projects

ALLNet

hybrid cnn for leukemia classification · 92% acc

published in ieee bibm 2021 (4 citations). trained a novel hybrid CNN architecture to identify acute lymphocytic leukemia (the most common childhood cancer) from white blood cell images; beat out contemporary CNN architectures (resnet, inception, vgg) and achieved 92% test accuracy.

identifying Parkinson's from a paient's voice

detecting parkinson's from audio features · 90% acc

ftc publication paper

published in future technologies conference. trained gradient boosting model to identify parkinson's from features calculated from patient audio data, while using mRMR to prune features (90% accuracy).

cnn for tracking eye movements

single-eye gaze estimation · 1.4–2.3 cm error

paper code

trained a squeeznet model on millions of images with aws ec2 to predict user eye movement from a single eye image (increasing efficiency and robustness in practical applications); single-eye tracking with 1.4 cm (two eyes) and 2.3 cm (single eye), achieving comparable results to mit/google benchmarks (12 citations).