Working at Carnegie Mellon Image Science Labs has been one of the most intriguing experiences of my life. The lab's culture stood out significantly from the labs I previously worked at in Davis, primarily due to the scale. Most labs at Davis tend to be relatively large, and my interactions typically involved working with a graduate student rather than engaging directly with the professor. At CMU, I had the privilege of meeting with Professor Sankaranarayanan once a week, and if I finished my weekly tasks early, I could arrange additional meetings to review my work. This level of direct engagement with a professor was a unique and positive aspect of my experience, something I hadn't encountered in my previous lab environments, where scheduling conflicts and missed meetings were not uncommon due to professors' busy schedules.
My summer project at CMU revolved around hyperspectral images. Unlike typical images that have three color channels (red, blue, and green), hyperspectral images comprise numerous matrices representing very narrow wavelengths of colors. In the dataset I worked with, there were 880 bands. A "spectra" refers to the list of values across all these matrices for a single pixel location.
In an image, you often find three primary spectra that represent the main materials in it. More complex substances can have more than three, but, as a general rule, starting with three spectra is a good way to approximate the number of primary spectra. My task was to use an autoencoder to identify these three primary spectra.
Autoencoders are a machine learning technique used to encode an image into a compressed form in the latent space and then decode it back using machine learning as well. This latent space is not human-readable. However, the autoencoder I employed worked a bit differently. While decompressing the image, it used linear algebra and hyperspectral image properties rather than machine learning. This unique approach allowed the compressed image to contain useful information. In my case, the autoencoder predicted the primary spectra of the hyperspectral image, shedding light on the materials within the image and the reactions that might have occurred when it was taken. This could be valuable for chemists, biologists, and climate scientists.
After completing this work, the next crucial step was to submit it to conferences in order to receive credit for it. I submitted it to conferences at Rice University and Purdue University, where I could deliver poster presentations. Unfortunately, these conferences were held virtually. However, at the Purdue conference, I had the opportunity to meet Professor Beecham, who expressed interest in having me work in his lab in the upcoming year—an offer I accepted.
In addition to the poster conferences, I submitted my work to more traditional conferences where they publish your abstract and related material. Initially, I received rejections from both. However, I also applied to the Student Undergraduate Consortium for the AAAI conference, which is generally considered more competitive than the standard student abstract submissions. To my surprise, I was rejected from the student abstract track but accepted to the undergraduate consortium. This means that my abstract will be published in the conference proceedings.
You can find my manuscript on my website, and soon I'll have a three-minute lightning talk uploaded, as AAAI suggested that it would be beneficial for explaining my research.May you be ever victorious in your endeavors.