At long last, I built a new site for myself. I'd like to actually blog more, so this is a lot more focused on actual text than the last one all previous sites of mine.

This site is totally static, it's just html pages, no database, no backend admin system. It gets built using Pelican, a static site generator like Jekyl and Hyde. I tried Hyde, but the current documentation is a mess, and Pelican seems to do everything I want. Except compile LessCSS automatically, like Hyde does. Maybe I'll add that.

Yesterday I launched hearhere.org, basically a mashup between youtube and eventful. It allows you to select venues they like (with a sweet sorting animation thanks to quicksand) and then creates an audio/video playlist of the shows coming up at those venues. The playlist will continue playing, event by event, automatically, so it can easily be used in the background.

Each playlist has a unique short-url so you can share your playlist with others (currently, even playlists that are identical have different urls, which is silly, but I opted for speed here, and can fix this later).

Another mode of operation is text search, where you simply enter the search terms you'd like, and hearhere gives you a playlist of them, which is useful for festivals and other lineups.

I wrote hearhere in python, and it runs on Google App Engine, which I've enjoyed working with so far, though I hope the pricing changes don't make it too costly to run it.

Incomplete features I intend to add include the ability to submit and vote on videos, so the videos displayed would be better/more relevant, comments, and a closer tie in with the eventful api to allow users to declare they're attending events. 

For my machine learning class I created a neural network and a visualization to show how it learns.

What is a neural network? Basically, it's a computer model inspired by the way your brain works. The simplest version only has inputs and outputs. The interesting part is the synapses between the inputs and outputs, which start with small random weights (think multipliers), and get adjusted in small increments over many "epochs" in order to cause the network's output to approach the outputs desired.

Where it gets really interesting is when you start adding more layers, besides inputs and outputs, called hidden layers. This allows the neural net to, in a sense, hold abstractions, and generalize from the training data. 

For this project, I took data from the last 70 years, and built a network with four layers, six inputs, four neurons in each of the two hidden layers, and one output. (Specifically, I built a feedforward back-propagating supervised network, which is what I'm discussing here. There are many other types of neural nets.) 

As inputs, I used the number of Democrats and Republicans in the Senate and House, the active strength of the US Military, and the unemployment rate for the previous year. As output, I used the unemployment rate for that year, as that's what I wanted to predict.

The video above shows the neural net learning, by adjusting the synapses between the neurons, which are animated in gray, and then running through the training data, and then the testing data (i.e. data that it wasn't trained on, in this case, the most recent 15 years).

In the end, my neural network was much more accurate than I expected, the error on the testing data was 11%, meaning that, in theory, if you give the network inputs for a given year (House and Senate makeup, military, last year's unemp) it will predict the unemployment rate within 11% . So if the true rate was, say, 8.3%, the network would give you something between 9.21% and 7.387%. Not too bad. (Frankel and other economist friends, feel free to chime in about the various ways this is wrong and makes no sense.) 

More details:
To do this, I used the excellent Encog library for java. I like python more than java, so messed with pybrain a lot, but Encog has every feature conceivable. And works with processing, which made visualizing easier.

Before I got to what you see in the visualization, I tried a lot of different, similar neural nets. First, I tried it without the temporal data, that is, without the last year's unemployment. This gave me a 17% error rate on the last 15 year data. Next, I tried it with a random selection of 15 years, and the error rate went down slightly. Finally, I added the extra input of last year's unemployment, and, with a ton of training (much more than in the video), achieved an 11% error rate on the test data. I also messed with a great amount of variation as far as number of neurons in the hidden layers, from 2 to 500, but 4 turned out to be best.

There were two areas in which I had trouble, one was data normalization using Encog, so I ended up normalizing the data myself using a python script I was using to aggregate the data anyway (neural nets like data to be between 0 and 1).

Second, was displaying the actual values moving through the network while it was being tested. I rewrote most of my code in attempt to add this (and learned how to use the iterable object correctly), as I think it would be really useful for understanding the way the network works, but I couldn't get the values for each layer to display correctly. With some time, I'd really like to add that.

