# Overview

At **Imperfect Information Learning Team**,
**Center for Advanced Intelligence Project (AIP)**,
**RIKEN**,
we are developing reliable and robust **machine learning** methods/algorithms that can cope with various factors
such as **weak supervision**, **noisy supervision**, and **adversarial attacks**.
This page hosts our program codes used in our published papers.

The page contains 4 top-level research topics:

*Weakly supervised learning*is aimed at solving a learning task from only weakly supervised data (e.g., positive and unlabeled data, data with complementary labels, and data with partial labels);*Label-noise learning*is aimed at solving a learning task from possibly mislabeled data (i.e., the dataset for training a standard classifier is a mixture of correctly and incorrectly labeled data);*Adversarial robustness*is aimed at improving the robust accuracy of trained models against adversarial attacks (i.e., tiny perturbations applied on the data to flip the model predictions).- Our published papers that do not fall into the above 3 topics are included in
*other topics*.

For more related machine learning methods/algorithms, please check the following pages of our strategic partners:

- Sugiyama-Yokoya-Ishida Lab @ The University of Tokyo, led by Prof. Masashi Sugiyama
- Trustworthy Machine Learning Group @ Hong Kong Baptist University, led by Prof. Bo Han
- Trustworthy Machine Learning Lab @ The University of Sydney, led by Prof. Tongliang Liu

Disclaimer

The software available below is free of charge for research and education purposes. However, you must obtain a license from the author(s) to use it for commercial purposes. The software must not be distributed without prior permission of the author(s).

The software is supplied "as is" without warranty of any kind, and the author(s) disclaim any and all warranties, including but not limited to any implied warranties of merchantability and fitness for a particular purpose, and any warranties or non infringement. The user assumes all liability and responsibility for use of the software and in no event shall the author(s) be liable for damages of any kind resulting from its use.

Weakly supervised learning

*Standard supervised learning* relies on fully supervised or fully labeled data, which means that every instance is associated with a label, in order to teach a model how to map an instance to its label.
In practice, however, collecting such data is often *expensive* in terms of budget and/or time window for labeling or sampling big enough data, or even *impossible* since it may cause privacy and/or fairness issues.
*Weakly supervised learning* is aimed at solving a learning task from only weakly supervised or weakly labeled data, e.g., the positive class is present while the negative class is absent.
Our weakly supervised trained models still try to *capture the underlying map* from an instance to its label and to *predict the true label* of any instance exactly the same as fully supervised trained models.

## Positive-unlabeled learning

Positive-unlabeled learning is aimed as solving a binary classification problem only from positive and unlabeled data, without negative data.

Analysis of learning from positive and unlabeled data (NeurIPS 2014)

Convex formulation for learning from positive and unlabeled data (ICML 2015)

Theoretical comparisons of positive-unlabeled learning against positive-negative learning (NeurIPS 2016)

Semi-supervised classification based on classification from positive and unlabeled data (ICML 2017)

Positive-unlabeled learning with non-negative risk estimator (NeurIPS 2017)

Semi-supervised AUC optimization based on positive-unlabeled learning (Machine Learning 2018)

Classification from positive, unlabeled and biased negative data (ICML 2019)

Scalable evaluation and improvement of document set expansion via neural positive-unlabeled learning (EACL 2021)

Rethinking class-prior estimation for positive-unlabeled learning (ICLR 2022)

## Unlabeled-unlabeled learning

Unlabeled-unlabeled learning is aimed as solving a binary classification problem only from two sets of unlabeled data with different class priors.

On the minimal supervision for training any binary classifier from only unlabeled data (ICLR 2019)

Mitigating overfitting in supervised classification from two unlabeled datasets: A consistent risk correction approach (AISTATS 2020)

On symmetric losses for learning from corrupted labels (ICML 2020)

Binary Classification from multiple unlabeled datasets via surrogate set classification (ICML 2021)

Federated learning from only unlabeled data with class-conditional-sharing clients (ICLR 2022)

## Complementary-label learning

Complementary-label learning is aimed at training a multi-class classifier only from complementarily labeled data (a complementary label incidates a class which a patter does NOT belong to).

