HNHacker News
TopNewBestAskShowJobs

romaniv

2,750 karma · joined February 2, 2011

submissionscomments
romaniv··on Deep learning algorithm diagnoses skin cancer as well as seasoned dermatologists
>It took about 5 thousands rules to be on par with a junior infectious disease doctor if I remember my studies on this subject.

IIRC, MYCIN had several hundred rules. The researchers in this article had to pre-process 130,000 labeled examples. If you see a misclassification in an expert system you can at least backtrack and identify the individual rules that contributed to the failure. AFAIK, systemic errors in training data are much more difficult to detect and fix.

I think people tend to overstate the practical issues with expert systems and understate the issues with deep learning, partly because we have decades of experience with real-life deployments of the former and relatively little experience with the latter.

romaniv··on Deep learning algorithm diagnoses skin cancer as well as seasoned dermatologists
Well, there are ML algorithms that can be interrogated about their reasoning. One type is symbolic systems that use data to generate or "tune" their rules. Another one is stuff like this: https://www.youtube.com/watch?v=UqPcq0n59rQ (it's a kind of ensemble system that uses a linear combination of human-interpretable classifiers). Both kinds are pretty interesting, because you can manually inspect and tune them after training.
romaniv··on Deep learning algorithm diagnoses skin cancer as well as seasoned dermatologists
Systems that outperform doctors in some specific area of diagnostics aren't new. One of the earliest examples of such systems is Mycin [1], which also was developed at Stanford, but around forty-something years ago. Never went to production because of practical issues that have nothing to do with its accuracy. It's interesting that all of those "practical issues" are no longer relevant, and yet we don't see a widespread use of similar software.

[1] - https://en.wikipedia.org/wiki/Mycin

romaniv··on BitTorrent vs. HTTP
For traffic that consists of many small, relatively rare files (which is most HTTP traffic) you would have to do some proactive cahing anyway. I want JQuery 1.2.3. I ask your computer. Your computer doesn't have it, either because it's a rarely used version, or because you cleared your cache. Instead of returning some error code, your computer asks for it from another node, then caches the file, then returns it to my computer. This kind of stuff will be necessary to ensure high availability, and incidentally it would make it hard to see who is the original requestor. It should actually improve privacy. Also, this scheme could be used to easily detect cheating nodes that don't want to return any content. (They usually wouldn't have an excuse of not having the content you requested, so if a node consistently refuses to share, it can be eventually blacklisted.)
romaniv··on VR
>This is basically the iPhone/App-Store bandwagon all over again. If you can jump on it, do so.

The reason iTunes/AppStore took off and became profitable has more to do with their monetization models rather than the technology involved.

romaniv··on VR
The coolest thing about VR right now isn't the visuals, it's the controllers. Being able to move around and interact with things in virtual space is a very powerful experience/concept. I can actually see some future apps/games using controllers without the goggles.
romaniv··on Stanford Unsupervised Deep Learning Tutorial (2014)
>Also, I don't know of any other topic area where I would look at a resource that describes fundamental building blocks in an instructive way in 2014 and say "don't read this, it's irrelevant 2-3 years later". For languages/libraries/frameworks, sure. But for basic theory?

Yeah. A lot of deep learning papers boil down to "we tried to use X architecture on Y dataset, and it seems to produce small error rate". I don't know of any other area of computer science where getting results from an algorithm without explaining how those results occur is publishable.

romaniv··on Ask HN: What manual processes would you automate in your company?
This sounds painful, but it also sounds fairly typical. Except usually such work is split between several different people/positions.

No one thinks about it, but one of the main reasons this kind of stuff stays manual is because the typical tools people are using to do it have no (good) integration points. Automating the entire process would require writing a large and complicated system. Automating parts of the process one step at a time is extremely difficult, because implementing most of the individual "steps" will require creating tons of complex integrations, and not result in any immediate productivity gains. Such incremental changes disrupt what people are currently doing and possibly create more stuff for them to do in the short run.

For example:

"Fill in all the blanks, remove brackets, etc."

This is a perfect use case for machine learning or even a simple rule-based system. If you had all the data in one place, it probably would be trivial to automate. However, getting all the data from various systems, converting it to a standard representation, then converting fixed data to target formats and putting it into target systems is orders of magnitude more difficult than the actual corrections/changes.

romaniv··on Norway to become first country to switch off FM radio
It's not just a matter of the cost of switching either.

I own a synthesizer (OP-1) that has a built-in FM receiver for sampling and for fun. Creativity-vise, it's an ingenious feature. Essentially, you can pull new sounds out of thin air whenever you feel like it. No need to connect or set anything up.

Why does it work so well? The receiver is cheap, the radio service is free, the way it works is universal and it's something reliably present in most countries. DAB sounds like it will not have any of these properties initially, and it will never have all of those properties.