Source code available here, though this doesn't really resemble what's in the video. Below is the console output for the most successful run of the network. Because the weights start randomly, each training is different from every other training.

Epoch #5447 Error:1.4999955997114394E-4 Neural Network Results for Training Data: 0.66,0.28,0.534,0.324,0.18,0.73, actual=0.4983663395292939,ideal=0.495 0.66,0.28,0.534,0.324,0.3859,0.495, actual=0.23436250492049088,ideal=0.235 0.57,0.38,0.444,0.418,0.9045,0.235, actual=0.08452120095631022,ideal=0.095 0.57,0.38,0.444,0.418,1.1451,0.095, actual=0.0815033122889647,ideal=0.06 0.57,0.38,0.486,0.38,1.2056,0.06, actual=0.08130640799774146,ideal=0.095 0.57,0.38,0.486,0.38,0.3025,0.095, actual=0.24327423747050617,ideal=0.195 0.45,0.51,0.376,0.492,0.1582,0.195, actual=0.20086019727739185,ideal=0.195 0.45,0.51,0.376,0.492,0.1445,0.195, actual=0.2008001851266592,ideal=0.19 0.54,0.42,0.526,0.342,0.1613,0.19, actual=0.27778673711288804,ideal=0.295 0.54,0.42,0.526,0.342,0.1459,0.295, actual=0.28814359093564434,ideal=0.265 0.48,0.47,0.468,0.398,0.325,0.265, actual=0.16153750109068066,ideal=0.165 0.48,0.47,0.468,0.398,0.3635,0.165, actual=0.13558489288725128,ideal=0.15 0.46,0.48,0.426,0.442,0.3555,0.15, actual=0.17602786013114474,ideal=0.145 0.46,0.48,0.426,0.442,0.3303,0.145, actual=0.2199931108910821,ideal=0.275 0.48,0.47,0.464,0.406,0.2935,0.275, actual=0.23021456016162994,ideal=0.22 0.48,0.47,0.464,0.406,0.2807,0.22, actual=0.253903192793502,ideal=0.205 0.49,0.47,0.468,0.402,0.2795,0.205, actual=0.2555298780817681,ideal=0.215 0.49,0.47,0.468,0.402,0.2599,0.215, actual=0.26123536179631845,ideal=0.34 0.64,0.34,0.566,0.306,0.2504,0.34, actual=0.27390713255535215,ideal=0.275 0.64,0.34,0.566,0.306,0.2476,0.275, actual=0.29709252965041594,ideal=0.275 0.64,0.36,0.524,0.35,0.2483,0.275, actual=0.2855653767340758,ideal=0.335 0.64,0.36,0.524,0.35,0.2808,0.335, actual=0.24978578065111492,ideal=0.275 0.67,0.33,0.516,0.352,0.27,0.275, actual=0.27536312822913894,ideal=0.285 0.67,0.33,0.516,0.352,0.2687,0.285, actual=0.2765217655180545,ideal=0.26 0.68,0.32,0.59,0.28,0.2656,0.26, actual=0.2554208345203682,ideal=0.225 0.68,0.32,0.59,0.28,0.3093,0.225, actual=0.1939832730643538,ideal=0.19 0.64,0.36,0.496,0.374,0.3375,0.19, actual=0.19772126359928496,ideal=0.19 0.64,0.36,0.496,0.374,0.3547,0.19, actual=0.16332311108949626,ideal=0.18 0.58,0.42,0.486,0.384,0.346,0.18, actual=0.16442127838161877,ideal=0.175 0.58,0.42,0.486,0.384,0.3066,0.175, actual=0.24245546463834086,ideal=0.245 0.54,0.44,0.51,0.36,0.2714,0.245, actual=0.2558851329275245,ideal=0.295 0.54,0.44,0.51,0.36,0.2324,0.295, actual=0.2954853700500669,ideal=0.28 0.56,0.42,0.484,0.384,0.2253,0.28, actual=0.2681779921075669,ideal=0.245 0.56,0.42,0.484,0.384,0.2163,0.245, actual=0.26151565006675204,ideal=0.28 0.61,0.37,0.582,0.288,0.2129,0.28, actual=0.3470994484698628,ideal=0.425 0.61,0.37,0.582,0.288,0.2081,0.425, actual=0.37127167300622715,ideal=0.385 0.61,0.38,0.584,0.286,0.2075,0.385, actual=0.3441559331912948,ideal=0.355 0.61,0.38,0.584,0.286,0.2062,0.355, actual=0.34953890305512214,ideal=0.305 