Learning from complementary labels (NeurIPS 2017)

Complementary-label learning for arbitrary losses and models (ICML 2019)

Learning with multiple complementary labels (ICML 2020)

Unbiased risk estimators can mislead: A case study of learning with complementary labels (ICML 2020)

## Partial-label learning

Partial-label learning is aimed at training a multi-class classifier only from partially labeled data (a partial label incidates a set of class labels one of which is the true one).

Learning with multiple complementary labels (ICML 2020)

Progressive identification of true labels for partial-label learning (ICML 2020)

Provably consistent partial-label learning (NeurIPS 2020)

PiCO: Contrastive label disambiguation for partial label learning (ICLR 2022)

Exploiting class activation value for partial-label learning (ICLR 2022)

Ambiguity-induced contrastive learning for instance-dependent partial label learning (IJCAI 2022)

## Pairwise learning

Pairwise learning is aimed at solving a classification problem from pairwise similarities/dissimilarities.

Classification from pairwise similarity and unlabeled data (ICML 2018)

Uncoupled regression from pairwise comparison data (NeurIPS 2019)

Learning from similarity-confidence data (ICML 2021)

Pointwise binary classification with pairwise confidence comparisons (ICML 2021)

Multiple-instance learning from similar and dissimilar bags (KDD 2021)

Learning from noisy pairwise similarity and unlabeled data (JMLR 2022)

## Self-supervised learning

Self-supervised learning is aimed at learning a representation from unlabeled data with various priors and pesuodo supervisions.

Contrastive embedding for generalized zero-shot learning (CVPR 2021)

Large-margin contrastive learning with distance polarization regularizer (ICML 2021)

Learning contrastive embedding in low-dimensional space (NeurIPS 2022)

Linearity-aware subspace clustering (AAAI 2022)

Robust audio-visual instance discrimination via active contrastive set mining (IJCAI 2022)

## Other

Binary classification from positive-confidence data (NeurIPS 2018)

Rethinking importance weighting for deep learning under distribution shift (NeurIPS 2020)

Meta discovery: Learning to discover novel classes given very limited data (ICLR 2022)

One positive label is sufficient: Single-positive multi-label learning with label enhancement (NeurIPS 2022)

Label-noise learning

Standard supervised learning relies on high-quality *clean labels*, which means that instances are with labels drawn from the *clean class-posterior probability*.
Nevertheless, if we require every instance to be associated with a label, our collected labels would probably come from non-expert annotators or be automatically annotated based on logs in practice.
Such lower-quality labels are called *noisy labels* and regarded as drawn from some *noisy/corrupted class-posterior probability*, resulting in a mixture of correctly and incorrectly labeled data.
*Label-noise learning* is aimed at solving a learning task from such possibly mislabeled data, where our models trained with noisy labels still try to *predict the true label* of any instance exactly the same as models trained with clean labels.

## Loss correction for class-conditional noise

Masking: A new perspective of noisy supervision (NeurIPS 2018)

Are anchor points really indispensable in label-noise learning? (NeurIPS 2019)

Dual T: Reducing estimation error for transition matrix in label-noise learning (NeurIPS 2020)

Learning noise transition matrix from only noisy labels via total variation regularization (ICML 2021)

Provably end-to-end label-noise learning without anchor points (ICML 2021)

## Sample selection/reweighting for class-conditional noise

Co-teaching: Robust training of deep neural networks with extremely noisy labels (NeurIPS 2018)

How does disagreement help generalization against label corruption? (ICML 2019)

Searching to exploit memorization effect in learning with noisy labels (ICML 2020)

Rethinking importance weighting for deep learning under distribution shift (NeurIPS 2020)

Sample selection with uncertainty of losses for learning with noisy labels (ICLR 2022)

