Density: Automatic Measurement of Room Capacity

Roddur Dasgupta *Henry Gu,  Ritika Dendi,  Stephanie Lampotang,  Shantanu Jhaveri
* Project Lead

Technologies that accurately and anonymously track the number of people in buildings are critical for improving the security and productivity of spaces, without violating the privacy of the individuals. Density (an SF-based startup) develops hardware and software that counts people entering and leaving doorways without capturing identifying information about people. Our ML collaboration with Density developing deep models that accurately and robustly detect people in doorways.

Kawasaki Disease: Rare Disease Diagnosis using Machine Learning (3.0)

Hayden Shively *Komal Patri,  Matt Evenson,  Ritika Dendi,  Jessie Zhang,  Washington Zhao
* Project Lead

Kawasaki Disease (KD) is a rare heart condition that causes inflammation and swelling in blood vessels throughout the body. Diagnosing and treating KD at an early stage is critical, because KD results in life-threatening heart problems for 1 in 5 victims, typically children under the age of 5. Our collaboration with UCSD’s Kawasaki Disease Research Center involves developing an image-based classifier for KD.

Gene Sequence Analysis

Isaac Gelman *Tomas Angelini,  Natalie Abreu,  Nicolas Perez,  Shannon Brownlee,  Shreya Havaldar
* Project Lead

Increasing amounts of genetic data are becoming available due to the rise of Next Generation Sequencing (NGS) technologies, which are revolutionizing the study of evolution, pharmaceuticals, and clinical medicine. NGS data is unique because of its size -- millions of reads in each experiment make it slow to analyze as whole dataset. The Sequence Read Archive (SRA), an international open-source biological sequence dataset from the NCBI, has uploads which are under-labeled, which poses a problem for researchers mining the genetic data. Our collaboration with USC’s Smith Computational Genomics Lab group involves applying machine learning, probabilistic models, statistics, optimal stopping theory to analyze sequences, label data, and help researchers advance biomedical science.

Deepfake Detection

Armaan Pishori *Brittany Rollins,  Nicolas van Houten,  Nisha Chatwani,  Omar Uraimov,  Shantanu Jhaveri
* Project Lead

Deepfakes are a rising threat to various social and political aspects of our society, given their potential to spread misinformation in the media. A CAIS++ team is participating in the Deepfake Detection Challenge, along with research groups around the country, to develop deep models that can help detect deepfake videos and manipulated media.

Games Reinforcement Learning

Zane Durante *Gireesh Mahajan,  Oscar Bashaw,  Gloria Chang,  Nathan Huh,  Lily Perry
* Project Lead

A CAIS++ team is exploring and implementing different reinforcement learning (RL) algorithms in the context of games, from basic RL approaches such as Q-learning to more complex deep reinforcement learning approaches.

Music Generation

Ada Toydemir *Brian Plotnik,  Karan Menon,  Priscilla Lee,  Tierra Buissereth
* Project Lead

A CAIS++ team is using ML to create a sample music generation program. When musicians produce music, they often use samples for sounds such as kicks, snares, hi-hats, or even sound effects. The character of a song can be drastically changed by the quality and uniqueness of the sample. Unfortunately, samples are expensive to buy, difficult to record or create, and take up hundreds of gigabytes of space on hard-drives to maintain. We will create a program that takes multiple samples as input and uses them as reference to generate new samples.

The Ladin Language: Preserving Endangered Languages through Phoneme Recognition (3.0)

Zane Durante *,  Leena Mathur *Eric Ye,  Jack Zhao,  Tejas Ramdas
* Project Lead

Approximately ~60% of the world’s current 7,000 languages is predicted to go extinct by the end of the century; the death of any language represents an irreversible loss of information across multiple fields, including linguistics, psychology, sociology, and anthropology. A CAIS++ team is continuing their work on developing unsupervised representation learning techniques for automatically extracting and preserving the phonemes of endangered languages. Their focus is on Ladin, a language spoken by about 20,000 people in the Italian Alps, in a research collaboration with USC’s Department of Linguistics.

FALL 2019

Homelessness Resource Allocation

Homelessness is at a crisis level in Los Angeles, with ~53,000 persons sleeping on the streets or living in emergency shelters every night. The Los Angeles Homeless Services Authority (LAHSA) can only house ~15,000 persons each year. To address this problem, they developed a “paper and pencil” approach to housing prioritization and matching, based on people’s vulnerability scores, past housing, and history of return to homelessness. Can an ML-based approach out-perform the city’s current scoring system? In Fall 2019, a CAIS++ team worked with a de-identified dataset of homeless adults to develop ML approaches that identify predictors of housing success.

Computer Vision for Self-Driving Cars

Self-driving technologies have the potential to revolutionize human transportation by improving efficiency and safety. In Fall 2019, a CAIS++ team worked with one of the largest open-source datasets provided by Waymo to develop deep networks for object detection (vehicles, pedestrians, cyclists, signs) and weather condition classification, based on camera and LIDAR information sensed by self-driving cars.

Evolved FPGAs for Machine Learning

Can FPGAs function as universal function approximators, like neural networks? A CAIS++ group is researching and implementing evolutionary algorithms on the programmable architecture of FPGAs -- the goal is to exploit the hardware to develop networks that solve ML problems faster than neural networks.