Skip to main content

Spring is springing

2011-03-19austinchickensfamilygardenhousejugglingresearchsan-antonioweather

Now that spring has arrived, the days have been passing in a flurry of activity both inside and out.

Spring

It’s hard to imagine now, but just a few weeks ago Austin was covered in an inch of snow and ice, and the poor Meyer lemon we planted outside appeared to have frozen solid, dropping all its leaves. In the intervening days, spring has gone from a promise to a premise to a full-on phenomenon—leaves have emerged from tree limbs, grass is reaching skyward, and the bees are out searching for nectar. At first, spring started in tiny places : a single branch on the side of an otherwise dormant trail, seedlings emerging from soil, a tiny leaf sprouting from the ground, green near brown. But in the past ten days or so, these tiny bits of new growth have pushed up and out, bringing with them daffodils, sage, and even tiny new Meyer lemon branches ! The weather in the past two days, even, has gone from the seventies to the eighties. All too soon we’ll be in the middle of summer.

Juggling

Several Saturdays ago, L and I went to the Texas Jugglefest 2011 public show. We weren’t really sure what to expect other than lots of bean bags, pins, and rings being thrown in the air. But the show was way more than that, and we both left feeling energized and optimistic. Highlights from the show were:

Juggling is a very strange activity. First you learn the basics, and then you kind of have a ladder of next-trick-to-learn. (My own progression stopped at passing three clubs, so I am shooting for four balls or four clubs.) An entire show of juggling acts sort of starts looking like a checklist after not too long : ok, now they’re going to do the rings, now four rings, now they do the trick where they stack the rings on their head, now they do five balls, now five clubs, etc. But I have to say that despite the repetition in the acts, there is a lot of training behind the scenes, and performing in front of an audience is impressive, no matter who you are. Combined with the stage magic between acts, and thanks to the kids in the audience, it was definitely a show worth attending. I’m glad that someone on the neighborhood mailing list sent out an announcement about it, because these shows are one of the things that makes living in Austin so nice.

Garden update

The Sunday after the juggling show dawned gray. When I finally brought myself to open my eyes and look out the window, a spitting drizzle was knocking against the glass. I lugged myself out of bed far later than I should have, and moped around the house for a minute. I looked outside again and noticed that the trash bin had spent the weekend on its side in the alley behind the house, so I put on my shoes and ventured outside to pick it up. Happily, the air outside was refreshing from the rain, and the resulting humidity radiated the warmth that was, in fact, peeking in through the clouds. It was a beautiful day, and just walking around outside a little brought me out of a listless funk.

We ended up making the most of the day. L went off to the studio, and I finally, finally finished a project in the garage that had been on my todo list for months. Then, with perhaps a little hubris, but mostly because of the gorgeous weather (the afternoon sun had burned off the clouds and brought warmth and light to the backyard), we undertook to dig out the second raised bed for the garden. Out came the mattock, shovel, and hoe, and we spent the next couple of hours moving a relatively small (tiny ?) amount of dirt a small distance across the yard. It was a completely exhausting effort. The mattock is (by design) heavy, and lifting it up and slamming it into the sod is, by turns, satisfying and overwhelming (not to mention blistering on the inside of the thumbs). L discovered that the hoe is a great tool for a follow-up, to loosen the chunks of sod and move them out of their stubborn hiding places. Then a tour with the shovel moves the dirt out of the ground and into a pile on the side of the lawn. Then, rest.

The week after we had the beds laid out, we ordered bulk garden soil from the Natural Gardener in Austin, and now we have three lovely raised beds in the backyard, where before there was only grass and scorched weeds ! We went to the Sunshine Community Gardens plant sale a couple of weeks ago and bought six tomato seedlings, a couple lavenders, three bell peppers, basil, oregano, and fennel, bringing our seedling population up to nearly two dozen. After the beds were ready, we waited out what looks to be the last of the cold spells for this year, and last weekend we put the seedlings in the ground ! Now I will fret and worry as the bugs devour the tender plants, but I hope that we’ll get a bunch of tomatoes, cucumbers, squash, beans, and chard out of the garden this year.

Just for the record, we planted these seeds in mid-February :

Around early March we had seedlings from the first four, especially the pumpkin and squash, which push up out of the ground with remarkable force. The pomegranate seeds never did anything, and the basil and mint seem to have just needed more time to produce their teensy sprouts.

