_hark

14 ms·

31.
Symmetric Power Transformers(manifestai.com)
18 points by _hark 2y ago|5 comments
32.
by _hark 2y ago
I worked on this experiment as an undergrad ~10 years ago during my freshman year! We built a Cherenkov radiation detector, focusing magnets, and did tons of simulations. This is all from memory, but I remember the beamline setup was to get
33.
by _hark 2y ago
We could discuss completeness of the sample space, and we can also discuss completeness of the hypothesis space . In Solomonoff Induction, which purports to be a theory of universal inductive inference, the "complete hypothesis space&
34.
by _hark 2y ago
I think what you're getting at is the construction of the sample space - the space of outcomes over which we define the probability measure (e.g. {H,T} for a coin, or {1,2,3,4,5,6} for a die). Let's consider two possibilities: 1.
35.
by _hark 3y ago
NNs are typically continuous/differentiable so you can do gradient-based learning on them. We often want to use some of the structure the NN has learned to represent data efficiently. E.g., we might take a pre-trained GPT-type model, a
36.
by _hark 3y ago
https://archive.is/HgqAy
37.
by _hark 3y ago
Sorry, but this is just incorrect. Go players have gotten stronger over time overall [0][1], and AI discovered many new ideas that all top pros have incorporated into their game-play (idk how to give a source for this, it's just very w
38.
4000x Speedup in Reinforcement Learning with Jax(chrislu.page)
131 points by _hark 4y ago|30 comments
39.
by _hark 4y ago
jax.vmap() is all you need?
40.
Radiance Fields in Space, Time, and Appearance(sarafridov.github.io)
7 points by _hark 4y ago|0 comments
41.
by _hark 4y ago
These kinds of non-constructive results are almost always present when you invoke choice (when it's actually relevant), and this is a well-known example of the non-intuitiveness of choice. If you don't like this, you're proba
42.
George Hotz Potentially Taking Internship at Twitter(twitter.com)
23 points by _hark 4y ago|13 comments
43.
by _hark 4y ago
This is basically a re-implementation of the recent work from OpenAI[1], where you first collect high-quality observation-action-observation transition tuples, train a model to predict the actions between observations, then use that trained
44.
Biofire exits stealth-mode to bring smart guns to market(techcrunch.com)
1 points by _hark 4y ago|0 comments
45.
by _hark 4y ago
DeepMind to OpenAI and everyone else[1]: > your hyperparameters are bad and you should feel bad It's amazing to me that such a big goof was missed by so many for so long. All these multimillion dollar language models and people just
46.
by _hark 5y ago
Would be great to see a "Hacker News Approved" buyer's guide for various products. Like Wirecutter but written by people who actually know what they're talking about.
47.
by _hark 5y ago
This work from Berkeley + Google Brain describes using RL and GANs for transferring motion styles from animals successfully onto robots. Super neat idea, and love to see ML being used more and more on real robots! Love the Cost of Transport
48.
Making Robots Move Like Animals(sites.google.com)
2 points by _hark 5y ago|1 comments
49.
by _hark 5y ago
Hmm, I don't think I misunderstood? I get that they're using MuZero to decide the bitrate for equivalent perceptive quality as a function of the content. Sure, once they decide on that using MuZero it's a valid compression an
50.
by _hark 5y ago
While an interesting use of applied RL, in some sense isn't this just another way to cast the compression/compute tradeoff? I.e. can't we just achieve the same effects by using another compression scheme which trades off loca
51.
by _hark 5y ago
Control policies learned via RL are starting to work in the real world! Typically policies learned via simulation tend to transfer poorly to the real world (the so-called sim2real gap), so I'm curious to dig into this work to see how t
52.
Huge Step Forward in Legged Robotics from ETH(youtube.com)
1 points by _hark 5y ago|1 comments
53.
by _hark 6y ago
You can make KataGo play moves that keep the score roughly even since it has a trained score head, e.g. kataJigo [1]. This will keep the game even to your level, a nice way to train. [1] https://github.com/sanderland/ka
54.
Seismic GAN: a walk through latent space(brantondemoss.com)
2 points by _hark 6y ago|0 comments
55.
by _hark 6y ago
Sounds interesting, I'll take a look. If you want to email me (grab from my website) so I can ask some questions about the data (if you have expertise here) that would be helpful.
56.
by _hark 6y ago
Wow I had no idea. Thanks for the info and link. The seismic community has been producing plots like these for decades, but they're just referred to as "wiggle plots", not a great name. Ridgeline seems like the best option go
57.
by _hark 6y ago
What's the horrible meaning of the phrase 'joy division'? Do people just not like that band anymore?
58.
by _hark 6y ago
Along these lines, I work in geophysics/AI right now and trained a GAN on some seismic data, which people usually visualize with ridgeline/joy plots (literally just realized people call them joyplots because of the joy division co
59.
by _hark 6y ago
What style of bot did Honinbo Warrior implement?
60.
Go AI Past, Present, and Future(brantondemoss.com)
1 points by _hark 6y ago|0 comments

More ›