Mikio L. Braun

ML and AI expert

@mikiobraun

@mikiobraun Twitter Memorial

19,827 tweets · 2008–2024 · 1046 threads

2010

RT @XKCD386: MMXI is this year's new netspeak. I give it about 365 days until no one cares about it anymore
Wer macht eigentlich die Personalplanung bei der Post? 7 Angestellte für ca. 200 wartenden Leute #fail
Replying to @ikai
Oh yes. Young communities seem to stuck to project names emphasizing the language. Java seems to have outgrown that only recently.
@michaelgrube the site seemed legitimate (some blog). Still I don't understand why anyone would put something like that on his server...
Website claims my computer shows infected behavior, recommends anti-virus tools, requires capcha to proceed. Based on what? My GET requests?
I wonder how long pinboard is sticking to its "sign-up fee = number of users * 0.001$" formula....
In case you're wondering, mloss.org and mldata.org are down till Sunday evening due to university wide network maintenance....
Replying to @ogrisel
Yes, we know. The whole infrastructure of our university is taken down for maintenance thus weekend :(
Booted into Win 7 to do some jblas work. Still update and reboot cycling after 20 min. #fail
By popular demand (= a colleague of mine complained once about 3 months ago), my blog's RSS feed now contains the whole post again.
Replying to @chl
@thinkberg Ich sehe, ich war nicht der einzige, der Fotos gefordert hatte ;)
Replying to @ChrisDiehl
You can upload any data type you wish. For some formats (e.g. ARFF) a bit more parsing is done.
I guess it's time for the year-end purge of unused blog post ideas. (Yes, this is a warning ;) )
While @thinkberg is in Tokyo, reading his copy of the Cluetrain Manifesto. Funny to read all that after Web 2.0 already happened.
New @twimpact backend pretty stable, although it seems we need to tweak a number of configs still. A few more servers could help, too ;)
Meanwhile, I'm looking at symmetric generalized eigenvalues contributed by Nicholas Oury for JBLAS.
. @thinkberg and I are patching together all the stuff to bring the new @twimpact infrastructure online... .
Replying to @GoogleAI
Oh no, @googleresearch is giving away Galaxy Tabs at #nips2010 and I'm not there... ;)
Test-driving our new improved trender for @twimpact. As always, real performance boosts come from better algorithms, not faster hardware ;)
RT @discoproject: Disco 0.3.2 is released! New features include chunking, user-defined attributes and authentication for DDFS
On the other hand, since I'm using debian stable, my OpenOffice is already two years old, of course... .
I'm always amazed at how OpenOffice manages to replicate the Windows User Experience down to all those little inexplicable glitches.
Methinks this Winter is taking his job quite seriously. Good to see a pro is in charge. Sort of.
2
Getting more requests for jblas on 64bit Windows. ATLAS is out of the question. Pure FORTRAN lapack-lite still better than nothing?
Received my copy of "Business Model Generation" Layout-wise reminds me of my daughter's picture books ;)
Replying to @ChrisDiehl
In retrospect, I think the biggest question is how to couple noisy components in a robust manner.
RT @lemire: Are we going to have a picture of Jimmy Wales forever on top of every wikipedia page? How much money must I pay to make it g ...
Hm... Then again, it's really hard to say :( Do writes distribute evenly over column families? Or do you only write to a few?
After some more testing: @cassandra's row caches can be large (~3GB in my case), and the factor is more like *10* * MemTable size * #CF
IMHO @cassandra's heap estimate (MT * 3 * #CF + 1GB + caches) is wrong. If you have large indices, you run out of mem before the 1st flush.
1
RT @francoisfleuret: We will have MASH mugs at NIPS for the braves who will have uploaded a heuristic! (no rush, we are updating the sit ...
IMHO a centralized email subtitutes a la Facebook Messages goes against core ideas of the internet like open protocols and fault tolerance.
1 1
RT @francoisfleuret: We are Bayesian machines, and the media feed us with highly biased data, continually. That is not good.
