HNHacker News
TopNewBestAskShowJobs

heyitsguay

1,081 karma · joined January 5, 2017

submissionscomments
heyitsguay··on Unity Learn platform free for three months
Whoa, JavaScript support is deprecated? News to me! I always use C# anyway but still seems like a shame.

And yeah, agreed, it's probably pretty clunky compared to what some WebGL-specific frameworks might give you. I don't know if anything out there has quite the UI setup or polish as Unity though. Do you know of any?

Edit: Saw your edit. That sounds super cursed.

heyitsguay··on Unity Learn platform free for three months
Unity has Javascript support, and can target WebGL for its builds.
heyitsguay··on AI for generative design: Plain text to 3D Designs
Do you think the resulting meshes are amenable to auto-decimation tools? Also (directed to the previous poster), while VR applications sound cool, i must admit my interest really lies in improving the asset generation tools themselves. The work done in the original post is super inspiring, I'd love to see a more robust asset generation pipeline built up from tools like that.
heyitsguay··on Failure of Self-Driving Truck Co Tells You All You Need About Self-Driving Vehs
As far as i can tell, investment helps with getting bigger, better research and engineering teams together.
heyitsguay··on AI for generative design: Plain text to 3D Designs
Cool! I've got some Unity VR experience, though i don't know much about incorporating ML pipelines into it. I'm guessing there's resources out there? So basically you'd want speech -> text -> object?
heyitsguay··on AI for generative design: Plain text to 3D Designs
Amazing! I do work in 3D computer vision and I've really wanted to get into 3D generative asset design as a side project at least. If the author or anyone else is interested in collaborating, I'd be down to talk and shop something.
heyitsguay··on Covid-19 Open Research Dataset
Watson is basically a brand name for IBM's data analytics consulting services. My understanding is they're not that great at it, they haven't scored any major wins outside of that Jeopardy run. I don't have any articles on hand but i seem to recall reading about some failures with a medical partner in particular, but then that's been a tough field for other big names like Google, too.
heyitsguay··on Rational Math Catches Slippery Irrational Numbers
Great layman explanation! I'd suggest one extremely minor modification to your ping-pong metaphor: Cauchy sequences don't necessarily have to bounce back and forth between being smaller than your target point and being larger than your target point. They can always increase, or always decrease, too. To give a concrete example, one possible Cauchy sequence for the square root of 2 would just be {1, 1.4, 1.41, 1.414, 1.4142, 1.41421, ...}
heyitsguay··on Court outlaws German Weather Service's free weather app
Don't German citizens pay for this information through taxes? Is this different from forbidding citizens from using the police instead of private security companies because they're providing security services "for free"?
heyitsguay··on Covid Notes from an Intensivist in the Seattle
If you ignore the medical technicalities which seem primarily intended for other doctors, the tl;dr is that social distancing ASAP is vital, and being young and healthy isn't a guarantee that you won't get very sick or die. Don't panic, but take it seriously.
heyitsguay··on US admits 'failing' on testing, says Fauci
Sure, I meant for the non-technical aspects that the previous poster mentioned - chain of custody, sample integrity.
heyitsguay··on US admits 'failing' on testing, says Fauci
Are there any differences for Covid-19 samples versus any of the other samples such labs would be routinely testing though?
heyitsguay··on Before Clearview Became a Police Tool, It Was a Secret Plaything of the Rich
I think they just meant marginalized groups, not People Who Are Bad.
heyitsguay··on WHO Director: "Covid-19 does not transmit as efficiently as influenza"
Interesting... i wonder what the data is to back up that assertion?

Another interesting tidbit in there was that very few cases (1%) appear to be totally asymptomatic, and that most infections apparently start showing some symptoms after a couple days. So there are a few assertions in total that seem to depart from what was previously understood, at least through mass and social media.

heyitsguay··on Ask HN: Who is hiring? (March 2020)
And as a separate add-on because the qualifications are pretty different:

National Institutes of Health | Postbaccalaureate Trainees | Bethesda, MD | Onsite | Full-time

General information: https://www.training.nih.gov/programs/postbac_irta

The Laboratory of Cellular Imaging and Macromolecular Biophysics (LCIMB, lab chief: Richard Leapman, PhD), National Institute of Biomedical Imaging and Bioengineering (NIBIB), at the National Institutes of Health (NIH) in Bethesda, MD is seeking two postbacs to join a research program for biomedical computer vision research for electron microscopy.