0.58,0.41,0.554,0.316,0.2031,0.305, actual=0.33600392793467077,ideal=0.29 0.58,0.41,0.554,0.316,0.2063,0.29, actual=0.3332303554084988,ideal=0.355 0.46,0.53,0.484,0.384,0.2101,0.355, actual=0.38391068153643754,ideal=0.38 0.46,0.53,0.484,0.384,0.213,0.38, actual=0.4706693478188674,ideal=0.485 0.46,0.54,0.538,0.332,0.2163,0.485, actual=0.4305970376393928,ideal=0.48 0.46,0.54,0.538,0.332,0.2184,0.48, actual=0.42942131121472155,ideal=0.375 0.47,0.53,0.506,0.364,0.2207,0.375, actual=0.3791308632978942,ideal=0.36 0.47,0.53,0.506,0.364,0.2233,0.36, actual=0.3408292536305523,ideal=0.35 0.55,0.45,0.516,0.354,0.2244,0.35, actual=0.3180255945111218,ideal=0.31 0.55,0.45,0.516,0.354,0.2209,0.31, actual=0.3054461292681537,ideal=0.275 0.55,0.45,0.52,0.35,0.2203,0.275, actual=0.30308948074905867,ideal=0.265 0.55,0.45,0.52,0.35,0.2144,0.265, actual=0.3008652986844994,ideal=0.28 0.56,0.44,0.534,0.334,0.2077,0.28, actual=0.3167321964928064,ideal=0.34 0.56,0.44,0.534,0.334,0.188,0.34, actual=0.33105436361549073,ideal=0.375 0.57,0.43,0.516,0.352,0.1775,0.375, actual=0.35928649461540163,ideal=0.345 0.57,0.43,0.516,0.352,0.1678,0.345, actual=0.3048710453119955,ideal=0.305 0.48,0.52,0.408,0.46,0.1583,0.305, actual=0.2624632144627692,ideal=0.28 ------------------------------- ------------------------------- Neural Network Results for NEW DATA: 0.48,0.52,0.408,0.46,0.1538,0.28, actual=0.23930077726065127, ideal=0.27, percentage error=0.11370082496055091 0.45,0.55,0.414,0.452,0.1504,0.27, actual=0.24239373755821966, ideal=0.245, percentage error=0.010637805884817685 0.45,0.55,0.414,0.452,0.147,0.245, actual=0.22977833679284043, ideal=0.225, percentage error=0.021237052412624118 0.45,0.55,0.422,0.446,0.1451,0.225, actual=0.22768245012752356, ideal=0.21, percentage error=0.08420214346439793 0.45,0.55,0.422,0.446,0.1449,0.21, actual=0.2253461254045889, ideal=0.2, percentage error=0.12673062702294444 0.5,0.5,0.424,0.442,0.1451,0.2, actual=0.21840667496538055, ideal=0.235, percentage error=0.07060989376433804 0.5,0.5,0.424,0.442,0.1478,0.235, actual=0.22297073742023873, ideal=0.29, percentage error=0.23113538820607327 0.48,0.51,0.41,0.458,0.15,0.29, actual=0.24921051256244264, ideal=0.3, percentage error=0.16929829145852449 0.48,0.51,0.41,0.458,0.1494,0.3, actual=0.25960230146033725, ideal=0.275, percentage error=0.055991631053319176 0.44,0.55,0.404,0.462,0.1455,0.275, actual=0.24845814093718874, ideal=0.255, percentage error=0.025654349265926538 0.44,0.55,0.404,0.462,0.1456,0.255, actual=0.23292273464134336, ideal=0.23, percentage error=0.012707541918884128 0.49,0.49,0.472,0.398,0.1451,0.23, actual=0.24803874970006634, ideal=0.23, percentage error=0.07842934652202753 0.49,0.49,0.472,0.398,0.1474,0.23, actual=0.24857095999001486, ideal=0.29, percentage error=0.14285875865512113 0.58,0.4,0.514,0.356,0.1493,0.29, actual=0.26828403415745933, ideal=0.465, percentage error=0.4230450878334208 0.58,0.4,0.514,0.356,0.1506,0.465, actual=0.5537600529394734, ideal=0.48, percentage error=0.15366677695723627 Average Error for Test data 0.11466037 Hey there. End of program.

