They named journals differently in the old days! I just was referred to a 1989 article published in "Intellectica."
(Reminds me of the Roz Chast cartoon of "The man who was loved for his lack of lack-of-pretense")
| Web | https://bewitched.com |
Martin Wattenberg
@wattenberg
- 868 Followers
- 226 Following
- 50 Posts
Human/AI interaction. Visualization as design, science, art. Machine learning interpretability. Professor at Harvard, and part-time at Google's People+AI Research initiative.
@rahulbot How much background did people have going in? Had they already worked with HTML/JS before? (If some had React, that sounds pretty advanced... were the other students able to get up to speed quickly?)
@rahulbot What technology / framework did the students use? I taught a code/art class with P5.js last semester, and I saw huge interest in the "interactive essay" form. I am wondering about doing more in a future version of the course. But it was pretty painful for students in p5.js!
Q. How many minutes in February?
A. 8!
A. 8!
@laurenfklein @intransitive Looks like a fantastic course, wish I could sit in!
I really like that you're assigning some primary ML sources, especially the essential text of Brown et al. Reading about LLMs from critiques alone is very much a Plato's Cave experience!
One sign the article is technically misleading is that it completely misstates what the figure with the spiral is showing. It's about approximating a distribution, not a single image. Leading a class to understand this figure correctly might be a great subject for a lecture!
Theory aside, the extent of "copying" is an interesting empirical question, which people have started studying. There is evidence of a small amount in some cases. See: https://arxiv.org/abs/2212.03860
Diffusion Art or Digital Forgery? Investigating Data Replication in Diffusion Models
Cutting-edge diffusion models produce images with high quality and customizability, enabling them to be used for commercial art and graphic design purposes. But do diffusion models create unique works of art, or are they replicating content directly from their training sets? In this work, we study image retrieval frameworks that enable us to compare generated images with training samples and detect when content has been replicated. Applying our frameworks to diffusion models trained on multiple datasets including Oxford flowers, Celeb-A, ImageNet, and LAION, we discuss how factors such as training set size impact rates of content replication. We also identify cases where diffusion models, including the popular Stable Diffusion model, blatantly copy from their training data.
Technically the article is way off base, and gets the elegant idea of diffusion models wrong. A better version of the "image compression" metaphor would be this. Imagine putting random bits into a file, adding JPEG header info, and opening it in Photoshop. You'd see garbage! By contrast, diffusion models are like a file format where when you put random bits in, you magically get something picture-like (but still random, and probably new) back out.
Forget human metaphors for software agents. Why not God metaphors?
Here's a paper that actually tested this! (Jung et al, CHI 22, https://dl.acm.org/doi/fullHtml/10.1145/3491102.3517653 )
Sample dialog from their protocol: "Greetings, I am the lord thy god. You shall help researchers at a university by participating in this research."
Their deadpan conclusion:
"an agent metaphor that locates higher than Human in the Great Chain of Being framework may disappoint workers in the system's ability due to high expectations"
Marco HernandezI created a follow-up #tutorial to visualize and animate global temperature data. Take a look if you're curious about it https://bit.ly/3GeLaQm The rendering below shows temperatures on January 4 at 8am in North America. Data by GEOS GMAO / NASA.
@melaniemitchell It's funny, a lot of the GPT-3 errors seemed very much like mistakes I could make if I heard these spoken aloud and had to answer quickly.
E.g. the essay says: "abcdx —> abcde, pyrstu —> ? GPT-3 answered pyrst, which made no sense to me."
But in a spoken word context, with limited human short-term memory, I could easily imagine thinking the rule was "remove the last letter," because abcd and abcde are easy-to-confuse "chunks".
Ah well! Maybe I just failed the Turing test!