Dear IDE coders: Please put code analysis in a background thread. It often feels like typing over a shaky 300 baud line...
It seems mingw-w64 (64bit gcc for cygwin) has finally made it into the official packages. Time to give 64bit jblas another try on Windows.
1
After some thinking and *gasp* proving I managed to drastically reduce the number of column slice requests in our @twimpact analysis code.
1
RT @twimpact: twimpact.com is catching up data right now … activating language trends next, then we will look into hashtag and generic t ...
As always, @justinbieber proves priceless in weeding out our @twimpact retweet analysis code...
I have to admit there is something about Compiz's window animations which makes me feel like I'm in fact living in the 21st century ;)
Replying to @DRMacIver
That's probably still better than compiling ATLAS ;) #frighteningitemsontodolist
Twitter is so much more fun on the bus compared to the subway: much better 3g reception to open all the links.
Finished reading "The Agile Samurai" by Jonathan Rasmusson. Very nice and lean introduction to agile software development!
Replying to @DRMacIver
You mean like "I had a very elegant proof of that remarkable fact but it didn't fit in 140 characters."
Replying to @DRMacIver
Ah yes, I know. I wanted to say "They should probably improve the compiler warnings", but it didn't fit in 140 characters ;)
Replying to @DRMacIver
No problem. I think it makes sense. Just took me a bit to figure out a bit why my code wouldn't compile... . Compiler warning?
Replying to @mdreid
Probably the same guys who said let's skip fall and directly go to winter over here in Berlin... .
@mdreid That actually happened to me. Only that the function was piecewise constant -> derivative zero almost everywhere.
Ob AVM weiss, dass @1und1 immer so tut, als wäre die Firmware der fritzbox Schuld an allem? #daskoenntemanauchgeschickterkommunizieren
Netbeans's scala plugin always grinds to a slow death after a few hours of coding. Giving IntelliJ IDEA a try...
Day Two with only two virtual desktops for coding feels oddly claustrophobic - but in a good way! ;)
Reduced number of virtual desktops from six to two to reduce preemptive context switches during coding. One for the IDE, one for the shell.
Doing some benchmarking with the new multi-threaded @twimpact architecture. Totally maxing out all resources on my poor dual core notebook.
Finally dropped maven's "jar-with-dependencies" in favor of "copy-dependencies" with some CLASSPATH magic. Much improved recompile times :D
In the movie "Social Network" Mark was always against ads because Facebook would "stop being cool". It seems that has been a long time ago.
Twitter needs a feature to search in your timeline. Looking for a link from two days ago, no chance. #hasthishappenedtoyou
1 1
Oddly, playing Mass Effect 2, going through all those loyalty missions feels a lot like taking care of your graduate students... .
Replying to @francoisfleuret
Congrats! That's great! Of course, our T-Shirts are already being printed ;)
Is it just me or is the new ubuntu font - as cool as it might look - not really suited as the default font for everything?
So true! RT @muratk3n: Programs must be written for people to read, and only incidentally for machines to execute. —Abelson & Sussman
Replying to @francoisfleuret
yes, you'd better! ;) I'll have a look at Reddit in exchange...
RT @codemonkeyism: I think the mongodb analysis shows the only real fault is with #foursquare to not watch memory capacity
RT @ccbuhr: High Level Group on Scientific Data delivers its report (bit.ly/ag7sMT) to the European Commission, @NeelieKroesEU likes it ...
Just finished preparing my talk for the "Validation in ML and Stats." workshop. Looks like it's going to be a lot of OSS evangelizing again.
Replying to @DRMacIver
Good point ;) Probably around the time Twitter and Facebook crush under their own weight.
If economy would be run like science, your competitors would have a say whether you'll get any shelf space for your new products.
1
Again back to benchmarking our @twimpact reanalysis process. Looking at detailed timing stats for tables vs. @cassandra operations now.
2
Yesterday's Facebook outage adds another argument for decentralization to the list: reliability. Ever seen email "go down"?
