1/ Tragically, the famous mathematician John Conway has recently passed away to COVID-19. To commemorate his life, here's a thread about his most famous creation, the Game of Life, which raises many provocative questions about the nature of complex systems.
2/ The Game of Life is a system consisting of an infinite grid of squares, that can be either black (alive) or white (dead). Every time step, the squares evolve according to three simple rules:
3/ 1) Any live cell with two or three live neighbors survives. 2) Any dead cell with three live neighbors becomes a live cell. 3) All other live cells die in the next generation. Similarly, all other dead cells stay dead.
4/ With these three simple rules, you can get all kinds of emergent patterns: some stay still, some oscillate, some move across the grid, and so forth.
5/ For example, the glider pattern will fly forever:
6/ You can build a "glider gun" that will spit out infinite gliders:
7/ And you can even make a "breeder" that will blaze a trail across the grid, spawning glider guns in its wake, each of which will spit out gliders forever:
8/ Though the Game of Life has been studied for decades, people are still learning more about what you can do in it.
9/ For example, in 2010, a pattern was discovered that can simultaneously clone and delete itself... every 34 million generations! It uses an "instruction tape" that guides the cloning of the pattern, including the cloning of the tape itself.
10/ And you can even build a universal Turing Machine in the Game of Life, complete with ticker tape. Thus, you can compute *any* computable function within it.
11/ If you believe the brain to be a computer (in the classical sense), you could therefore theoretically simulate a brain within the Game of Life.
12/ And since the rules for running the Game of Life are themselves computable, you can also build a simulation of the Game of Life *within* the Game of Life:
13/13 Many lessons to learn from this, but to me it's a humbling reminder that you can get amazingly complex and unpredictable behaviors from the simplest parts. #conway#gameoflife
• • •
Missing some Tweet in this thread? You can try to
force a refresh
NEW PAPER 🧵: Deep neural networks are complicated, but looking inside them ought to be simple. In this paper we introduce TorchLens, a package for extracting all activations and metadata from any #PyTorch model, and visualizing its structure, in as little as one line of code. 1/
The core function is log_forward_pass. Just pass in any PyTorch model (as-is, no changes needed) and input, and this one line of code gives you a data structure with activations and metadata of any layer, along with an automatic visualization of the model’s computational graph 3/
1/ NEW PREPRINT #2: Your visual experience is full of colored shapes. How does your brain encode colors, shapes, and the way they’re combined? biorxiv.org/cgi/content/sh…
2/ In this study we examined 1) do separate brain regions encode these features, or are they encoded by the same regions? And 2) if the latter, how are these features encoded relative to each other: independently or interactively?
3/ We applied fMRI MVPA to responses in early visual cortex and higher-level ventral stream regions sensitive to either shape or color, examining stimuli varying in color, and either orientation or curvature.
1/ NEW PREPRINT: Much work has examined the “binding problem” of how the brain encodes combinations of visual features. But how do convolutional neural networks do it?
2/ In this paper, we looked at how CNNs encode color/shape combinations: are these features encoded independently or does the representation of one feature interact with the other feature?
3/ Specifically, we used a variation of RSA to look at whether encoding of color varies across different shape features.
@General_Jagoff @MattWalshBlog Go to the actual study in PNAS. It was so deeply flawed that the authors had to issue a correction specifically stating that it doesn’t show what people are taking it to show (that black people aren’t more frequent victims of violence).
@General_Jagoff @MattWalshBlog Meanwhile there are plenty of carefully controlled studies showing that black people *are* disproportionate victims of violence... e.g. journals.plos.org/plosone/articl…
1/ Just finished the fantastic (if pretentiously titled) biography "William James in the Maelstrom of American Modernism" by Robert Richardson. James is one of those people who grabs your attention and gets more interesting the more you learn about them. Some thoughts
2/ First, and most obviously, the guy is phenomenally quotable, on any number of topics... a smattering of examples (some of his better ones are longer than Tweet length and rely heavily on his obsession with italics, unfortunately):
3/ On socializing and the self: "A man has as many social selves as there are distinct groups of persons about whose opinion he cares."
Pointing out hypocrisy in political opponents can be fun/satisfying (feels like a "knockdown argument"--got em!) and has its uses but 1) it often replaces the work of making the case for an alternative vision, promoting laziness (1/4)
2) "inconsistent" isn't the worst thing an ideology can be (e.g. consistent sociopathy/sadism is bad; why should "at least they're consistent" always have a positive connotation?), and (2/4)
3) when overemphasized it can lead to timidness/cynicism about having costly principles at all (South Park effect); can't be called a hypocrite if you never commit to something that sometimes demands personal sacrifices. (3/4)