Aaron Hertzmann
@aaronhertzmann
www.dgp.toronto.edu/~hertzman
This suggests a fun exercise: watch fictional sympathetic robots, but imagine them as mere chatbots: Data providing simulated relationships to humans; C-3PO + R2-D2’s bickering as generative performances of personality; WALL-E mindlessly following its programmed behaviors long after humans left. 5/
Very interesting! It seems like there's a common trend here; a similar thing happens in visual art when medium-quality models make the most unintentionally-interesting pictures: dl.acm.org/doi/abs/10.1... osf.io/preprints/ps... (artbreeder images are deliberately manipulated for effect)
Chaotic systems like double pendulums are especially hard to get right and make accurate predictions. So simulation of even moderate size systems can be nearly insurmountable. We can predict accurate behavior of the weather with some probability, but it's far from a complete simluation. 12/
Maybe the brain is some kind of non-digital computer? This might be true, but not useful. We don't have much theory for analog computers beyond outdated technology like slide rules. "It's an analog computer" sounds like "it's a computer." It's really misleading. A brain is just a brain. 8/
I agree, but I think biology/mechanisms play a role in how we value others. For example, modern values about animal welfare often relate to beliefs about their natural behaviors. My dog doesn't have the right to vote. Anyways, I think it's worth arguing both fronts (social and substrate).
It comes down to the point from your paper about different usages of "computer". When people say that LLMs can be conscious because the brain is computer, that isn't about analog computers.
FYI I am a computer scientist and I taught undergrad and grad courses in computer science for many years. I have worked through the steps of Turing machines by hand. I understand the difference. That's not what the debate here is about.
By "computer," we normally mean the real, digital computers that convert everything to 0s and 1s and process them with logic circuits, all synced to a fixed clock cycle. You can separate what a computer computes from how: the same software can run on many kinds of computers. 3/
On the last day of SIGGRAPH this year, we will hold a new technical workshop called "From Human Vision to Machine Vision", with talks ranging from low-level human visual perception, to applied perception, art & photography, and to machine vision. #SIGGRAPH For more info, see: fromhv2mv.github.io
And here's a progression of Piet Mondrian's work over a decade before he got to his most famous abstract work. 8/
Some scientists have claimed that the aesthetics of abstract art relate to real-world natural image statistics, which I agree with. And I think this is visible in the way that some abstract artists began realistic and got more and more abstract; here's a progression of Richard Diebenkorn's work. 7/
Here are strokes optimized to be classified as faces, in 2020, by @mmariansky.bsky.social 6/
As a model for this, consider the work of @drib.net , in which he optimized strokes and blobs to maximize state-of-the-art image classifiers (in 2018) to produce desired categories. This is his "ceiling fan." medium.com/artists-and-... 5/
Even when there are recognizable outlines, a lot of the detail seem like textural scribbles, not recognizable individual details. 3/
Sometimes I sketch really fast, in ways that produce drawings that seem recognizable, but often with few recognizable low-level features. Often, objects are recognizable but the individaul details don't make much sense. 2/
After all, what is the alternative? Maybe you alone are conscious. Or maybe there are a few, undetected philosophical zombies amongst us. Either way, it would seem that consciousness plays no meaningful role in human behavior or abilities, despite shared biology. 4/
Now that I've looked up Thorndike's law, I want to be clear that I was not claiming to have discovered it. It is the main axiom of every animal trainer and book I've encountered (all positive-reinfocement based). It probably comes from Karen Pryor, who wrote these relevant words in 1999:
I’m not sure but it would be some generalization that requires long-term planning in unfamiliar situations. See also the post:
The best book on behaviorism for animal training (and also humans) that I’ve read is: Don’t Shoot The Dog by Karen Pryor, the trainer that led the way in positive reinforcement training. I didn't learn how to train dogs, but I learned to better understand the training skills I had learned. 10/
Banishing "knowledge" from your dog vocabulary lets you describe dog behavior better. A dog does not know "he gets a treat if he sits", he has learned the habit of sitting in response to sitting, and to expect a treat. A dog doesn't "know" their name, they have learned to respond to their name. 3/
Dogs can be described well behaviorally: they learn from conditioning (including training); they learn associations (this thing good, that place safe, etc.); they have internal states, and so on. Dogs do not know about the future or the past and they cannot plan or solve multistage problems. 2/
@rodneyabrooks.bsky.social called them "suitcase words." www.technologyreview.com/2017/10/06/2...
I got 42.64, but I have a lot of practice using HSV color sliders to match colors (from practice with digital drawing), and I can think of a lot of words to describe color that help me remember a color
Artists also use conceptual knowledge. For example, a person's eyes are halfway between the top & bottom of their head. But it's so easy to get this wrong, like in this sketch I made! Here's are illustrations of one technique for avoiding the problem, including sketches by Veronese, 1568. 13/
Learning to draw also involves learning eye-hand coordination skills, like the "Target Locking" behavior shown in this clip from John Tchalenko's film "Capturing Life." This is a common drawing behavior, but people don't know about it because people don't know what their eyes are doing. 12/
What happens if you restrict eye movements? Here’s a 30-second "blind drawing" I made, without looking once at the drawing. I took the photo a moment later. The “errors” are typical of blind drawing: individual shapes are accurate, but their proportions and placements are not. 10/
These visual limitations explain why artists typically move their eyes so frequently. Here’s a person copying a picture, recorded with an eye tracking device. The black rectangle show where the his eyes are looking at any moment. 8/ (Clip from "Capturing Life" by John Tchalenko)
One clue is that human vision is severely limited, both in terms of what we see at any moment, and how much of it we remember. And, we are *totally unaware* of how much we do not see, until it is pointed out. Suppose you wanted to draw a picture from this photo of a houseplant, 5/
Here's Henri Matisse drawing a portrait of his son. Watch his head and eyes, and how often they look back and forth between his drawing and his son. Why does he look back and forth so frequently—why can't he just draw his son from memory? And where exactly is he looking? 4/
