Meeting #3

Friday, 2/25/2022

Table of Contents

1. Agenda
2. Deliverables
   1. Presentation Guidelines & Objectives
3. Notes
   1. Discussion
   2. Groups
   3. Journal Club Possibilities

Meeting recording available here.

Agenda

1. Housekeeping 5:10-5:20
2. Agenda
3. Journal Club Announcement
4. Social Media
5. Discord Tutorial - Joining as a Member
6. Discussion Questions 5:20-5:50
   • Key components of thought
   • How much do you trust your brain?
   • Perception and truth
   • What do trees feel? - Is Siri intelligent?
7. Card Games - resolving the spoons tournament 5:50-6:00
8. Presentation Groups and Work Time 6:00-6:30
9. Ideas for Journal Club 6:30-6:50

Deliverables

Have a presentation ready with your group for next week (Meeting 4)!

Based on the form interest, the following research groups have been created:

Next meeting, each group will give an informal presentation sharing their research. If you want to be involved in another topic, let us know and we’ll add you!

Presentation Guidelines & Objectives

You will research a given topic in small interdisciplinary groups, then give a small presentation on what you learned. The idea is to discover what experts are doing in your topic/field, and where our team may be able to add value through our own project.

Touch on the following questions, not necessarily in this order:

• How does this topic relate to the idea of intelligence
  • Does it have neuroscience ties? Computer science ties? Other related fields in which it is relevant? (These ties are not requirements for the topic, just an open question.)
• What are some interesting recent papers, projects, or advances in this topic?
• What are at least 2 project ideas or directions that would build upon the progress of this field?
• Any libraries/technologies required/could help with implementation?
• What do we need to know more about to better understand this topic?

Notes

• Discord tutorial - make sure to go to #role-react and click the brain emoji to be given the Member role.
  • Clicking the Member role button unlocks presentation group channels where you can collaborate with your group members.
  • Generally, make sure to stay on top of #announcements channel.
  • All meeting notes, zoom recordings, slides, and more are on the website: interactive-intelligence.github.io
• Journal Club - scheduled for Monday, 5-6 PM, Sieg Hall 332 (see schedule page for most up-to-date information).
• Interested in helping us out with social media? Message us!

Discussion

Should we trust our brain?

• Shepard’s Tone - the perception of continuously increasing/decreasing pitch.
  • Optical (and audio) illusions as examples that we do not perceive reality or objective truth
• Do you trust your brain?
• Success rate of information being given.
• Whenever possible - try to offload everything from the brain - memory, etc.
• Complementary relationship between human and machine properties
• Rely on data and things other than our brain - Hume and induction
  • Is maximum likelihood estimation the best we can do? If we have only ever seen white swans then must we logically assume all swans are white?
• Preciseness of measurement.
• Abstract and precise truths. Humans are generally good at abstract truths (e.g. “the United States is a country”); machines are generally good at precise truths (e.g. “it is 23.4837… degrees Celsius”, “12345 + 98374 = 110719”).
• Metaphorical truth - something that is not necessarily true but is useful.
• Attaching emotions to events - instinctive judgements.
• Outliers
  • Is our idea of reality constructed by majority? What if the outlier is ‘correct’? How do we choose whether to include outliers or not? What about the gradient of outlier-ness - i.e. which data points are ‘more typical’ than others?
    • Out of Distribution Detection - what is the distribution of very high-dimensional datasets? See Balestriero, Pesenti, LeCun: “Learning in High Dimension Always Amounts to Extrapolation”
  • Idea that the more independent observations of the same phenomena occur the more likely to be true?
    • Example of DNA and transcribing proteins being analogous to computing and executing programs, these phenomena independently emerged, so does that evidence an objective reality
    • What about all the people that report seeing elves when they take DMT (independently reporting same hallucinations)? What is their ‘truth’/’experience’?
    • Standpoint epistemology - see Harding, “Rethinking Standpoint Epistemology: What is ‘Strong Objectivity’?”
• Should we trust AI? Can we trust AI?
• Court cases - how emotions can bias the truth, skewing events, etc.
• Voices in your head - who is write? Who knows about what we think?
• Both cells and CPUs share similar features that have been independently developed.
• Godel’s incompleteness theorem
  • If math - perhaps the most ‘objective’ system known to humanity - is incomplete, what is objective? Can anything beyond the complete calculation of the universe be objective (and no model capable of ‘pure’ objectivity)?
  • Objectivity is still based in general consensus, math founded on the idea that we can all come to a common representation of what ‘one’ object is / we can all count the same
• Aristotelian models - we like discrete representations and dislike fuzzy understandings.
• Legal and symbolic systems - Not a mistake that legal systems have been mentioned as examples for truths (e.g. abstract truths), they are designed to classify the world into good / bad
• Can we measure brain waves and neural activity during computation of something “objective” like “1+1” - how similar are the computations as a metric for ground truth?
• Hierarchical model of perception in both neuroscience and neural networks: as information is processed by different modules of the brain, higher order abstraction and meaning emerges.
• If we can’t perceive something, that doesn’t make it wrong/nonexistent/incorrect.
  • Cats can see wavelengths of light we don’t have perception for
  • How can you compare intelligence when the sensory inputs aren’t even the same
  • Survival as a rough metric for how accurately our perception matches some underlying reality?

Groups

• Everyone has been assigned a group. The group will conduct independent research and present on their research.
  • Goal: understand where the field is, how it relates to AGI, and how we can contribute with possible projects
• Initial documents and communications are happening in the drive and in the discord.
  • See #announcements for further details

Journal Club Possibilities

• Again, see you all Monday Feb. 28th, Sieg Hall 332 5-6pm for a lowkey first journal club!
• Chaytan’s ideas for possible journal club topics, non-exhaustive list, share any ideas wherever – #papers-and-ideas or #general or any other channel works
  • Information theory textbook, Jason recommends certain chapters
  • Interview w Ben Goertzel on AGI + accompanying podcast
  • Classic neural network papers or neuroscience papers, RL theory papers?
  • Neuromorphic hardware Shih-Chii Liu: Neuromorphic electronics, A historical perspective (Telluride Neuromorphic 2020)
  • Intel Loihi chip design Asynchronous Circuits
  • Recorded lecture: Towards Conscious AI
  • Less technical video about AGI: PBS, Can We Build a Brain?
  • Awesome merging of neuroscience and ML: Predictive Coding Models of Perception
  • Nature paper on Predictive Coding: Neurons learn by predicting future activity
  • Cell paper on DL from neuroscience: Informing deep neural networks by multiscale principles of neuromodulatory systems
  • Taking suggestions for any subjects you’re interested in or passionate about!
• Have one person present or take charge of a paper that is shared beforehand and present during the Journal Club.
• Will incorporate a mix of Youtube videos, publicly available recorded lectures, etc (tend to be broader in scope) with deeper dives into papers.
• People that want to present what they have learned or know about given subjects can do that as well on any given week.