Successful candidates will work alongside a team of LCIMB machine learning researchers to develop new computer vision techniques applicable to large-scale 3D structural modeling of cells and organelles. Depending on interests and background, postbacs may assist in creating training data, developing software to enable microscopists to apply new research tools to their biological imaging problems, and conducting computer vision research of their own within the scope of our lab’s projects.

The postbaccalaureate program is open to recent college graduates who wish to spend 1-2 years doing research before continuing to graduate school. This program is open only to USA citizens or permanent residents. Please see the Eligibility heading at https://www.training.nih.gov/programs/postbac_irta for full eligibility criteria.

The ideal candidate will have a BA/BS in computer science, applied math, or a similar area, and will have prior experience in image processing, computer vision, and/or machine learning. Postbacs will be working on Linux workstations with Python and common scientific computing and deep learning libraries, so experience with these tools is preferred. Candidates should be able to effectively communicate their work to a multi-disciplinary research team of biologists, microscopists, and data scientists.

To apply, pplicants should complete the online postbac application at https://www2.training.nih.gov/apps/publicforms/pbt/forms/log.... Applicants should also contact Dr. Richard Leapman at leapmanr@mail.nih.gov to indicate their interest in joining our lab. Anyone with questions can message me at matthew.guay@nih.gov.

heyitsguay··on Ask HN: Who is hiring? (March 2020)
National Institutes of Health | Postdoctoral Fellow | Bethesda, MD | Onsite | Full-time

The Laboratory of Cellular Imaging and Macromolecular Biophysics (LCIMB, lab chief: Richard Leapman, PhD), National Institute of Biomedical Imaging and Bioengineering (NIBIB), at the National Institutes of Health (NIH) in Bethesda, MD is seeking a postdoctoral fellow to join a new research program for biomedical computer vision research for electron microscopy.

Successful candidates will work alongside LCIMB staff scientists to research new computer vision techniques applicable to large-scale 3D structural modeling of cells and organelles, with a focus on areas including but not limited to: volumetric segmentation, robust multi-dataset segmentation, and self-supervised pretraining for computer vision neural networks. They will help supervise graduate students, postbacs, and interns working on the project, and guide team members who will integrate new research developments into software tools that microscopists can use for biological research applications.

To be eligible for postdoctoral training at the NIH, individuals must hold a doctoral degree (PhD, MD, DDS, or the equivalent) and have no more than five years of relevant research experience since receipt of their most recent doctoral degree. U.S. citizens and permanent residents are appointed to Postdoctoral Intramural Research Training Awards (IRTAs); citizens of other nations are appointed as Visiting Fellows.

We are looking for candidates with a PhD in computer science, applied math, or a similar area, with a focus on image processing, computer vision, and/or machine learning for imaging applications. Candidates should be comfortable working in a Linux environment, with basic Bash scripting experience. Candidates should also be familiar with Python and common scientific computing and deep learning libraries. Experience with high performance computing is preferred. Exposure to biology is preferred but not necessary. Supervisory experience is also preferred - for current graduate students, supervision of undergraduate or summer students is ok. Candidates should be able to effectively communicate their research activities to a multi-disciplinary research team of biologists, microscopists, and data scientists.

To apply, applicants should submit their curriculum vitae, a cover letter describing research interests, and names and contact information for three references to Dr. Richard Leapman (my boss) at leapmanr@mail.nih.gov. Anyone with questions can message me at matthew.guay@nih.gov.

The NIH is dedicated to building a diverse community in its training and employment programs. We encourage applications from members of underrepresented or marginalized communities.

heyitsguay··on FastMRI leverages adversarial training to remove image artifacts
I just gave a talk on similar work being done in microscopy: https://leapmanlab.github.io/nihai/jan20/ .

The tl;dr (in microscopy but apparently also in mri) is AI imaging can evidently enable new concrete solutions to intractable imaging problems, but the failure modes are really treacherous. The example on slide 39, taken from another excellent review paper, does a great job illustrating the problem. I think these methods will get more trustworthy, but i wouldn't stake my life (or my paper's prestigious research results) on them at the moment.

heyitsguay··on Generative Teaching Networks: Accelerating Neural Architecture Search
Probably just an empirical comparison with other NAS strategies and hand crafted architectures right? The whole area of research is still ruthlessly empirical.
heyitsguay··on Generative Teaching Networks: Accelerating Neural Architecture Search
That sounds very different from what Uber is doing here, which is basically accelerating training with synthetic data to accelerate otherwise standard neural architecture search tools. The focus is on the data synthesis network.