RT @stef: #neologism "Mergency" - the state of panic induced in team-mates when Stef suggests merging his code into the development branch
RT @kazuhito: えっ RT @oogatta: ええええええええーーーー。 RT @engadget: Microsoft: Windows Phone 7 to be GSM-only until first half of 2011 http://bit. ...
My 3 yr old daughter discloses more sensitive information to complete strangers at the playground than all Facebook policy changes combined.
1
Tag 3 ohne DSL. #1und1 zeigt immer noch keine erkennbare Aktivität. Soviel zum Thema Qualitäts-Garantie. #fail
I more and more find myself thinking science would be more effective if people used blogs, open repositories, and reputation systems more.
@kufleisch OAuth sollte aber doch für die Twitter eigene Android-App kein Problem sein, oder?
Attending a big project meeting, I'm amazed at people's ability to simultaneously discuss and do emails for hours. ;)
We're getting closer... RT @thinkberg: it is alywas strange to see a new webapp in raw (w/o css) … twimpact 2.0 trends #twimpact #nosql
Replying to @mdreid
Acutally, it's hard to tell whether you're joking or talking about real apps ;)
RT @jboner: RT @remeniuk Have just got an official permission to use Scala/Akka in production (I'm working for O2 Germany, Business Inte ...
Ok, now that @echofon has stopped working in Firefox, let's give tweetdeck another try... .
Replying to @al3xandru
My tip is to give Cassandra more memory than MemtableSize * #ColumnFamilies suggest and tune the CMS threshold properly.
@grahamashton Actually not that sure whether Alex groks Ruby fully. Misses the point how you can build DSLs using Ruby's dynamicity.
Just saw that results of my #cassandra benchmarks will be reflected in the new default settings in 0.6.6 and 0.7.
1
Replying to @ochafik
I guess that's because sbt does not need to start a new scala instance.
And another run of the #cassandra benchmarks, this time with the same version of the JVM. Thankfully, I've automated everything ;)
Having the mental capacity to write and maintain chaotic code does not qualify as being good at programming in my book.
Decided to run the @cassandra experiments in true benchmarking style. Now it's a mix of jruby/jmx4r/bash and probably some gnuplot... .
I think after this week I've done enough GC tuning for a lifetime. At least it feels like it... .
RT @graemerocher: RT @adrianmw75: Intel CEO: "We need antivirus, can someone buy me McAfee?" Few hours later: "Done." "Great, which vers ...
How can Python people argue that their indented blocks are so much cleaner and then have all that "self. ..." visual noise?
Why does rdoc take the first file as start page, but docs for multiply defined modules from the last file?
Oh no, worked with some old object files from my attempts to compile ATLAS for 64bit :( Release of jblas 1.1 pushed back to next week... .
Three reboots so far to bring my Windows 7 installation up-to-date. When will this stop? Can't just restart system services on the fly?
Alright, time to boot into Windows to cut a new release of jblas... . jblas 1.1 will have SVD, and some bugfixes.
Have to admit waiting for 4sq and Gowalla server replies is a good way to kill time, though. ;)
Been test-driving foursquare and Gowalla in parallel to understand the hype. Not sure if I really "get it"...
While watching @cassandra filling up all available RAM, I've been hacking on a caching mechanism for jblas. Looks quite promising.
1
Hm probably it's also a mix of concurrency + memory issues between native and Java code. Beautiful ;)
1
Hacking a bit on a newfangled cached DirectBuffer for improved jblas performance. All the fun of chasing memory errors you can think of ;)
Looking at @cassandra's memory usage curves reminds me of waiting for one's favorite learning algorithm to converge... .
Replying to @nikete
Ok, what if people say in papers they're better than the state-of-th-art. Hm... probably I should stop nitpicking here ;)
I wonder whether you can actually be better than the state-of-the-art? Don't you advance the state-of-the-art once you're better?
1
At least after all the tweaking, cassandra gives a solid 110 analyses per second, already with about 100 million tweets in the db ;)
#Protip for retweet analysis: @justinbieber for high frequency, @tweetmeme for largest number of distinct tweets.
