Transcribers: Chih-yu Chao & Varun Aggarwal
At that time, language was approximated using the following model:
controlled lang. -> model + model-focused interpretation
The open questions in relation were habitability, learnability, generalizability.
There is a shift to Broad language capability now, the model being:
broad coverage & mappings -> layered mapping -> language oriented logic
* Layered mapping
C-structure -> F-structure -> linguistic semantics -> conceptual semantics -> contextual semantics
Open Questions: what layers, handling ambiguity, language use KR interaction
Examples of ambiguity: I like "Jan", untiable knot, the duck is ready to eat, every proposer wants an award
Ambiguity at all hierarchical level of the model leads to combinatorial explosion.
Probable solution: Use learning to kill choices asap. It gives fast computation, but not necessarily the right result.
* Some open challenges
Larger structures need to be incorporated - linguistic, ontological, historical
Learning by reading as humans do: leverage the scale, use teaching texts and tutoring interactions, solve problems at the end of chapter, analogy to previous knowledge
Address the issue of scale - one million facts fallacy
* Some predictions
What happens after 5 years
*Logic is still important in AI
Influence: Minsky, Newell & Simon
Earlier view of AI: ultimate intelligence will be society of programs and different researchers would work on these programs.
Later view: Approach of learning coglomeration of NNs, Bayesian, etc. Logic analytic procedures will replace some of the learning methods.
Logic shall be faster than human counterpart. Evolutionarily brain is a pattern recognizer and learning is built on top of it. Using logic analytical abilities, one can find regularities and codify. Not possible in general by common sense.
Prominent stuff:
General approaches: problem solving, expert systems in the olden times. Operating systems have roots in AI.
machines were viewed "smart" (allowing multiple users simultaneously)
Expert system in the 60's: Resusable and generalizable, more programmable
In 70s: Language (Lisp), Giant lisp machines, general purpose machines
60's - 70's: Budget cut, downturn in speech understanding, Approach: take what's out there and make something new
80's: Take the great ideas to the military (industry wasn't a big player)
where do we go from here/what darpa cares about?
* Pioneers Turing, McCarthy, Minsky, Rochester, Shannon
* How one thought about AI?
Any aspect of learning/feature of intelligence can be precisely described for a machine to simulate it
* Second Generation McHack VI, Shakey the Robot, Shrdlu, Freddy, Hearsay systems, expert systems (Dendral, Mycin, ...)
* Current State of AI: Deeper AI, more specialized and technically and mathematically rigorous
* If I were younger what would I pursue (Current interesting things):
* "HLAI-Friendly" projects
Speech Recognition
Situation today: far away from human-level recognition
Goal: recognize, interpret, and execute spoken language input to the computer. Make speech the preferred mode of communication.
Challenges: sources of variability, too many sources of knowledge
How research progressed:
Early on: Use of syntax
Then use semantic
1976: organize and use knowledge: BBN, use of graphs, knowledge represented as graphs, find path through network to recognize speech. (task level knowledge: hearsay system, chess commands)
Mid 80's: FSG and HMMs based systems; Speaker Independent Speech recognition systems (SPHINX); how to organize and represent speech,Dragon and Harpy systems; integrated representation provided a single abstract model, leading to a great conceptual simplicity; Beam Search: optimal search requires consideration of every path at huge cost
Solved the problem of unlimited Vocabulary dictation: statistical language modeling
Challenge: non grammaticality in spoken language
* Land Marks and commercialization: dragon dictate and naturally speaking, IBM via voice, nuance-based tellme 800 services, microsoft speech server (voice dialing)
Speech recognition requires interdisciplinary teams
Future challenges
Raj Reddy: where are we & where are we going
a1: necessary to put deadlines for to achieving certain AI goals?
Raj Reddy: history of having roadmaps to get funding
Robert Kahn: internal deadlines, fine; respect to the funding agency, can't have the general motion...
