Meeting #7
Monday, 5/9/2022
Table of Contents
Lineup
Presenter | Paper/Topic |
---|---|
Andre | Measuring Intelligence |
Papers Discussed
- “Computing Machinery and Intelligence”, Turing. 1950.
- “Ascribing Mental Qualities to Machines”, McCarthy. 1979.
- “Using Deep CNNs to Prove That I Look Better Than Tom Cruise and Shah Rakh Khan Combined”, Bharadwaj. 2022.
- “On the Measure of Intelligence”, Chollet. 2017.
Presentation Slides
Open in a new tab here.
Recording
Forthcoming
Notes
Awesome notes courtesy of Janna!
Introduction
How do people measure intelligence?
- Framework to ascribe human qualities to machines and programs
What does it mean to say that a machine is “conscious”?
- Equation for intelligence
- ARC is a concrete dataset where we can optimize intelligence
Turing: Computing Machinery and Intelligence (1950)
Can machines think?
- Do not answer on the basis on democracy or intuition
- Imitation Game
Interrogator
How do we build machines to do imitation game?
What is the machine vs. human?
- Originally articulated in terms of gender
- Most people during this time thought that machines can’t think
- Identified learning as a key to AI
Practical:
- Only binary yes or no answer
Justifying the Imitation Game
- Seperates physical and intellectual capacities
- Truly measure the intelligence of the system
- I.e. what does it mean to have a language to think things?
- Does it know what an “apple” is if it hasn’t experienced yet?
- I.e. what does it mean to have a language to think things?
- Truly measure the intelligence of the system
- If you could artithmetic faster, you are a thinking machine
Only care about input → output
- How can we have a machine that demonstrate human thinking behavior but not be thinking?
- General philosophy has been favorable in scope of history
- Anti model agnostic
- “There’s no way that this race or animal can think like we do!”
- Anti model agnostic
Digital Computers
- Can do anything that humans can do — human computers are just higher than today’s calculators
- Digital computers vs. Nervous system
- Both electrical
- Thoughts from nervous system are just complex hierarchies of digital ops
Learning Machines
- Simulate a child‘s mind
- Simulate its education and evolution into an adult brain
McCarthy: Ascribing Mental Qualities to Machines
How can we ascribe mental qualities like beliefs, intentions, and wants to machines?
What is legitimate to ascribe?
- You can say that a machine is “conscious”
- Term for the “self”
→ Why ascribe things at all?
- Helps understand the structure of the machine and its temporal behavior
- Want to describe the machine and its state
- Need language of mental qualities to describe machines that represent higher level organization
- Computers can perform abstracted tasks
- Go one level further where its demonstrating complex qualities
Separate mental qualities from motivational structures
- When you think of “feeling”, you might associate that with motivational structures, however
- Use mental qualities to understand the internals of the system
Systems with mental qualities
- Thermostats, self-reproducing cellular automata, computer time-sharing systems, programs designed to reason
Chollet: On the Measure of Intelligence
We need explicit intelligence metrics
- Concrete metric → easier to measure → move towards AI
uring test and variants
- So much influence over intelligence of machines
- Intelligence is tied to how much the interrogator knows
- As we become more acquainted to non human intelligence
- Generator fooling discriminator
Two different understandings of intelligence:
1) Task specific skill
- Mind is an arrangement of ~static specialized mechanisms fine-tuned through evolution
- i.e. How do our eyes evolve to see well?
2) Generality and adaptation
- Mind is a general purpose algorithm
- Arbitrary experience → knowledge and skills
Chollet believes that both views are flawed.
Task specific skills of DL:
- Primary way is bench marking to standardized metrics
- Metrics are key to the modern deep learning
Metrics for task specific skill:
- “AI Effect”
- When AI does something new, it’s not really “thinking”
There is no single task X such that skill in X demonstrates intelligence
- When the machine beats world champion, we don’t think that machine is intelligent
- Machine built to do that one thing is not very “intelligent”
- Narrow skills are impressive in the context of generality
- Humans playing chess is impressive because it’s built upon cognitive skills
Generalization in AI:
System-centric generalization
- Interpolation
- Can the system generalize stuff based on what’s given?
Developer-aware generalization
- Extrapolation
- Can the machine see beyond the system?
- Give machine a new point (x=) which is beyond what the system has seen
- Can the machine see beyond the system?
Current Efforts for Broad AI Evaluation:
Generalization in Reinforcement Learning
- Exposed to environment via exploration
- Only for trivial modifications
- Multitask benchmarks
- How does the machine perform with other skills?
Where do we fall short?
Surpassing humans in skill
- “Moonshots”, AlphaGo & AlphaZero, DotA2, “AI beats human”
Developing broad abilities
- Learning to learn
- Acquiring new skills
- General, flexible
New Perspectives
How do we evaluate a skill?
- Chess can be abstract when humans play it, but it does not need abstraction
- Think of the difficulty of abstraction
- Learning hard-coded knowledge from data
- Rigorously control the priors, experiences, and g-word (generalization)
Universal intelligence is a scam; must be anthropocentric
- Obtain universal intelligence vs. Simulated human intelligence
- Progress should be benchmarked against human intelligence
- Recognize that this anthropocentrism is not greedy
- Property not restrictive to humans
- When comparing AI, compare it to human mode of thinking intelligence
- Recognize that this anthropocentrism is not greedy
- Progress should be benchmarked against human intelligence
Priors - where intelligence starts from
- Most intelligence is acquired; it is not innate
- Acquired via interaction through environment
- Learning algorithm
- Acquired via interaction through environment
- Human cognitive priors
- Low level — teeth chatter
- Meta-learning — causality (making sound when one hits the table)
- Knowledge — visual object-ness, Euclidean spaces, goals, social
Intellectual progress paved so far
- Quantify the strength of adaptability
- General AI must be benchmarked against human
- There is no meaningful concept of intelligence that humans can pursue
The intelligence of a system is a measure of its skill-acquisition efficiency over a scope of tasks, with respect to priors, experience, and generalization difficulty.
- Experience → Skill
- Great way to think about intelligence
Photos
Also courtesy of Janna