For my video sculpture and computational cameras class projects, Eric Hagan and I created an interactive piece using a flashlight we modified to output only IR light, a simple webcam we modified to only see IR, a projector, and a lot of processing code.

The idea came out of a video sculpture I created that allowed the audience to illuminate a section of the climatic scene from the film 'The Shining' using an old flashlight. We decided to use the a similar technique, but enhance the technical aspects, film our own footage, add layers, and add something for the audience to search for, using the beam of the flashlight.

We raided the photography department a few floors up for developed, exposed film. Unfortunately, I'm not sure what worked best, as we were raiding the trash, and none of the film remnants were labeled. I suspect it had to do with ISO and developing process, but a couple pieces we found worked perfectly, letting almost no visible light through, and almost all of the IR. 

On a nice spring day, we filmed our footage in Washington Square Park. We masked out a couple of elements, and brought it into Processing. Unlike my previous project, our IR worked so well, we were able to put the camera with the projector, facing the surface we were projecting on, which made for a very natural interaction. After experimenting with the built in blend functions for masking/creating the spotlight affect, we wrote our own, which worked very well and was surprisingly fast.

We ran into some problems trying to add our extra element, both technically and content-wise. For content, we eventually decided to have it be a heart, somewhat to break away from the 'scary video sculpture' aesthetic, and somewhat because Waldo didn't seem worth the effort. Technically, we discovered a serious bug in Processing's copy function, trying to copy a frame of an animated, transparent gif, to a frame of video (not to the screen, mind you, that would have been easy, but we had more processing of the video to do). The copy() function in processing is supposed to respect transparency, but it turns out there is a know bug that it does not. After trying a lot of hacky work arounds, we choose to simply use the blend() function, which worked, but caused a slight flicker.

But it worked. And worked well. I'd sort of like to revisit the project, and go back to building something scary. A lot of people wanted us to put a camera on the other side of the wall and do x-ray vision, and that would be fun too. 

Source code here, though for size reasons, the source video is not included. It should be easy enough to drop in your own, and add some static masks. 

(More about my original comperas final project here.)

For my Design Frontiers in Biology and Materiality class I experimented with spherification of edible liquids (and one pureed delicious solid). Given that you're on the internet, and can't taste my spheres, the next best thing is to check out the photos I took.

Spherification uses a reaction between sodium alginate and calcium to create a thin skin around the sphere. The liquid you want to create spheres out of is mixed with the alginate, and then dropped into a bath of water and calcium (I used calcium chloride). Reverse spherification uses those same chemicals, but, you guessed it, in reverse. This is necessary for liquids that contain calcium already, as they would thicken too much if mixed with the alginate.

I bought all my chemicals from lepicerie.com. If you're going to try this, definitely be sure to purchase food grade chemicals.

Liquids I made into spheres: Water, red tea, apricot nectar, apple cider vinegar, wine (pinot noire), Sriracha hot sauce, and pureed bacon. I reverse-sphered: coffee , balsamic vinaigrette, and milk. 

I used a ratio of about 1 gram of alginate per 200ml of liquid, which in retrospect was a little much. Wine, for instance, thickened a bit more than I'd have liked, so while the spheres were tasty, biting into them was a bit more like eating an alcoholic gummy bear than a caviar. The tea and apricot didn't thicken as much, and had the desired effect. Bacon failed miserably and looked disgusting. Sriracha was so thick to being with that I diluted it some, but it was still to thick to form spheres in the air, dropped out of a syringe into the calcium bath: hence Sriracha spaghetti.

