A record of the chat during the birds of a feather discussion of AI and Snap! at SnapCon

Computer Science AI in Snap!
1

There are some interesting discussions (including the suggestion to create this sub-forum) and links here.

19:01:26 From James Bologna : The virtual conference: https://www.ai4i.org/aik12 (If you are interested, please consider presenting a session!)
19:02:04 From James Bologna : Also love the topic of teaching AI topics without the use of computers!
19:02:29 From Brian Broll : @Tilman Michaeli has some resources on that :slight_smile:
19:06:25 From Brian Broll : I think that is a huge question. There seem to be tradeoffs on each side
19:08:01 From Mr. Stempel : Introducing the topic of neural networks by having students think about their own learning and how their brain works would be doubly interesting as a way to build meta-cognition skills and mindfulness about themselves as learners…
19:08:56 From Brian Broll : Good point, @Mr. Stempel!
19:09:15 From Brian Broll : A Blocks-Based Introduction to Text Analysis Snap!Con 2020
19:10:25 From James Bologna : Agreed @Mr. Stempel - starting with some of @Tilman's resources to abstract to the core concepts
19:10:53 From Brian Broll : GitHub - NetsBlox/Custom-Python-Services: Custom NetsBlox Services in Python! This enables you to create your own services for use with select groups of users.
19:12:27 From Brian Broll : A completely other approach to teaching ML coming from the math side: http://brollb.github.io/general/netsblox/2020/01/24/autodiff-in-netsblox/
19:12:32 From Rich Nguyen : Thanks @brian
19:13:02 From Brian Broll : Yeah, no problem!
19:14:07 From Rich Nguyen : TensorFlow.js is amazing!
19:14:26 From Rich Nguyen : https://www.tensorflow.org/js
19:15:01 From Brian Broll : applying style transfer to costumes is cool :slight_smile:
19:15:13 From Mr. Stempel : I've had this Google service flagged to figure out a way to have students experiment with machine learning. If anyone has any other issues, I would love to collect for a Community of Practice I'm a part of…https://teachablemachine.withgoogle.com/
19:17:11 From Mr. Stempel : Sorry, not issues, resourses...
19:17:29 From Ken Kahn : The work I was talking about is at https://ecraft2learn.github.io/ai/
19:17:55 From Mr. Stempel : @Thanks!!
19:18:00 From Mr. Stempel : @Ken
19:20:14 From Mr. Stempel : @Ken, I think the interdisciplinary approach is so important. As a general STEM teacher, I want students to find applications in whatever their passion or preferred disciplines/careers are…That's how I keep them engaged in non-chosen elective course. I honestly believe as Tilman is saying that these skills will be essential regardless of their career or discipline.
19:20:55 From Brian Broll : I think that is a good idea!
19:21:04 From Mr. Stempel : @Rich, I like that approach. With my STEM students we won't get far beyond block coding.
19:21:07 From Tilman Michaeli : Brians textmining for example is doing exactly that for a University Level :slight_smile:
19:21:14 From Brian Broll : The challenging thing is that I don't think there is a silver bullet
19:21:17 From Leonardo : Where can the community go to continue the conversation?
19:21:27 From Leonardo : Is there a piazza that can be created?
19:21:27 From Darek Dorozalski : thanks
19:21:34 From Mr. Stempel : @Ken, great idea in the SNAP forum
19:21:37 From Rich Nguyen : Great idea to continue this in the forum
19:21:38 From James Bologna : Thanks Ken!
19:21:43 From Rich Nguyen : Thanks Ken!!!
19:21:44 From Sara Sonia : Thank you!
2

I don't think I believe this. When I attempt introspection, I don't think about neural nets, and a fortiori I doubt that kids who didn't grow up at the AI Lab think about neural nets. Maybe they think about something on the lines of Society of Mind, but even that isn't really available to introspection. You can of course pretend, to please your teacher, kind of like pretending you made the flowchart before you wrote the program.

3

This discussion thread began with a comment I made that Aaron Sloman wrote (in the late 70s) that as wonderful as it is to support children in building virtual worlds and games that helping them try to build simulated minds would be even more so. The connection with meta-cognition and introspection skills is plausible here. But today they are simulating brains not minds. So the question of whether there might be a meta-cognition boost seems much less likely as Brian suggests here.

While I mostly agree with Brian there are times when I see some support for Mr Stempel's claims. I've come to appreciate the idea of a distributed semantics for words (You shall know a word by the company it keeps (Firth, J. R. 1957:11).) When I think about how a neural network does vision I recall a live drawing class where it was so hard to draw based upon what is in front of my eyes when it is so much easier to draw the high-level interpretation of a scene that my vision system made available to my consciousness. The way neural networks work provide some insight into why drawing instructors suggest exercises such as looking at a scene through a narrow tube or upside-down.

But I haven't begun to look at millions of images to learn to tell apart cats and dogs. :smiley:

4

Aaron Sloman, that's a name I haven't heard in a while.

They told me to draw with my eyes closed, kind of the opposite of your exercise. But more to the point, your introspection doesn't tell us anything about kids' introspection; you know too much. I'm pretty sure that kids in your drawing class are not thinking about neural nets; if they're introspecting at all they're thinking "he wants me to see the trees, not the forest." High level of abstraction, way above the hardware level!

5

Yeah, I guess there's something similar in extrospection, such as in physics (or all the natural sciences as matter of fact), where experimental physicist's data is always theory-ladden, in other words, prior-knowledge--ladden.

A kid could do (i.e. reproduce, step by step; following a given plan/'recipe') exactly the same experiment as an (academically knowledgeable, grown-up) scientist would, and even look at the same experimental data, but the kid wouldn't be able to say what she/he/they were looking at.

Because without the (prior, domain-specific) knowledge both the extrospective and introspective "data" aren't (real) data at all (such a "data" is instead just a mess of some, low-level, sensory input, I guess).

I still remember when I was a kid and I envied adults their (seemingly) superhuman, almost mysterious, understanding...

EDIT
In many ways I still don't understand a lot of what I observe in the world and in myself. I wish my understanding would be greater.

6

Ha ha. I agree. Good one.