1
Seems like our performance issues with cassandra where mostly due to different settings for rpc and query timeouts.
Cassandra performance not quite satisfying. Memory behavior much better with lower memtable thresholds, but now it just hangs for minutes...
Seems like information on configuration in Cassandra's Wiki is sometimes outdated, e.g. changes in names of configuration parameters.
Returning from @Dagstuhl seminar on Learning in Dynamic Environments. Strong focus on Reinforcement Learning and Neural Networks.
Trying to print out my DHL online stamps. Technology nightmare: Javascript embedded into PDF wanting to call home for authentication.
Never underestimate a full-screened #emacs with speedbar as an Ersatz IDE. Little context awareness, but it's fast! ;)
Turns out @justinbieber is really useful for debugging Twitter analysis apps because he gets retweeted so often ;)
Please tell me that fsc does not get confused when used in parallel from two different directories... . #scala
Apparently, cassandra likes concurrent access (~4K w/s -> ~15K w/s with 8 threads). Now I "only" need to multithread our twimpact code.
While the heat keeps me immobilized, watching Cassandra dealing with some test data. More optimization required, it seems.
Installing gcc on Mac via mac ports. Please remind me again why we're compiling everything from scratch on earth's most controlled platform?
1
Replying to @halfthebird
Thanks for the remark. You can post to the jblas-users Google group, but I'll add the bug.
Replying to @headius
I fully agree :) Hopefully someone will eventually write a "better" maven... .
Also trying each repository verbosely... . RT @headius: Sigh. Add "watching Maven download dependencies" to that list of 3AM time-wasters.
I think most of our colleagues suspect that we're just making up those buzzwords as we go along. ;)
RT @twimpact: We're also playing around with ActiveMQ, Camel and Akka for better robustness, scalability and flexibility. ^MB
"Have you tried 'fsc -shutdown'?" is the "Have you tried to reboot your machine?" of Scala.
Programming feels a lot less like construction but more like redressing lately, mostly making different frameworks match.
1
Our MLOSS workshop just finished. We certainly had the most code per slides of all workshops ;)
Great! Free wifi at Ben-Gurion airport. Waiting for transfer to Haifa. Tonight: dinner with our workshop's invited speakers.
Looking into LIFT's ORM layer to get some inspiration for how to best use Cassandra in @twimpact from Scala.
Now that the computers are analyzing, time to work on that project proposal. Would rather prefer if it were the other way round ;)
RT @ikai: I've written a lot of hacky code over the years; what I'm about to write might just put all those hacks to shame
Replying to @raghuiyer
Thanks. I'm still learning myself and Scala's type system is really quite complex, but also quite elegant.
@thinkberg Hatten eben beim Mittagessen auch den ersten in freier Wildbahn am Nebentisch gesehen!
Just learned that navigation was unlocked on Android's Google maps in Germany. Now if only I had errands to run in Berlin ;)
Fixing issues with the Atom feed of my blog. The fun of "blogging like a hacker" ;) #jekyll
@muratk3n Guess one cannot blame a researcher for protecting his ideas for business, though. ;)
Hm... I wonder why I cannot post status updates to facebook from my browser anymore. Someone telling me to quit facebook? ;)
Just subscribed to German games on Google calendar. Time to get ready for the world cup... .
Added experimental flag --ptatlas to jblas to link against the multithreaded version of ATLAS. Up from 9.2 to 14.5 GFLOPS peak on my laptop.
3
Debugging linkage issues with jblas. I know why so few people dare to do Java projects with multiplatform native code... .
Replying to @DRMacIver
One day you'll have to tell me why you became less interested in Scala... .
Hm. Anyone else experiencing 50% packet drops en route to www.google.com at 209.85.250.1? #mtr
Just returned from a five-day trip to Tokyo to meet with Japanese social media people. I'll write a blog post once the jet lag wears off ;)
Second day of our #twimpact seminar at mitsue-links in Tokyo ends. Very interesting meeting. Always nice to talk to the actual customers!
1
Rigged up a small script with maven and pynotify to deal with long scala compilation times.