## Other techniques for class-conditional noise

SIGUA: Forgetting may make learning with noisy labels more robust (ICML 2020)

Class2Simi: A noise reduction perspective on learning with noisy labels (ICML 2021)

Understanding and improving early stopping for learning with noisy labels (NeurIPS 2021)

## Instance-dependent noise

Part-dependent label noise: Towards instance-dependent label noise (NeurIPS 2020)

Tackling instance-dependent label noise via a universal probabilistic model (AAAI 2021)

Confidence scores make instance-dependent label-noise learning possible (ICML 2021)

Instance-dependent label-noise learning under a structural causal model (NeurIPS 2021)

Instance-dependent label-noise learning with manifold-regularized transition matrix estimation (CVPR 2022)

Learning with noisy labels revisited: A study using real-world human annotations (ICLR 2022)

Adversarial robustness

When we deploy models trained by standard supervised learning, they work well on *natural* test data.
However, those models cannot handle *adversarial* test data (also known as *adversarial examples*) that are algorithmically generated by *adversarial attacks*.
An adversarial attack is an algorithm which applies specially designed *tiny perturbations* on natural data to transform them into adversarial data, in order to mislead a trained model and let it give wrong predictions.
*Adversarial robustness* is aimed at improving the robust accuracy of trained models against adversarial attacks.

## Algorithm perspective

From the algorithm perspective, we would like to design *smarter loss/objective functions* for training standard models.
Specifically, given the knowledge that the test data may be adversarial, we carefully *simulate some adversarial attack during training*.
Thus, the model has already seen many adversarial training data in the past, and hopefully it can generalize to adversarial test data in the future.

Attacks which do not kill training make adversarial learning stronger (ICML 2020)

Geometry-aware instance-reweighted adversarial training (ICLR 2021)

Probabilistic margins for instance reweighting in adversarial training (NeurIPS 2021)

Reliable adversarial distillation with unreliable teachers (ICLR 2022)

Adversarial robustness through the lens of causality (ICLR 2022)

NoiLin: Improving adversarial training and correcting stereotype of noisy labels (TMLR 2022)

Adversarial training with complementary labels: On the benefit of gradually informative attacks (NeurIPS 2022)

## Model perspective

From the model perspective, we would like to design *smarter models* that can be robust even when they are trained by standard supervised learning.
Indeed, some models are *born to be more robust* than other models, because they possess special layers and components designed for removing or reducing the negative effects of adversarial attacks.

CIFS: Improving adversarial robustness of CNNs via channel-wise importance-based feature selection (ICML 2021)

Learning diverse-structured networks for adversarial robustness (ICML 2021)

Synergy-of-experts: Collaborate to improve adversarial robustness (NeurIPS 2022)

## Detection perspective

From the detection perspective, we would like to check whether a set of data has been attacked;
if the given data have been attacked, we want to know further which one has been attacked and which one has not.
For the data coming from the attacker but not the nature, we can *make more careful predictions* or even *abstain from giving predictions*.

## Application perspective

From the application perspective, we would like to apply the adversarial robustness to case studies. In particular, adversarial attacks are the ways of evaluating the vulnerabilities of machine learning models; adversarial training is the way of making the machine learning models robust against such vulnerabilities. Given a method empowered by machine learning approaches, we can leverage the adversarial attacks/training to evaluate/enhance the method's robustness.

Adversarial attacks and defense for non-parametric two sample tests (ICML 2022)

Towards adversarially robust image denoising (IJCAI 2022)

Other

Active feature acquisition with supervised matrix completion (KDD 2018)

Guide actor-critic for continuous control (ICLR 2018)

Do we need zero training loss after achieving zero training error? (ICML 2020)

Few-shot domain adaptation by causal mechanism transfer (ICML 2020)

Variational imitation learning with diverse-quality demonstrations (ICML 2020)

Robust imitation learning from noisy demonstrations (AISTATS 2021)

Incorporating causal graphical prior knowledge into predictive modeling via simple data augmentation (UAI 2021)