Also, the system you describe sounds impractical for any of the complex learning tasks you suggest, especially if it hasn't even done much simpler things yet. Why would machine code be the right level of abstraction for a vision or robotics problem?

heyitsguay··on Growing Neural Cellular Automata: A Differentiable Model of Morphogenesis
I'm wondering these things too. You say in the paper that your net is quite small - I wonder if you could first extend it to regenerative sprite animations for some fun procedural NPCs in a roguelike? If you can efficiently add or change a pixel or two per frame in a sprite+collision mask, imagine battle sequences where things get hacked off, maybe regrow if you wait too long, etc.
heyitsguay··on Sometimes, requiring explanability is counterproductive
I think the type of explainability Yann LeCun is describing in this post is not the same as what people really want when they talk about explainable AI. To use his example, I think most people would be fine with an AI that said "we're doing X because that's what heuristic Y prescribes for the situation", and getting from where we are now to something like that is the challenge. FWIW i see a similar dialogue in biomedical contexts quite a bit, where there's some cool but hard-to-trust work being done in data-driven image restoration (https://www.biorxiv.org/content/10.1101/236463v5.abstract). There's still a lot that could be done short of programs providing a priori mathematical proof for all models.
heyitsguay··on Let's remove quaternions from every 3D engine (2018)
Computable functions have that same property, and that's the space that programs work in. It doesn't matter if the composition elements are linear, only that they can be computed on your hardware within some time budget.
heyitsguay··on Let's remove quaternions from every 3D engine (2018)
You don't need to restrict yourself to mathematically-nice algebras of operations or whatever if you're doing 3d graphics though. If you want to represent a translation of x by vector v in 3D, just do x+v.
heyitsguay··on A Proof of Useful Work for Artificial Intelligence on the Blockchain
Right, I understand that. I'm saying from the point of view of the task submitters, the PoW approach is unnecessary and adds time and cost overhead to what is ostensibly the main goal of their participation: getting an ML model trained.
heyitsguay··on A Proof of Useful Work for Artificial Intelligence on the Blockchain
If the order were deterministic that would be true, but I haven't found any solutions to control GPU task scheduling at that level. Maybe one exists? I don't know enough about low-level GPU coding to say.
heyitsguay··on A Proof of Useful Work for Artificial Intelligence on the Blockchain
If your algorithm uses reduce-like ops for float values and is run on a GPU, your results won't be deterministic even with controlled RNG seeds, since floating point addition is non-commutative.
heyitsguay··on A Proof of Useful Work for Artificial Intelligence on the Blockchain
Wouldn't it waste even less energy to not use a PoW scheme at all though? Just set up a webpage that connects people with tasks to people with hardware.

On another note, digging in closer, the proposed training scheme (Section 4.3) sounds like it would be incredibly inefficient for any large neural models - you're computing your own updates via gradient descent, but also getting compressed updates from peers on the network, and then broadcasting your own updates out. For reference, at work I might be training a model with, say, 30M parameters, and I get about 4.5 training iterations per second on a GTX 1080. What would that drop to with peer broadcasting added in?

They mention only sending weight updates that under/over-flow a certain value, is that enough to avoid, e.g., a 5x drop in training speed? Given that training time is already a big bottleneck for ML work, this seems critical. Section 7.2 seems to address this, referencing Supplementary Materials (Appendix F). The numbers there don't look great - with the optimizations they mention, for GPU timing in Figure F.3 it looks like local training ops account for about 21ms, and the protocol overhead is something like 100ms. I also don't see an apples-to-apples comparison of "here's how long this example takes to train locally vs. with this protocol", which isn't a good sign.

heyitsguay··on A Proof of Useful Work for Artificial Intelligence on the Blockchain
For people submitting tasks, it seems like this would be hard to make cheaper than, say, spinning up an AWS instance or similar. Is that assessment correct? If so, would there still be an incentive to use this setup?

Edit: I see they address this, claiming 30% cheaper than cloud services and better ROI for miners than Bitcoin. That said, wouldn't it still be more efficient to skip the blockchain stuff and just have a marketplace where people with tasks can rent time from participating devices and pay with whatever currency?

heyitsguay··on Wolf Pups That Play Fetch
You're probably being downvoted for telling a just so story: https://en.wikipedia.org/wiki/Just-so_story . It's a particularly pernicious trap when trying to understand behavior and evolution!
heyitsguay··on Digital Twins Are Reinventing Innovation
Cool! What companies do the twinning? Is it handled in-house by the companies applying it or are they partnering with specialists?
← PreviousPage 12 of 15Next →