After dropping them in the bath, I waited about one minute, before removing them, and rinsing them in a water bath. This last step is important, because calcium chloride does not taste good. In the future, I may experiment to take them out sooner in some cases, to get a thinner shell

For the reverse-spherification I used calcium lactate gluconate instead of the calcium chloride, with the same ratio. It worked very poorly, as you can see here. The problem with the reverse is that, instead of a thin shell, you get a very thick, clear shell. This could be avoided with an decreased alginate in the bath, but that would probably necessitate a thickener in the spheres, such as xanthan gum.

A debate seems to exist about whether or not it's necessary to blend the alginate with your liquids, or just gently mix it to avoid air bubbles, as well as how long to wait do degas your liquids. I come down firmly on the side of an immersion blender, and learned that waiting to degas isn't that important, at least in my experience. Spheres with air trapped in them looked more visually interesting than the ones I degassed by leaving them in the refrigerator overnight, and had no disernable difference in texture. 

For the final project for Computational Cameras, I decided to  expand on my most recent work for that compcameras: an always-on video recorder with a circular buffer, in essence, a program that lets you decide to record events to a video, after they have occurred.  

Will and I wrote that, the source of which is available here. We'd intended to get it running on an Android phone, but didn't get it working in time for class, so we only had a proof of concept running on a laptop.

For the final, I decided to tackle this problem again. TL;DR: It's hard. Mobile development is really hard, if you're trying to do anything sophisticated involving the sensors on the device. 

 I checked out Appcelerator Titanium, which is a wrapper for Android and iPhone, but it doesn't give you low enough level access to the camera or video frames.I tried Processing for Android, but the video libraries are totally underdeveloped.  I talked to Dan Shiffman, and he said that stuff would be fixed up in the nearish future. I messed with Katai briefly as well, but had the same issues. 

So I was left with writing either Java for Android or Objective-C for iPhone. I'd never even looked at Objective-C, and am somewhat comfortable in Java, but have no real love of it. So I messed with Java first, got disheartened when it wasn't as easy as I had hoped, and tried switching to Obj -C. Which was, perhaps, not the best idea, as Obj-C is not an easy thing to pick up in a couple of weeks. 

In the end, I gave up. I combined comperas with my video sculpture final, and created a more whimsical, artsy project than I have previously. But I learned a lot about mobile development, and absolutely plan to pursue it this summer. Over the next month, I plan to learn as much Obj-C as possible, before starting a mobile project that has we just received VC funding for. Cocoa or bust! 

1. Netflix. An obvious one, but one of the most successful in terms of actually getting me to take it's suggestions. For instance, this week, despite years of urging from all of my friends', I finally caved in to watching Battlestar Galactica when Netflix recommended it when I couldn't decide what to watch. As the media landscape changes, and we move away from the idea of channels, or the channels become to numerous to use as a shorthand, recommendations will become more and more important.

2. iTunes Genius. I don't use this one all that much, but it is handy, as my music library grows to be over a month of music, the idea of the computer generating a smart playlist gains appeal. It's far from perfect, and often fails to create a playlist based on the song I pick if the original song is at all obscure, which sort of defeats the purpose. But it's getting better. I found this article about how it works. . Evidently they use term frequency-inverse document frequency which seems pretty cool. 

3. Ads, Adwords, Facebook Ads. I'm assuming the major ad platforms use a recommendation engine to tailor their ad placement to specific customers, for specific ad campaigns. These could be...better. I suspect they approach the problem from the standpoint of recommending customers to ads, rather than the opposite, but I'm not sure this matters.

4. The New York Times emailed me, and told me they that I would be allowed to keep accessing the Times website for free for a year, courtesy of the Times and Lincoln. What? I'm not sure what data they have on me that caused them to offer me this, but whatever data/algorithms they used screwed up. I don't have any money or clout, or whatever would make it rational for them to give me free access.

In general, these algorithms fail to provide me with useful data. I'm much more likely to trust the recommendation of a friend, and in most cases, will simply ignore computer recommendations. Netflix is a notable exception.