Leon's MSc Thesis: Efficient Bayesian Neural Networks for Outdoor Semantic Scene Understanding Tasks in Robotics
The full thesis can be viewed here.
Deep neural networks often suffer from overconfidence and slow computation. My thesis focused on two aspects:
- making networks more efficient (i.e. faster inference speed)
- perform Bayesian inference on these networks
Camvid dataset: 367
, 101
, and 233
train, val, test images respectively, trained and tested with resolution 480 x 360
. You can download the dataset from here: https://www.kaggle.com/datasets/carlolepelaars/camvid.
The segmentation network is built using inverted residual blocks, with symmetrical encoder-decoder.
A novel Bayesian inference technique is proposed using stochastic depths.
A bayesian neural network wrapper class Bayesian_net
is used to set and toggle the dropout and stochastic depth layers, essentially overwriting the default model.eval()
for certain layers. The bayesian forward pass is computed as follows:
where
A set of network variations were tried better understand the encoder-decoder architecture.
- no-skip: remove skip connections between the encoder and decoder
- upsample-skip: add skip connection everytime the network upsamples
- dense-skip: every encoder is connected to a decoder
- add-skip: use the add operation instead of concatenating
The Bayesian networks trained with stochastic regularization achieves much more calibrated uncertainties.
If you find this work useful and use it in your work, please cite it as follow:
@article{yao2023msc,
title = "Efficient Bayesian Neural Networks for Outdoor Semantic Scene Understanding Tasks in Robotics",
author = "Yao, Linghong",
journal = "leonyao.net",
year = "2023",
month = "Sep",
url = "https://www.leonyao.net/projects/msc"
}