Actually I'm one of the organizers, but I couldn't have said it bettet than @mlsec #mloss #icml2010
Two hours phone of conference over a Google docs spread sheet, and our program for our MLOSS workshop at ICML 2010 is set. Stay tuned.
Also, it's not that hard to have relatively informative project site. At least put your name in there, and link to docs, API, and source.
Seriously, please package your software into <project-name>-<version-number>.tgz and unpack to <project-name>-<version-number>/ !
1
Sifting through reviews for our machine learning open source software workshop at ICML 2010... .
Had nice discussion with Marc Toussaint over lunch at our group's retreat on Bayesians vs. Frequentists. Got some ideas for new blog post.
Replying to @headius
I guess the situation might be the same for all HTC phones with HTC Sense.
Replying to @headius
Ah, no, I have an HTC Desire which comes with it's own (on-screen) keyboard. And so far, I've only seen "¦" but not the "|".
Replying to @headius
BTW, if I'm not mistaken the HTC engineers put the "¦" instead of the "|" into the keyboard. Do you think IRB could auto-translate?
Wow, our first experiments with Cassandra seem like we'll be able to reanalyze our @twimpact data base in about one week.
Hm. Sdcard just became unreadable on my HTC Desire again... I thought a Linux based device wouldn't require so many reboots ;)
Didn't knew that Twitter has its own Android app. Looks nice, supports lists. But HTC Peep has the nicer (larger!) widget!
1
No luck with the wireless (rt3090) and Ubuntu 10.04... . Back to my old custom-compiled 2.6.33 kernel... .
Replying to @ChrisDiehl
It was. At one point I was discussing how different calling conventions affect the stack with some colleagues... .
Alright, had to fix the stub generating code... . Stay tuned for a bug-fix-release of jblas.
Aha, it all seems to boil down to gfortran support c99 style direct handling of complex numbers (as opposed to g77). Argh.
Debugging a nasty bug in the JNI Java->C->Fortran interface of jblas. Next step: Compiling Fortran code to assembler. :(
Replying to @wrede
ich glaube, wenn die Google-Kontakte halbwegs aufgeräumt sind, dann klappt's auch.
Just could not resist buying "Enterprise Integration Patterns" by Hohpe and Woolf. Eventually, I should also read another ML book ;)
Replying to @ryotat
I'll be there for our mloss workshop, too. Looking forward to seeing you!
As much as I like my new HTC Android phone, contacts sync is just broken. Impossible to have consistent edits.... #fail
How a bit of <scalaClassName>scala.tools.nsc.CompileClient</scalaClassName> can cut compilation time from 30 to 7 seconds... #maven #scala
Preparing standard solutions for our machine learning lecture. Actually writing down math on paper feels oddly archaic ;)
1
After roughly one and a half weeks of back-to-back meetings finally some time back at my desk.
Replying to @ChrisDiehl
I quite liked CSS - The missing manual. No idea if it's still available, though.
Seems like I lost about 20% of my subscribers in the latest migration-induced RSS feed glitches on my blog. Sorry... .
New blog post: "What I have learned from Twimpact" on how important fail tolerance is for networked code
1
Just had a nice Lunch with Holger Arndt of @ujmp (Universal Java Matrix Package). Will definitely take a look.
1
Man, this week feels like the Return of the Administrative. Never head to deal with such a mess before... .
1
Got a new Seagate 7200.4, 500GB for my laptop. About 40MB/s write, 60 MB/s read with bonnie :D Win7 required 4 restarts to feel safe again.
Linked jblas against the multi-threaded version of ATLAS for fun. Got 32 GFLOPS (single prec) on a 12-core machine ;)
1
Replying to @DRMacIver
A friend once showed me a paper on a compile-time interactive debugger for template errors. That feels so wrong ;)
Replying to @DRMacIver
I know exactly what you mean! Always confused by lack of install. And whatever happened to concise error messages?
Trying out eigen. There is something inexplicably satisfying in writing a few lines of C++ and Makefile code and that's the whole project ;)
Reading "Programming Scala" by Wampler & Payne. Really like that it's written for people who already know a bunch of programming languages.
