AAAI26 | Is the Information Bottleneck Robust Enough? Towards Label-Noise Resistant Information Bottleneck Learning

This repository provides the official implementation of LaT-IB, a method designed to address a key limitation of the classical Information Bottleneck (IB) framework: its vulnerability to label noise.

LaT-IB reformulates the IB objective in the latent representation space and introduces a new principled objective that improves robustness under noisy labels. Furthermore, we provide upper and lower bound analysis to theoretically justify the information meaning and effectiveness of our formulation.

We conduct extensive experiments on two representative tasks to validate LaT-IB:

• Image Classification
• Graph Node Classification

The results consistently demonstrate that LaT-IB achieves better robustness and generalization under noisy supervision.

./
├── IC (image classification, LaT-VIB)
│   ├── LaT_VIB.py
│   ├── README.md
│   ├── attack.py
│   ├── data
│   │   ├── animal_10N.py
│   │   ├── cifar.py
│   │   ├── datasets.py
│   │   └── utils.py
│   ├── log
│   ├── model
│   │   ├── __init__.py
│   │   ├── modelIB.py
│   │   └── resnet.py
│   ├── test_attack.py
│   ├── utils
│   │   ├── fmix.py
│   │   ├── loss.py
│   │   ├── mix.py
│   │   └── randaug.py
│   └── visual.py
├── NC (ndoe classification, LatT-GIB)
│   ├── LaT_GIB.py
│   ├── Noise_about.py
│   ├── README.md
│   ├── arg.py
│   ├── data
│   ├── data_loader.py
│   ├── log
│   ├── loss.py
│   ├── model.py
│   └── utils.py
├── README.md
└── tree.py

---

Environment requirement

IC

• torchvision
• PIL
• sklearn
• matplotlib (optional)

NC

• torch_geometric
• torch_sparse