Journal of Neural Engineering, 20(5):056041, 2023
Compares recurrent networks, Transformers and a hybrid of the two for predicting the firing rates of recorded neurons forward in time and of neurons the electrodes never reached. Plain recurrent networks outperform both Transformers and the models built specifically for neuroscience, and a Transformer encoder adds further gains only when placed after the recurrent layers. Ranked first on 2 of 4 Neural Latents Benchmark datasets.
ACM Transactions on Sensor Networks, 18(1), 2021
A video object classification system for embedded devices that trades accuracy against latency at runtime, adapting as video content changes and as other processes contend for the device. I worked on this as a summer research intern, training the local executor offline on AWS EC2 and evaluating the model on an NVIDIA Jetson TX2 board.