Extrapolative Continuous-time Bayesian Neural Network for Fast Training-free Test-time Adaptation

This is the author's official PyTorch implementation for ECBNN. This repo contains code for experiments in the NeurIPS 2022 paper:

Extrapolative Continuous-time Bayesian Neural Network for Fast Training-free Test-time Adaptation

Project Description

Human intelligence has shown remarkably lower latency and higher precision than most AI systems when processing non-stationary streaming data in real-time. Numerous neuroscience studies suggest that such abilities may be driven by internal predictive modeling. In this paper, we explore the possibility of introducing such a mechanism in unsupervised domain adaptation (UDA) for handling non-stationary streaming data for real-time streaming applications. We propose to formulate internal predictive modeling as a continuous-time Bayesian filtering problem within the context of a stochastic dynamical system. Such a dynamical system describes the dynamics of model parameters of a UDA model evolving with non-stationary streaming data. Building on such a dynamical system, we then develop extrapolative continuous-time Bayesian neural networks (ECBNN), which generalize existing Bayesian neural networks to represent temporal dynamics and allow us to extrapolate the distribution of model parameters before observing the incoming data, therefore effectively reducing the latency. Remarkably, our empirical results show that ECBNN is capable of continuously generating better distributions of model parameters along the time axis given historical data only, thereby achieving (1) training-free online adaptation with low latency, (2) gradually improved alignment between the source and target features and (3) gradually improved model performance over time during the real-time testing stage.

Method Overview

How to run

Experiment

NOTE: All our experiments are run on an AMD EPYC 7302P 16-core CPU and an RTX A5000 GPU.

Experiment settings:

• Python 3.8.13
• PyTorch 1.8.1
• CUDA 11.1

You may need to install the following packages:

pip install easydict
pip install progressbar

Running experiments

Streaming Rotating MNIST $\rightarrow$ USPS

• The streaming rotating MNIST and USPS data are generated and processed according to our paper. Please download our processed Streaming Rotating MNIST $\rightarrow$ USPS dataset following MIT License. You can put the data as follows:

Streaming_Rotating_MNIST_USPS
`-- data
    `-- MNIST
        |-- train_seq.pt
        |-- valid_seq.pt
        |-- test_seq.pt
    `-- USPS
        |-- train_seq.pt
        |-- valid_seq.pt
        |-- test_seq.pt

• To better understand our problem setting of test-time streaming domain adaptation, we visualize our data as follows. The left shows the streaming data (MNIST) on the source domain, while the right shows the streaming data (USPS) on the target domain:

• To train and further evaluate our ECBNN on source-testing and target-testing set: (To ensure the reproducibility, use 'CUBLAS_WORKSPACE_CONFIG=:4096:8' as a prefix)

cd Streaming_Rotating_MNIST_USPS
CUBLAS_WORKSPACE_CONFIG=:4096:8 python run.py --model_name ECBNN

Multi-view Lip Reading

• We followed the paper End-to-End Multi-View Lipreading, S. Petridis, Y. Wang, Z. Li, M. Pantic. British Machine Vision Conference. London, September 2017 to download and process the dataset. Currently, the official download page does not work. You may contact Guoying Zhao ([email protected]) to get a copy of the dataset. After downloading, you can put the data as follows:

OuluVS2
`-- data
    `-- unipadding
        |-- train.pt
        |-- valid.pt
        |-- test.pt

• The lip videos from $0 \degree$ view are the source domain while the lip videos from $30 \degree, 45 \degree, 60 \degree, 90 \degree$ views are the target domain.

• To train our ECBNN model from scratch (To ensure reproducibility, use 'CUBLAS_WORKSPACE_CONFIG=:4096:8' as a prefix)

cd OuluVS2
CUBLAS_WORKSPACE_CONFIG=:4096:8 python run.py --model_name ECBNN

Citation

If you use ECBNN or this codebase in your own work, please cite our paper:

@inproceedings{huangextrapolative,
  title={Extrapolative Continuous-time Bayesian Neural Network for Fast Training-free Test-time Adaptation},
  author={Huang, Hengguan and Gu, Xiangming and Wang, Hao and Xiao, Chang and Liu, Hongfu and Wang, Ye},
  booktitle={Advances in Neural Information Processing Systems},
  year={2022}
}

License

MIT license