Raj Reddy: lessons learned: its not that the agencies don't appreciate it; it has to go to the congress, and it's understandable in areas like astrophysics, semiconductor industry
a2: (respond to Raj) in astrophysics, it's the purpose of needing some equipment, not deadline
Raj Reddy: how one can specify deadlines?
Don Loveland: roadmaps, not deadlines, except for the very near terms
Raj Reddy: set the necessary conditions to progress for AI
Robert Kahn: if you were dealing with the search for intelligence, what kind of deadline would you put on that?
a3: (why interact with computers) elderly, companions
can't do speech - too slow
planes are not better than birds
in the early days, people did things in small groups; will it become that we can only do things in large groups?
Don Loveland: set the state of parameters to exceed the capabilities of human
Nils Nilsson: large groups for big things like telescopes, small wild groups for ideas, there should be a mechanism for funding; to achieve human level AI, we shouldn't neglect that what we want is more investigator-initiated funding
Daniel Bobrow: another problem -- AI is not science, but engineering discipline; you have tasks to achieve; can't say whether ideas will scale until you build them; can be sure with just an idea or a small implementation, need to integrate new ideas in the larger engineering context, what individual behaviors to transit into large engineering context
a4: (to Nils and Raj) hard to cumulate great ideas; Hawkins reinventing the blackboard model, which resurrected in other forms
Nils Nilsson: angry at Hawkins also; however there are some interesting ideas in the book, hopefully be referred by future researchers
a5: any single thing that surprise you that we did/didn't accomplish?
Nils Nilsson: tremendous success on propositional satisfiability models for calculation, solving P-SAT problems very fast
Don Loveland: resolution, grounded system, move the motion of theorem proving back to grounded system
a6: large funding goes to instrumentation, why not get more people setting integrated AI? is there anything that will be helpful for integrated AI?
Raj Reddy: all kinds of reasons.. shallow instrumentation
Robert Kahn: infrastructure in the discipline; sparks that never get followed upon, big aid to community in making progress
a7: (blackboard issue) how people come to know about earlier work (big gap, copyright), any plan to make it more accessible?
Raj Reddy: it's a social system we never worked on
a8: (Nils' system) do we also need to think about expanding the range we're trying to model? emotional modeling hasn't received much attention; is it worth exploring?
Raj Reddy: agree!
a9: different ways of communication; speech recognition: it depends on the context, goal, etc.
a10: can't have a unified infrastructure, problem is that framework is not compatible to different ideas, ontological commitment, each idea is incompatible with someone else's experiment; should interact in an appropriate way
a11: (follow up) Pyro: python architecture for plug-and-play AI robotics work; good for integrating isolated small groups
a12: lots of data, change in AI publication -> machine learning & induction, to what extend statistics will continue to be the strong driving force? methods to learn the data?
Daniel Bobrow: important social issue in the AI community; true understanding -- not the matter of being right on average, but right on particular...; use statistics when you have enough information; how to get enough backbone
Don Loveland: we had logic and structure, causal graphs; still using them but slowed down; things will balance out
a13: (follow up) agree it's a social issue; community is put ahead of communicating within itself; modern model (graphic search) has the essence of Dendral; needs better inspection and realize that it's similar to those dreams
Robert Kahn: need to understand the procedure of how the declarative world works; need to understand what exactly needs to be spelled in order for integration; need to know what it is and what to rely on --> the success of the internet
balance between procedural and declarative
a14: (comments) emotions, architectural, not shallow; integration, don't have time to find out other pieces of knowledge; "Learning how to Read", also require being wanting to communicate; language learning in human grows in so many aspects, can't succeed unless we include all these
a15: watch a 2-year-old -- the actual phenomena -- to get the inspiration; children learn how to close the sensory motor loop, how to do research on that? affordable robot with manipulative capabilities... things that will enable small researchers to conduct research
a16: need lots of data, not only for language problems, but for vision problems as well; how to undertake data from multiple AI areas?
Daniel Bobrow: what constitutes data vs. how long it takes to get good data; where to use the data & what the data based on/the characteristics of the data
Transcribers: Chih-yu Chao & Varun Aggarwal