Problem: Deep neural networks often suffer from overconfidence and slow computation. My thesis focused on 1) making networks more efficient and 2) perform fast Bayesian inference on these networks

Methodology: A segmentation network was built using inverted residual blocks. A novel Bayesian inference technique is proposed using stochastic depths.

Technical Detail: The network (built in Pytorch) is designed to be scalable in terms of parameters, ranging anywhere from 0.5M to 10M parameters.

Results: 74% mIOU on CamVid test set, with only 10M parameters. It outperformed many SOTA methods in terms of parameter-performance trade-off. The proposed Bayesian inference with stochastic depth also achieves higher uncertainty calibration compared to dropout-based Bayesian methods.