The Holiday Season & the Coffee Strikes Back

Well kicking the coffee lasted until Dec 25th... turns out my parents drink way too much coffee, and give and receive too much coffee to say no while I am back East. Not sure whether I'll give it a go again this year - I was still getting headaches on day 5 of kicking.

I'm not really a gadget person, but the end of the year brings about best of tech reviews which can be interesting. PC World has a review that lists Microsoft Office 2007 as the #1 most innovative product, but their piece seems very old guard to me - Intel, Sony, and Dell are also on the list. David Pogue's top 10 features of the year piece is more interesting - he includes the Zune's wireless song sharing feature and then sinks it for its 3 day/3 song expiration caveat.

Different sources will give you lists of scientific stories for the year - for instance Scientific American's top science stories and Technology Review's emerging technologies list. My vote for tech breakthrough of the year would be John Bower's work with optical interconnect on silicon. Communications and computation are intertwined and one always limits the other, improved bus speeds will do much to speed computation tasks.

It's interesting whether and when these types of university level discoveries make it through to mass production in the world. Many of the things I've read about, like a new type of fiber optic cable with vanishing attenuation invented at MIT more than 5 years ago, simply seem to disappear. Others, like lasers, make it into our disc readers and popular use without much fanfare.

Technology management programs at universities try to help graduate students take their projects into entrepreneurship, but at the class level I don't see much talk about this type of thing. Maybe, to be realistic, this only happens at Stanford.

Kicking the Caffeine Addiction Day 2

Overwhelmed by all of the Xmas shopping that I've put off until the last minute - I was on my feet driving around Santa Barbara/Ventura counties all day. It's day two of quitting coffee. Yesterday went smoothly enough. However, last night I was miserable - tossing around in bed, flailing my limbs, suffering from nightmares.

Today I had heachaches, dizziness, nausea, pain, felt lightheaded with symptoms intensifying in the afternoon (50 hours no coffee). I had to stop home and lie down and take a couple of aspirin. Some 'chai' tea helped - a small dose of caffeine to tide me over. Starting to feel bad again this evening.

I drink coffee mostly to help me think, read, write. At what price? They say there is no major health risk involved with coffee addiction except for people who have heart troubles. Somehow I imagine in the future they'll discover that coffee takes away your soul. I LOVE IT I LOVE IT I LOVE IT

Break my coffee addiction

Well it's the end of the quarter - wound up my last exam in Information Theory... given that I have a few weeks off, ostensibly, 'tis the time of year to shake my one true vice - addiction to coffee. I once read somewhere, I think it was in a Carl Sagan book, that mathematicians can be thought of as machines that convert caffeine into equations.

Years ago I learned that caffeine helps me think faster and get more done. Also years ago coffee went from something I'd drink when settling done to study, to a ritual morning and oftentimes afternoon beverage. Coffee breaks.

Now I succumb to headaches, crankiness and I even fill like I'm getting sick when I go a day or two without a cup of coffee. But again I'm going to prove to myself that I have control over my habits - and break the addiction, at least until the new academic starts again in January...

Here is my strategy:

I usually drink one 'grande' or medium cup of strong-ass Starbucks, Peets, or Java Jones joe a day... plus a bonus espresso or small coffee in the afternoon if feeling low. One 16 oz. cup a day wouldn't qualify as a lot of caffeine if it weren't for Starbucks & co.

-> drink green tea as a substitute ... often in the first few days, until I cut even that out altogether. (I don't drink sodas or energy drinks so I don't have these to worry about)

--> take aspirin if headaches get bad

---> work out --- try to go biking, running & surfing at least once a day

----> drink lots of water. The idea is that I can pee out the toxins.

Well let's see if I can do it (again I've quit for 2-3 week periods maybe a half dozen times). I've read that drug addiction actually causes a change in the wiring of neurons in the brain, I wonder how fast these can be altered? Or is it once an addict always an addict?

-------------------
Unrelated great news - my stipend check is in! Thank you Uncle Sam, now I can get a few extra weeks of 5% juice off of that Paypal money market account.

Visual Speech

As technology enables us to become creators of 'content' rather than just consumers it may be new modalities of communication emerge - not merely new mediums (IM, VOIP, mobile telephony, blogs, podcasts, webpage based).