1
BTW, Linux 2.6.33-rc7 was still extremely unstable after suspend/hibernate on my netbook (LG X130). Don't try this at home! ;)
Prototyping new streaming analysis framework for @twimpact by writing down the Ruby code I'd like to have in the end. Quite fun, actually.
1
Migrating my blog to jekyll by hand. This actually makes more fun than it's supposed to do... ;)
Trying to script my Java ML framework from JRuby. Running in all kinds of type erasure oddities... :(
Trying out 2.6.33-rc7 on my netbook to see whether stability after suspend/hibernate has improved... .
We probably should quickly do something about @omgfacts, @idothat2, @ohjustlikeme and friends on twimpact.
Replying to @DRMacIver
That sounds really awful, I hope you'll get better soon! Same thing with my daughter... But she's only three, of course.
RT @DRMacIver: Randomized testing is awesome for finding bugs, but when mixed with concurrency it sure doesn't give you much insight... .
Replying to @headius
I'm really looking forward to jrubyc producing real Java classes for seamless jruby <-> java interop!
Spent the last few days connecting twimpact to twitter's search API. Looks pretty stable so far.
Replying to @DRMacIver
Alright, just pulled your changes. :) I also bumped the version to 1.0 and set it to look for jblas.jar, just check.
Replying to @headius
One question, is polyglot about writing pom.xml in another scripting language, or about using maven dependencies from jruby?
Replying to @DRMacIver
I see. Seems like doing it by hand is the default (and only) way to do it...
Replying to @DRMacIver
jblas-ruby being no exception, doc-wise, of course. I'll hopefully fix that for the 1.0 release.
Replying to @DRMacIver
BTW, still no sign of "pull from fork" button. I'll probably do it tomorrow, by hand, that is... .
Replying to @DRMacIver
Oh, yes, please don't get me started on the general state of documentation on ruby projects ;)
Replying to @DRMacIver
Although I have to admit, I'm not sure whether to view a matrix as [[1,2,3],[4,5,6]], or as its linear list of members...
Replying to @DRMacIver
Yes, I fear it is... . The idea is that to_a gives something array-like and to_ary a real array...
Replying to @DRMacIver
"The data is out of date. We are showing a cached version". I'll probably check in later ;)
Replying to @DRMacIver
Thanks for the link! I'm using it for most of my stuff. Works for me ;) Hm... But how do I pull from your repo?
Replying to @DRMacIver
Yeah, sorry about that. Although I wonder how you got that message. Shouldn't happen through high-level routines...
Replying to @DRMacIver
Yes, jblas-ruby should feel sufficiently rubyesque... I guess I can add the "*i" -> "*!" with a bit of code...
Replying to @DRMacIver
BTW if you have any comments/suggestions/bugs concerning jblas or jblas-ruby, just let me know.
Replying to @DRMacIver
I also wanted to improve some parts of the design. Basically, double and float matrices don't work together as I'd like them to.
Replying to @DRMacIver
That's definitely on my todo list! It just took a bit longer to get jblas to work on all those OSes... . I'll keep you posted.
Trying to cross compile ATLAS from Linux to Win64 using mingw-w64. This better work soon :( I probably should charge for Win64 support ;)
When I asked my colleagues whether I can finish the Win7 install in 45 minutes, they laughed at me :( Next try tomorrow ;)
BTW, I should stress that we're only crawling what's public. It's only that people sometimes retweet stuff from private accounts... .
We just got our third request to remove stuff from our cache at @twimpact. Probably means we're becoming more known... ;)
Adding SSE2/SSE3 autodetection to jblas JNI code... Learning more about gcc assembler and position independent code than I ever wanted
Replying to @headius
I know it's pretty ugly. The real problem is activesupport/dependencies expecting only ".rb" files. Liked jrubyc, nice and clean!
Allright, found out that rails' autoloading assumes ".rb" files, instead of ".class". I'd rather stop before patching activesupport :(