopy is a dog.
  • What does this imply, friendly? all the time?
• Coherence

Broadened language understanding, acquiring "deeper" knowledge from text, and commonsense inference call for "language-like" representational capabilities

Do We Need a Common Framework for Investigating Architectures?

Aaron Sloman

There's some common terminology but generally words and phrases are used with different meanings by different people
There are no agreed formalisms, diagramming conventions, theorems

Many people investigate the best, or the right, or the most useful architecture, but not clear what the search space is...

* Architectures and Niches
An architecture is a spec at some level of abstratctions... components may be virtual machines or physcial components.
A niche is a specification of the requirements againt which the system can be evaluated. Like architectures niches can have different levels of abstraction and be more or less detailed

* I am guity too!
Ideas about the H-Cogaff architecture (for adult human-like systems) have been developing for 15 years: complicated diagram with many boxes and arrows
Closely related to Minsky's emotion machine, but driven by different interests

* Specifying an architecture
A specification of an architecture can be recursive and will typically include:

• Information about the main types of components
• The kinds of ontologies used
• The forms of representation used
• The mechanisms that are used to perform various functions
• The boundary between a system and the environment can be viewed differently for different purposes

* Design space and niche space
a design can be related to many possible niches and virse versa

* Varieties of trajectories

• trajectory in design space
• trajectory in niche space
  (for instance, architecture for kids to adults)
• caveat: biological evolution is discontinuous

How Can a Robot Learn the Foundations of Commonsense Knowledge from Its Own Experience with "Blooming, Buzzing Confusion"?

Benjamin Kuipers

* Human-level AI requires vast amount of knowledge and a lot of ontologies
Where did the ontologies come from?

* What is the mind of the robot getting before requiring ontologies?

• getting both sensory and motory pixels
• low-level perceptual information

* How do we learn these high-level ontologies?
They don't get there by being programmed, they must come from some learning process, or evolution
Evolution is a learning process
Individual gets the benefit from learning
Learning as species and learning as an individual
Surprisingly, individuals learn things superficially

Genetic code isn't big enough to specify everything the visual cortex needs to know
The machinery gets learned pre-natal by experience...
We know its important.

* Which categories/ontologies that we must learn?

• high-level ontologies is a good idea... particular ontologies?
• space, time, actions, objects, predicates that are tied to actions and objects
• closing the sensory loop
• low-level continuous reactions
  symbols of ontologies are built on top of those
• non-linguistic animals, pre-linguistic children
• part of he value is that it provides convenient access to incomplete knowledge (look at the intermediate state)

Interesting problem, lots more to be done.

New Challenge Problems for Research in Heuristic Search

Richard Korf

The term "search" becomes ambiguous:

problem-solving search, web IR search, etc.
same? different? do they help each other?

* Problem-solving search -- 2-step algorithm:

1. find a problem that has not been solved yet
2. write a computer program to solve it

* What are some of the new problems to replace the old problems?

• Games:
  • chess -> bridge, poker, real-time strategy game, ambiguity in games
  • puzzles (towers of hanoi) -> 4 or more pegs, optimal solution length -> becomes a search problem
  • rubik's cube solved in 20 moves
  • satisfaction problems (sudoku puzzles)

• Math:
  • ramsey numbers
  • party problems -- at least 3 people know or don't know each other
  • operation research problem

• Computer science:
  • query timing
  • relation database

• Planning:
  • over subscription planning -- cost function itself is difficult to compute

• Computational biology:
  • sequence alignment

How Must Logic Be Modified for Representing Common Sense?

John McCarthy

(Attendees: Don Loveland, Vladimir Lifschitz, John McCarthy, Leslie Valiant)

(What extensions to mathematics logic are need to represent human-level AI)

The proposals included nonmonotonic reasoning, approximate concepts that can't have if and only if definitions and contexts as objects...
The need for nonmonotonic extensions was agreed. Because of Godel's 1929 completeness theorem, extensions can be genuine only if they produce some sentences not true in some models of the premises.
The need for extensions to deal with approximate concepts was questioned. Maybe only weak axioms are needed.

Afterthought: this seems true for the syntax of sentences involving approximate concepts. The semantics seems to me more mysterious...
Needing contexts as objects was also questioned

Research on Integrated Systems for Human-Level Intelligence

Pat Langley

... if we agree this is desirable
(referred to Nils' Dartmouth talk)

Testbeds: community testbeds to encourage research in the area
Robotics: if no infrastructure, use synthetic environments

Should come with sets of challenge tasks

Tools: open source, to build and rebuild stuff
with good communication protocols to make integrated agents

Do research on the components
Evaluate on the whole system

Reward structure: grand challenge or smaller version of that

Publications: AAAI integrated systems track
Different content than the traditional AAAI conference

Education: fragmentation of AI has led to narrowness...
big picture of AI is not there for PhDs
Testbeds will make this easier (students get excited about it)
Foundation AI class to give the overview, long term goals

Funding

Promoting AI

Eugene Freuder

(practical issues)
Are we doing as good as other fields?

* Why promote?

• need 20 PhD students to generate 1 good idea
• list of questions
• how do we capture public attention
• what is role of grand challenges
• generate better support for international collaborations
• interaction among sub-communities
  
• successes and failures of first 50 yrs
• where we want AI to be in the future
  
• additional activities the fellows would like to organize
• more events?
• is there a more active role for the fellows?

* Individual suggestions

• motivating problems, competitions, grand challenge
• brochures with 10 open questions
• new edition of AI handbook
• popular science books
• documentary
• facilitate PR and lobbyists
• better educate members for lobbying
• provide NSF successful stories
• put people forward for prizes in other communities
• virtual laboratories
• infrastructure

Is it Time to Resurrect the Original Shakey Robot Project Using Current Technology?

Marty Tenenbaum

* Why?

• a forcing function for the science needed to do human-level AI
• logic and language recently added
• need to understand foundations rooted in perceptual motor loop

* Focus on real vs. artificial problems

• dealing with the complexity and messiness of the real world
  (vision and manipulation in unstructured environment)
• addresses timely themes (cognitive architectures, continuous learning, social interaction)
• it's exciting to kids, the public, funders
  150K attended robocup competition in japan

* Why not?

• language, social interaction seem more central to human-level AI than perceptual motor loop
• engineering quagmire
• blind people, children without arms and legs have a rich theory of mind, but not children deprived of social interaction

* What?

• Not Shakey - basic mobility is done
• Not Stanley - too special purpose
• Scientific: want a general purpose robot that can be trained, basic capabilities of a child, teach a variety of things
  robots interact each other and humans
• Application: service robots interacting with humans in the home

* How

• Core cognitive Substrates
• visual object recognition of a 2 year old child
• manual dexterity of a 6 year old
• language capability of a 4 year old
• social capability of a 10 year old

* Needs

• Robotics is hard because it needs infrastructure
• Funding

James Hendler: There's still a lot to be done; we don't have to rewrite our textbooks...

Transcribers: Chih-yu Chao & Varun Aggarwal