A teenager in his bedroom can do a halfway decent job throwing together a music track with the basic appealing elements - rhythm, suspense/resolution, melody ... or script up and shoot clips that do the same in the movie domain, capturing storytelling, plot, character, conflict etc...

The pros just do it better - I'm not saying the masses will replace the pros in the media we intake as a culture - but the collective effort allowing us all to be songwriters and filmmakers, singers and actors, etc... will cause a profound impact on how we interpret what we see or hear.

Some homemade content creators may only be looking to make a buck - eg Ask a Ninja on Youtube, but I think that most are really only creating in order to share the interpretation of their world with their friends, family and colleagues. It is all about communication, and it's sexy to have a collection of songs, mixes and videos to share with the world.

Communication... is it possible then for ever-present unlimited bandwidth to create for us a new modality for communicating - beyond the major ones of writing, speaking and gesticulating? Visual speech, in which we access a collection of photos, pictures, and other symbols, for use in discourse may emerge as an abstract way of representing our moods, desires, needs, and imperatives.

As is we have the variety of 'emoticons' in IM and we also have photos to represent our social page personal profiles (it's us!). What I am envisioning is something beyond even this.... For instance, it may be that in the future we can simultaneously interact with the world and view some sort of information display - and it may be possible to project to any or all people observing us a series of images that represent partially what we wish to communicate.

Sexy pictures of Facebook, pictures of drinking, etc... all slightly subconscious methods of doing this already.

As humans and machines co-evolve will we acquire the visual analog of speech?

Data Mining Projects part Two

ok where did I leave off???....

Student Projects in Data Mining----

#4) The Netflix Million Dollar Prize

A few months ago Netflix announced a million dollar prize would go to the group who could outperform their current movie recommendation software (Cinematch) by 10% in predicting how their customers would rate new movies (they do so on a 1 to 5 star system). Cinematch currently scores at about0.95 RMSE.

The student group tried a number of approaches, the best of which involved finding users who have viewed similar movies and gave similar predictions... they reported below .9 on RMSE on the training set - but when running the probe RMSE shot way up... of course generalization is a huge design issue in these types of problems - but some people in group discussion argue that Netflix has purposefully filled their test data with outliers...

The group in class has not tried using data on the movies themselves, for instance from the internet movie database (IMDB)... I'm sure someone somewhere has...

I'm kind of skeptical about the whole prize anyway - it seems like a gimmick by a company on a badly leaking ship (technology will make their business model obsolete). On the other hand scalability in data mining algorithms is an open research issue and this contest and its large training set (100 million ratings) might contribute something in the end after all...

To date the best score, wxyzconsulting.com, has shown 6% improvement (~.89 RMSE).... I think that puts them in the running for the $50,000 progress prize.

#5) EEG analysis

My project. I'll write more about this at a later date.

#6) Retinal Detachment analysis

By this time I was a little out of it as the caffeine was beginning to wear off after 2 1/2 hours+ of non-stop presentations. The gist of this effort was to develop methods to determine via images of detached retinas (in cats) the state of health of the cells. Cells were segmented, features in the cells extracted using MPEG-7 texture descriptors, labeled, and a Kullback-Leibler divergence distance was applied.

The dataset comes from an intensive area of research at UCSB in bioinformatics.

#7) Blog Mining

A group, building off of experience as interns at Google, undertook the task of mining and associating similar bloggers using Windows LiveSpaces as a dataset. Essentially, they clustered by keywords to create categories with some interesting results -> usernames closely matched the blog topics in some instances, e.g. jesuslover for religious blogs having frequent keywords - faith, hope, love, peace etc...

I see this moving towards companies contacting 'experts' given a certain category and hiring them to help promote items, or just selling directly to them - kind of Google's context analysis driven marketing on steroids.

#8) Delay Testing

Applying statistical analysis to predict device (silicon chip) failure based on input-output delay times for an exhaustive array of input vectors. This project suffered from the fact that the data was generated via an emulator - i.e. the results may be meaningless, but the overall technique seemed applicable. Unless I missed something I don't know why many in the class were so critical of this.

#9) Drunken Recommender Based on Delicious

Petko Bogdanov ... one of those brilliant guys that actually seems to have a personality, or at least a sense of humor - came up with the idea of randomly walking through related Delicious tags to provide link recommmendations. The drunken metaphor seems very Bulgarian/Eastern European to me.

