Leon's MSc Thesis: Efficient Bayesian Neural Networks for Outdoor Semantic Scene Understanding Tasks in Robotics

The full thesis can be viewed here.

Problem

Deep neural networks often suffer from overconfidence and slow computation. My thesis focused on two aspects:

1. making networks more efficient (i.e. faster inference speed)
2. perform Bayesian inference on these networks

Dataset

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.

Methodology

The segmentation network is built using inverted residual blocks, with symmetrical encoder-decoder.

A novel Bayesian inference technique is proposed using stochastic depths.

Bayesian forward pass

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:

$$ p(y=c|x, \mathcal{D}) \approx \frac{1}{D}\sum^{T}_{i=1}p(y=c|x, w_t) $$

where $w_t$ represents an instance of a realised weight tensor, and $p(y=c|x, \mathcal{D})$ is the softmax probability for class $c$ outputted by the network with stochastic regularization.

Results

Network architecture variations

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

Comparison to state-of-the-art

Uncertainties

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"
}