[Notes] CSCI 585 discussion 5 AL

20 Apr 2020 | CSCI585, English/英文, Note

Dangers of AI… We discussed the first five of these, last week:

bias deepfakes easy foolability lack of explainability unchecked power automation surveillance, loss of privacy … AGI - take a deep breath!!! Aaaaaaaaaaaaaaaaaaagh.

AGI: Artificial General Intelligence, ie. a human-like intelligent system.

AGI has been the ‘next gen AI’, since the birth of AI (in 1950) :) :(

So, where is it?

PS: a lot of what follows are my opinion… Too much HYPE, too early on! A series of early ‘wins’ made researchers get carried away. Here is a nice set of writeups on this. Tests for AI Turing Test: https://en.wikipedia.org/wiki/Turing_test

Loebner Prize: https://en.wikipedia.org/wiki/Loebner_Prize

https://developer.amazon.com/alexaprize

Aside: Searle’s Chinese Room Experiment: https://en.wikipedia.org/wiki/Chinese_room

Most of these tests/prizes are irrelevant when it comes to developing AGI! Language - is NOT about words! No amount of NLP [alone] is going to fix this.

Images - are NOT about pixels

So, what is it, then? The world has structure, phenomena, behaviors, cause-and-effect… WE LEARN THESE, THROUGH CONTINUING EXPERIENCE!!

We use a mix of induction (generalize -> likely conclusion), deduction (reason -> guaranteed conclusion), abduction (hypothesize -> best-guess conclusion) in our daily lives… People, Part 1 Misled/clueless/hype-y:

Stephen Hawking Bill Gates Elon Musk Ray Kurzweil … People, Part 2 Researchers and groups, exploring various aspects of AGI:

Paul Rosembloom (SOAR, Sigma) Arthur Toga Daniel Dennett Gary Marcus Alison Gopnik Demis Hassabis Judea Pearl Leslie Valiant Joscha Bach Yann LeCun NELL BabyX Allen Institute Josh Tenenbaum, also https://www.youtube.com/watch?v=7ROelYvo8f0&t=5085s Jeff Hawkins Kenneth Stanley (Neuroevolution) Karl Friston Giulio Tononi OpenAI DeepMind OpenCog … The above, spans the gamut - studying the brain, building brain-like hardware, brain-like software, musing philosophically, studying infants, pondering cause-and-effect, quantifying consciousness, proposing mind architectures, etc, etc. NO ONE HAS THE FULL PICTURE (a blueprint for creating AGIs) YET!

In addition, Marvin Minsky’s ‘Society of Mind’ is an intriguing collection of ideas related to intelligence.

My take on all this

brain structure MATTERS!

embodiment MATTERS! Embodied cognition is how nature does it

situatedness MATTERS! The brain is in a body that lives in the world (in an environment). VR provides a controllable environment for safe, modifiable, repeatable explorations…

memory STRUCTURE matters - how things are coded, retrieved…

continuous, 24x7x365 experiencing - record EVERYTHING

WE develop hunches/heuristics - from experience!

we can learn a LOT (on building AGI) - get clues from babies, children, animals! an inner ‘copy’ of the world, manipulatable a richly connected ‘graph’

emotions

attention

consciousness, thoughts, feelings, self, free will… can result in authentic creativity, incl art etc develop a comprehensive cognitive architecture testable items (on an AGI architecture) - conscience/morality/empathy, disorders, diseases, language, learning modes…

SDCs, robots… ALWAYS 2nd class - NO INTRINSIC MOTIVATION

summary: experience (v) -> ?? -> ?? -> ?? -> experience (n), and ‘4E’ (embodiment, embeddedness, extendedness, enactivism)