Petko complained about Delicious' API - how his crawler would get booted often for polling too often. Seems to me that Delicious, known for its server issues, also tries to keep some control of its data, maybe moreso since Yahoo bought them... kind of weird considering its supposed to be one of the defining social networks...

in that sense maybe it's parallel to gov't databases - they'll collect your data from everywhere: tax forms, credit card companies, phone companies, healthcare providers but you'll never have access to your own information!

and one day a 37 dimension hyperplane might decide you are a terrorist and you'll have to clear an extra 2 hours of hurdles before boarding a plane. wait - this already happens.

God Save the Outliers.

Data Mining Student Projects

Ok so I'm feeling a little down about the information age - and the realization that many of us will be tied to desks and computers for 40 years of working. Where's the human-to-human interaction, the sense of spirit and risk and adventure?

The Renaissance man is dead.

If I could I'd make a million dollars or so from computers and then never use them again for anything other than a publishing outlet.

So today my graduate seminar course on data mining and pattern recognition concluded, with a 4 hour session of student group project presentations. Thankfully, the topics were not pre-assigned but rather the primary criterion for topic selection was 'non-traditionality' in the data set... So we had presentations ranging from analysis of EEG signals to mining social networks for recommending bookmarks to studying retinal detachment.

Here is an outline of the projects presented, which may in a way be a way of sampling current hot research topics in electrical engineering and computer science... (except wireless networks are not represented):

DATA MINING STUDENT PRESENTATIONS PART ONE

- Atomic Decomposition of Musical signals.... my friend Bob Sturm's dissertation topic.

Speech analysis may be dead. It's been compressed to death, recognized and synthesized to a T, and now you see prominent speech experts like Larry Rabiner attending music signal processing conferences. hmmm

Bob Sturm has one of the most unique approaches to representing waveforms in terms of a non-orthogonal decomposition using an algorithm known as matching pursuit. In this project he looked at ways to cluster collections of atoms (think mother wavelets with different scales, offsets and amplitudes) in meaningful ways. A flute signal's atoms, for instance, may be clustered into notes consisting of atoms that make up the harmonics.

The big question everyone wants to know is: why? and what can this representation be used for really? In one sense the technique is the opposite of computer composer Curtis Roads' granular synthesis method... and so new methods for granular-based 'sampling' for music might fall from this.

I want to mention that Bob gives some of the best presentations I have ever heard other than by really famous experts.

- Image-based retrieval

Another group was interested in improving methods for retrieving relevant images given keyword searches in the Cortina database - can relevance feedback be used to eliminate non-relevant results (ie. if you search for dog, and dog leashes are returned, can user feedback eliminate this?)

I don't think you ever have users willing to do this... you may hire human experts to do this, but the cost to cover the most popular searches seems prohibitive.

I should mention that Cortina uses lower level image features to return its results - like edge, texture and color information, along with meta data like surrounding text, website category, etc... is it a huge open research area how best to combine lowlevel features with higher semantic-level information? I don't think Google Images uses low level features at all, and the results there are usually excellent... - they have so much data to work from!

- Automatic photo annotation for personal photo albums

This is a student's doctoral thesis project. Her technique only applies to outdoor photos - it is possible to do a rough segmentation based on a mean-shift approach to divide images into two main regions, and then train a classifier to automatically label these regions into the categories of sky, snow, waves, grass, trees, etc... The big question is are these particular tags of any use?

People are probably more interested in identities of the people in the photo or adding tags that can't necessarily be garnered from the image itself - ie which place the photo was taken or what occasion it was. Still, automated annotation can never hurt - why not be able to pull up all the best sunset photos in your collection for instance, or all the outdoor ones?

3 down, 6 to go in PART TWO!

Technorati Profile

The Unknown Function of the Spleen

Until recently doctors did not know which purpose the spleen serves. Imagine an entire organ's functioning unknown through the end of the twentieth century! Well I suppose they've figured it out now - removing it predisposes people to certain infections. So it must serve to prevent those infections. Oh, and it removes dead blood cells.

With the overwhelming avalanche of facts, and the death of the Renaissance Man -> only the specialists really understand any one field, it's become impossible to sort out what we do know, what we don't know, which folk wisdoms have some scientific validity, and which previously scientific consensuses are no longer held.

It is not only an information labyrinth we're forced to navigate in modern life - but a knowledge explosion as well.

