NCSA-NVIDIA AI Hackathon

October 5-6, 2019
3013 ECE building
306 N. Wright Street
Urbana, Illinois

The first NCSA-NVIDIA AI Hackathon of the semester is co-organized by the Gravity Group, Innovative Systems Lab, and NCSA Industry, and co-sponsored by NCSA SPIN and NVIDIA. The main goal of the hackathon is to let talented U of I students, postdocs and staff showcase their skills in a friendly competition while working on challenging problems involving deep learning on a state-of-the-art compute platform designed for AI. This will be an intensive 2-day experience culminating in the final presentation of results on Sunday afternoon. Courtesy of NVIDIA, the winning team will receive two Titan V GPU cards. The second-place team will receive one Titan V GPU card.

To participate, we ask interested students and staff to sign up at https://forms.illinois.edu/sec/5277314 by Monday, September 30 and indicate which of the four projects described below they would like to work on. Students accepted to participate in the event will be notified on Tuesday, October 1.

Hackathon Schedule

Saturday, October 5

8:30am — Teams registration and light breakfast

9:00am — Overview of the Hackathon rules, challenge problems, and a brief intro to HAL computing environment

9:30am — Teams assemble in break-out rooms and start working on the problems

Noon — Lunch (pizza will be provided)

1:00pm — Teams continue to work on the challenge problems (snacks will be provided)

5:00pm — Teams meet in 3013 for a brief status update

Sunday, October 6

8:30am — Teams continue working on the problems (light breakfast will be provided)

Noon — Lunch (pizza will be provided)

1:00pm — Teams continue to work on the challenge problems (snacks will be provided)

4:30pm — Teams present results

Wednesday, October 9

Winning teams are announced and presented with prizes

Hackathon Projects

Project 1: Early Genre Detection

Problem: Train a model to classify songs according to the genre they belong to. The model will be evaluated on both latency (early high confidence prediction) and accuracy. By early prediction it is meant that the model should go through the audio temporally and should make a prediction without going through the whole song. Input to the model should be raw audio.

Music dataset for the problem: https://github.com/mdeff/fma. Helpful lightweight dataset for prototyping: https://github.com/Jakobovski/free-spoken-digit-dataset.

Project 2: Wait a Minute! I Know That Song! (Stacked Models / Conditional Prediction)

Problem: When you are listening to music, often you can tell the genre before the song. Train a separate model that predicts song ID conditioned on the genre prediction from the first model (or conditioned on the ground truth genre). Results will be evaluated on both accuracy and innovative formulations of introducing latent variables to condition a prediction on a prior. Input to the model should be raw audio plus one piece of metadata (i.e. the genre).

Bonus: Demonstrate whether conditioning on genre boosts accuracy (i.e compared to a model only trained on raw audio).

Music dataset for the problem: https://github.com/mdeff/fma. Helpful lightweight dataset for prototyping: https://github.com/Jakobovski/free-spoken-digit-dataset.

Project 3: Human Segmentation

Problem: Train a model to perform amodal segmentation of humans. Results will be evaluated on accuracy of predicted regions that describe visible and occluded human body parts.

Datasets: Data available at sailvos.web.illinois.edu

Project 4: Feature Recognition in Digital Elevation Maps

Problem: Develop and train a model to recognize features in Digital Elevation Maps (DEMs). The provided dataset contains pre-processed DEM tiles with labeled regions of interests. The features of interest are soil erosion regions.

Dataset: Data will be provided at the time of the competition on HAL cluster. Example image can be found at https://wiki.illinois.edu/wiki/display/~kindrtnk/Hackathon+Dataset.