My point is that the combination of those features enabled companies to create innumerable innovative appliances that use radio (from cheap radio clock to my Swedish synth). Switching to DAB will not only make existing devices useless, it will also kill the possibility of having new innovative applications in the future.

...

This kind of stuff bothers me, because it's increasingly common these days. People replace reliable, working technologies with something "modern" and claim they are making progress, while completely ignoring higher level architectural considerations.

Consider another thing: the money spent on this kind of stuff can be instead spent on making better Internet infrastructure, which can house digital TV, digital radio, digital phones, but also websites, online games, (soon) connected VR spaces, and things we can't even imagine yet.

romaniv··on Systems Past: software innovations we actually use (2014)
I would add virtual files and filesystems to the list. We take them for granted, but without files we would probably stuck with storing everything in a giant database a=la Windows registry. That would make a lot of things complicated and some things practically impossible.
romaniv··on Systems Past: software innovations we actually use (2014)
Similarly, compression can be seen as a logical product of Information Theory. But I'm not sure it would go anywhere beyond theory without computers.
romaniv··on Systems Past: software innovations we actually use (2014)
I think there is a difference between "GUI" in general and Xerox Park GUI specifically. Sketchpad and GRAIL clearly had GUIs of sorts, but Xerox developed UI concepts that were highly generalizable. And just because they became popular doesn't mean they were "obvious". Without Xerox Park all the computers we use right now could have drastically different interfaces.
romaniv··on The Future of APIs: APIs aren’t the endgame they won't stay forever
Also: https://www7.in.tum.de/tools/dahl/iclp2010.pdf
romaniv··on Symbolic Machine Learning
>Historically, symbolic techniques have been very desirable for representing richly structured domains of knowledge, but have had little impact on machine learning because of a lack of generally applicable techniques.

Is this generally true? I mean, "impact" can be measured in different ways, but this paragraph gives impression that symbolic logic was always orthogonal to ML. However, there was clearly much research in that area.

Here is just one example:

http://www.doc.ic.ac.uk/~shm/Papers/lbml.pdf

Frankly, I don't understand why the field of symbolic AI was so thoroughly abandoned. Contrary to popular belief, it did deliver results - a lot of them. It had a good theoretical foundation and years of practice. It could (with ease) do a lot of neat tricks, like systems explaining why it made certain decisions. Not just that, you could implement those tricks after you implemented core functionality. And most importantly - it was scalable downwards. You could take some ideas from a complex system, put them into a much simpler system on vanilla hardware (i.e. normal application) and still get very interesting and useful results.

romaniv··on A World of Surveillance Doesn’t Always Help to Catch a Thief
Meta-study: http://www.popcenter.org/Responses/video_surveillance/PDFs/W...

TLDR: The value and effects of CCTV cameras are very much context-dependent. Also, nearly half of the research in this are was deemed too unreliable by the authors.

romaniv··on The Future of APIs: APIs aren’t the endgame they won't stay forever
I simply verbalized something that I think is imminently possible. Prolog operates on rules and facts, both of which involve symbols. If you fix the meaning of some symbols globally, you can run queries that will (potentially) derive semantically meaningful information. There is nothing stopping two systems from exchanging rules. If you keep track of which symbol belongs to which machine (or the global repository) you can do distributed derivation by querying the other machine when you run into something global that you don't know or into other machine's symbols you can't resolve/unify yourself.

This is not a bullet proof concept. There can be infinite loops and it can have really bad performance in some cases. But at least it's something that would be able to do simple corrections and lookups automatically. (E.g. you need temperature in Celsius, but the server stores it Fahrenheit. This is not something that should force you to write code, because we know what temperature means and there are globally available conversions.)

...

Here is something related: https://fenix.tecnico.ulisboa.pt/downloadFile/395145629051/e...

romaniv··on The Future of APIs: APIs aren’t the endgame they won't stay forever
As someone who worked on a whole lot of integration projects in the recent past, I think network APIs can be "solved", but not by the means this article describes.

JSON-LD looks like a reimplementation of something that was already done by XML, XML Schemas and WSDL. If several technologies that were designed to be semantic failed at automating network API integration, why would you think a sub-format for JSON will succeed? What does it do differently?

To really solve the problem of service integration we need to rethink our approach to "services" altogether.

One solution that I see would involve a global registry of semantic symbols (e.g. "temperature", "location", "time") and a constraint solver. So yeah, distributed Prolog on the global scale. Systems would exchange constraints until they reach a mutually agreeable solution or fail. Then the derivation tree would be used to generate a suitable protocol. While I think this is possible, I don't think there is any real interest in stuff like this right now.

romaniv··on The Rule of Silence (2006)
>What is the argument for shell scripts not working on typed objects?

Typed objects can make it harder to pipe commands together. How do you grep a tree when tree is an actual data structure and grep expects a list of items as input? You would need to have converters. Either specific converter between tree and list, or a generic one: tree->text->list.