Chickens

Two weeks ago L and I went over to Buck Moore Feed & Supply and bought three tiny chickens ! We have kept them in a box for the intervening days, during which time they have gone from little three-inch puffballs to six-inch gangly teenager-looking things. They still chirp like chicks, singing themselves to sleep every evening, but soon enough they will be big, probably ornery, chickens, hopefully eating the bugs from the garden plants and giving us a few eggs now and then.

Family visit

Last weekend my dad and stepmom came to visit, so we all took the occasion to drive the hour and a half down to San Antonio. The renowned riverwalk had piqued our curiosity, so we wanted to see what it was like. It turns out that we were there the weekend before Saint Patty’s Day, so the river had been dyed green. Combined with the narrow channel, the low banks, and the crowds of people ambling about, the entire scene looked something like Disneyland-meets-Vegas. It was surreal, but I really enjoyed the walk. I love urban landscapes that are actually used by people, and San Antonio has definitely got the use.

Despite being more like Disneyland than I had expected, the riverwalk was not too far from my imagination. What I was not expecting, though, was that San Antonio is a very old city, by American standards, and so there were loads of other interesting things in town. The downtown area surrounding the riverwalk is old and in various states of renovation and disrepair, and the crowds from the more touristy blocks spill out into the surrounding streets and make the city feel very European. The Alamo is in San Antonio, and the highlight there (for me) was the beautiful grounds surrounding the “Shrine of Texas Liberty” – even including a possum hanging out on one of the yucca plants. The best, though, were the ruins of four original Spanish missions in the southern part of town, generally built in the mid-1700s and mostly in ruins now. Surprisingly, though, the usable parts of the buildings are still maintained by some part of the Catholic church, and set aside as National Park land in some capacity, so they are all beautiful working areas that welcome both tourists and worshipers.

My dad is a fountain of know-how around the house, and my stepmom loves to garden, so we spent a couple of the days of their visit just hanging around at home doing a bunch of little projects. We labeled the circuits on the breaker outside, fixed a hanging lamp, installed an exterior outlet, weeded two huge flower beds, and installed drip irrigation in the raised beds. The house is in significantly better shape now than it was a month ago !

Research

It’s been a little slow going on the research front these couple of weeks, but things have been moving along in fits and starts. I’ve been coordinating with B to get the lane driving simulation to a place where it consistently learns useful policies. Over the past week I ran a number of small simulations in the walter world, where the learning agent must aim for targets and avoid obstacles. The world is normally static, but I added a new feature that allows the obstacles and targets to move around randomly at a fixed speed. By measuring the distance to the nearest target during evaluation periods, I found that the learner can develop a working policy even when the targets are moving at nearly the maximum available speed. Also, interestingly, the learner failed to develop a working policy for some speed conditions, but not all of them. In addition, when the learner did develop a working policy (so that it was able to remain near the targets throughout the evaluation periods), the distance-to-target metric went through several discrete plateaus of similar performance, divided by sudden change in policy – quite reminiscent of punctuated equilibrium. My take on the simulations is that learning is highly dependent on exposure to critical parts of the state-action space, and so unless the learner gets this critical expsure, there is some sort of ceiling on overall performance. I’m sure there’s a better way to phrase this, but that seems to be the behavior so far.

At any rate, the plan for the next week or so in the driving world is to separate the tasks in the lane driving world by controller, so that we can create a pedal controller that depends only on distance to target, and a steering controller that depends only on the lane occupancy. I’ve also included skeletal code for perceptual arbitration, which is exciting because that’s sort of the entire point of the reinforcement learning simulation for this project.

On my own work, I’ve been pretty stalled for the past couple of weeks. My work machine has been training up a Smith/Lewicki codebook for several days now (!) but seems unable to get below about 33 % reconstruction error on the development sound when the codebook vectors are allowed to shrink and grow. (A different Smith/Lewicki codebook that I trained earlier gets down to 23 % error with constant-length vectors.) I have, however, coded up a Python translation of the Taylor/Hinton/Roweis deep belief network that they used to model human motion capture data ! I plan to sandwich this network between the motor and perceptual representations in my learner to produce a full model of vocal mapping. This week I will be testing and verifying the deep belief nets code, and training up a matching pursuit (or, probably more likely, a Smith/Lewicki codebook) on the FFT power spectra, to see how well that can model speech production. It remains to be argued how this model can incorporate (a) plant changes like vocal cord length changes during development, and (b) frequency mapping to adapt parent-frequency speech to learner-frequency speech. I think these points are important but can hopefully be ignored in the first pass of such a model.