I was recently challenged to fit the entirety of Bohemian Rhapsody into 8Kb of data. I thought it might be impossible, as the original file is 5.7Mb. This would be a compression of 712.5x. However, the challenge said nothing about audio quality, so I still thought it might be possible. To begin, I cut the sample rate to 2k, the lowest any audio format I found supported, changed to opus format, and set compression to 10.
I reduced it to 120kb. That is 475x. The audio was terrible, practically turning Bohemian Rhapsody into death metal. I then decided to try using midi. Using a compressed midi, I reduced Bohemian Rhapsody to exactly 8Kb: 712.5x compression!
My roommate and I have been working to make our dorm room as amazing as possible. We currently have an Alexa device up and working (had to finagle that one with the school's network) and some very nice carpeting. I'm also creating a custom desk and two tables (pictures below). I'll update with a tour of the room once it is finished.
In partnership with Harvard's Center for Astrophysics (CfA), I have begun creating a planet classifier. The plan is to use a deep LRCN network to classify based on light curves. In the past week, I have created a PCA/SVC model to establish a baseline. In theory, the LRCN should do better, but the SVC is already significantly over chance. There were three categories: planet, eclipsing binary (often confused for planets), and junk. The ROC curves and PCA Dimensions vs Accuracy are below. Considering it's a three class problem, a 70% accuracy on validation data is pretty darn good. : )
I made this box about a year ago, but I never uploaded it. I recently rediscovered it upon a long overdue storage cleaning. It's a beautiful enclosure I made for the radio transmitter for my drone. I tuned the settings to emphasize the burn effect on the side of the wood. It created a very nice alternating pattern that I'm pleased with.
This is what I was working on over the Summer!
New Affectiva cloud API helps machines understand emotions in human speech
This is from a SolidWorks class I took last year, but I thought it was worth a share.
You can run some fun simulations in there!
It was obvious all along...
Problem: The case is overheating.
Solution: Remove case.
Want those excellent thermals at high loads? Buy a dalocalhost cardboard box!
So what does the average face look like? This may seem like an odd question, but it is one I had to answer the other day, and the results were surprising enough that I thought I should share them. I took 16,000 images, made them grayscale, and took the mean of each pixel value across all the images. What turned up was utterly terrifying. So it turns out the average face is creepy and possessed.
Meta Joke: Or could it be that everyone is just scared of being average?
I finally upgraded my computer setup. Until now, I have been running everything on a 2013 Macbook Pro. This includes all my machine learning training and intensive programming. Over the past two weeks, I scoured Craigslist and similar websites to find used parts to build a new computer. This computer, going from the cost of new parts alone, would be worth about $2000-2500. I paid under $1000. Its dual GTX-980s (one is not shown below) will allow my models to train as never before. All the parts are functioning, but I still need to figure out how to configure in SLI. I'll keep you posted!