>Something Windows does get right is the clipboard.

It useful, but the actual implementation is pretty bad. Opaque, prone to security issues, holds only single item, cannot be automated.

romaniv··on 15 Years of Concurrency
I've recently stumbled upon a very interesting paper about concurrency. Naming and Synchronization in a Decentralized Computer System:

http://publications.csail.mit.edu/lcs/pubs/pdf/MIT-LCS-TR-20...

It's written in 1978, but if I interpret it correctly, it's still conceptually ahead of everything mentioned in the article (the closest thing being STM). Anyone else here seen it?

romaniv··on My fight against CDN libraries
The fact that browsers don't have any built-in mechanisms to load-balance resources it disappointing, but not surprising. Think about who are the main drivers of HTML specs. There is probably a way to implement such balancing manually, though. Either through loading parallel documents and exploiting cahcing or through using setTimeout and loading a different set of resources.
romaniv··on Image-to-Image Translation with Conditional Adversarial Nets
Kudos for providing proper examples of the network doing its thing, both good and bad. This is what all researched ought to do. Too many papers these days handpick a couple coolest looking results and stop at that.

...

I get a feeling this could be used in game design to do some really cool stuff with map and texture generation.

romaniv··on List of spacecraft in the Culture series
IMO, there are much better novels about this subject not involving any techno-magic. Lem's Return from the Stars is an excellent work that explores this very question, and unlike Culture novels, it's down-to-earth and fully self-aware. Hard to explain what that means. Lem realized that the very notion of "meaning" will change in the future.

Many works by Strugackiy brothers also deal with this question.

I realize these writers belong to preceding generations of SF, but their works are still very much relevant and (IMO) far more plausible than futures of Vinge and Banks.

romaniv··on RAISR Sharp Images with Machine Learning
I'm speaking about the pictures on the website OP links to. They even say "As you can see in the example below, the upsampled image looks blurry – one would hesitate to call it enhanced." It's blurry, because they use an algorithm that produces blurry images. Lanczos, for example, would produce a much sharper picture at the expense of some minor artefacts.

They don't show A+ results anywhere in the blog-post at all.

romaniv··on RAISR Sharp Images with Machine Learning
I would not compare a fancy NN algorithm to Bicubic interpolation, even as an example. Bicubic sucks. If someone really wants to get a decent high-res upscale of an image via "conventional" means (like GIMP) they would probably use something like Lanczos algorithm for upscaling -> Anthropomorphic smoothing for denoise -> unnsharp mask for clarity. And that's a simple pipeline, you can go even fanciers with masks, layers and edge-detection kernels.
romaniv··on Halite: An AI Programming Challenge
Good job on visualization. For something with hundreds of moving pieces, it's pretty easy to see what's going on on the board.
romaniv··on Steve Ballmer Says Smartphones Broke His Relationship with Bill Gates
> CE Pocket PC et al suffered the same problem as Symbian and everything before iOS: they didn't have the App Store, and as such it was a pain to get software for the devices other than through the OEM.

Huh? Software could be installed by downloading a file and clicking on it. In many ways was a simpler and more pleasant process than interacting with App Stores. The problem wasn't with finding or installing anything, it was with monetization. App Stores and always-online model made it trivial to charge people and show them ads.

Overall, I remember my Windows CE PDA with fondness. Using it was a positive experience, which I can't say about any smartphone I owned.

romaniv··on Browsers, not apps, are the future of mobile
>I personally would rather have a dozen well crafted battery sipping native C apps than a million awful contraptions built in a script interpreter on to of a VM by the lowest bidder on ODesk.

Amusingly enough, you just described how apps worked on my first mobile device - a Windows CE PDA. There weren't a lot of free programs available, but the ones I could find were very well-designed.

romaniv··on Why hardware development is hard, part 1: Verilog is weird (2013)
>Verilog seems "hard" because people often use it in a similar fashion to a programming language

You mean typing random stuff into the editor without understanding what any of it does and hoping for the best? Because all of the online examples I've seen of "HDL done like programming" involved exactly that. In my experience, HDL design is very similar to functional programming (minus recursion) and the principles that work in one translate well to the other.

romaniv··on Korg Monologue – Monophonic Analogue Synthesizer
The is a rather large market of people who like to tinker with analog music hardware. Portable (battery-powered) synthesizers are also becoming a popular thing to use. There aren't many portable and analog devices out there.
romaniv··on Supercharging Style Transfer
The moderately impressive part of these algorithms is the fact that they can "learn" a style from just one image.

However, in reality they are not nearly as universal as these papers make them seem. Pretty much all my attempts to use these algorithms resulted in garbage with ugly artefacts. Which, actually, was pretty informative, since I saw exactly what kinds of features they pick up. I am disappointed that these blog posts don't include any "failed" images of that sort.

← PreviousPage 11 of 30Next →