Even worse - a technologist's current exposition on the state and moving trends behind modern life get dated FAST. I am currently reading Robert Lucky's Silicon Dreams, partially because I am enrolled in an Information Theory class and his book helps me bring the concepts to life. However I have to skip over major portions of the book (its 17 years old) - now that the internet, mobile phones, email, pervasive video, VOIP, mp3 players have revolutionized the way we live and work.

Brain Computer Interface

The year is 2045. What was once called the digital divide has now become the 'innermind divide' - people who can afford it now surgically opt for a brain interface chip in their teenage years. This chip can perform fast, accurate imaging of the brain... Scientists are still at a loss to produce a full model of the brain's functioning, but people with this implant actually learn new neuronal pathways to control and receive information from the interface. It is connected wirelessly to world network, of course, and able to access information about everyone and everything important to you which for the richest of people gets displayed via a nanoscopic retinal implant fused to a portion of the optic nerve.

Of course I am probably the ten millionth person to write about this... but this technology will develop. It is only a matter of time.

One of my more interesting research projects at UC Santa Barbara involves an important aspect of the above vision - pattern recognition of meaningful cognitive events using signals measured from brain-based sensors, in my case an EEG array. Pattern recognition methods continue to advance, now out of the box software packages can run today's modern techniques with all the bells and whistles attached. SVMLite is a great open source package for support vector machines, and I've seen packages that run Brieman's random forests and different Adaboost-style ensembles.

Controlling the stimuli is a challenging problem for BCI in real settings, as eye and muscle movements can destroy signal-to-noise ratio in EEG measurements. Chamber-based imaging systems, PET and fMRI for instance, are impractical for real world BCI. DARPA's augmented cognition program is hyping fNIR, functional near-infrared imaging, as the way to the future... it seems tissue absorbs near infrared light at varying levels depending on the presence of oxygenated hemoglobin. The idea then is: strap a head band with IR LEDs on one side, photoreceptors on the other, and measure light absorption in the skull. Of course the modulation schemes and absorption models are important, and I'm not claiming building a tempo-spatial map of oxygen in the brain from LEDs, digital filters, and communications schemes is a trivial thing - but here is a way to essentially have a mobile fMRI device for a few hundred dollars, maximum, in components and software.

The question is: does it really work? I've been told fMRI has drawbacks too concerning which regions of the brain it can accurately image, and that oxygenated hemoglobin is not exactly analogous to cognition, but the thing I'm most curious about is whether it can do what EEG cannot - create a strong signal for BCI.

Some big people in the AI research world are getting involved in this including Eric Horvitz at Microsoft... maybe their products suck but their research direction doesn't...

My Background

At the time of writing this I am a 2nd year Ph.D. student at the University of California Santa Barbara, studying Electrical and Computer Engineering - specifically signal processing and communications. I am interested in web and search, content analysis of video, pattern recognition, data mining, computer vision, music signal processing, cognition and statistical approaches to neuroscience, and well... maybe I better just stop.

I am lucky to be an NSF IGERT trainee - I receive a stipend from the NSF to devote myself to interdisciplinary research in interactive, digital multimedia. The "digital" seems to me redundant. My field historically addresses issues in compression and channel coding of signals that are not memoryless - speech, audio, video, etc... I am more interested signal analysis for knowledge deduction -> can patterns be recognized in multimedia from low-level and semantic features that lead to more fruitful human intake of such streams?

At times I rail against life as a graduate student - the open-ended freedom causes me to place a lot of unnecessary pressure upon myself. Other days I feel as if a human element is missing somewhere in the algorithm-creation business. I might be better off editing a travel magazine.

But let's take things as they are for the time being ... onwards!

Diving Right In

First, why a blog? and not a livejournal or myspace or even a Facebook?

Livejournal - which has been around the longest seems suitable for travelers or day-to-day diary style entries. It might devolve for that around here, but I'd like to keep these posts focused on goings on in the 'Information Age'

Myspace/Facebook/Friendster/Orkut - are for social networking...

The ability to offer information is marketing.... being perceived as the expert and building a network is how to succeed in today's world. Success, to me, does not mean salary or financial wealth - instead I see it as collected potential to help solve the world's problems or 'access' to the issues of the day.

Will the day come, wanting that success - where I am forced to choose between it and loved ones?