2022:

Prostate cancer therapy personalization via multi-modal deep learning on randomized phase III clinical trials

Andre Esteva, Jean Feng, Douwe van der Wal, ..., Amirata Ghorbani, et al.

Nature Digital Medicine

June 2022

[Paper]

Determining a patient's optimal therapy using a multi-modal deep learning architecture.

2021:

Who’s responsible? Jointly quantifying the contribution of the learning algorithm and data

Gal Yona, Amirata Ghorbani, James Zou

AAAI/ACM Conference on AI, Ethics, and Society (AIES)

May 2021

[Paper]

Introducing a new algorithm for simeltenous valuation of data and model

Data valuation for medical imaging using Shapley value and application to a large‑scale chest X‑ray dataset

Siyi Tang, Amirata Ghorbani, Rikiya Yamashita, Sameer Rehman, Jared A. Dunnmon, James Zou, Daniel L. Rubin

Nature Scientific Reports

Apr 2021

[Paper]

How Does Mixup Help With Robustness and Generalization

Linjun Zhang*, Zhun Deng*, Kenji Kawaguchi*, Amirata Ghorbani, James Zou

International Conference on Learning Representations (ICLR)

May 2021

[Paper]

A theoretical point of view for Mixup training

Improving Adversarial Robustness via Unlabeled Out-of-Domain Data

Zhun Deng*, Linjun Zhang*, Amirata Ghorbani, James Zou

International Conference on Artificial Intelligence and Statistics (AISTATS)

Apr 2021

[Paper]

We show with theoretical and empirical analysis that robustness of machine learning models against adversarial examples can be improved by adding cheap unlabeled data.

2020:

Neuron Shapley:

Discovering the Responsible Neurons

Amirata Ghorbani, James Zou

Neural Information Processing Systems (NeurIPS)

Dec 2020

[Paper] [Code]

We propose an efficient method for computing the contribution of individual neurons towards different aspects of a deep learning model's behavior (accuracy, robustness, etc).

Distributional Shapley:

A Distributional Framework for Data Valuation

Amirata Ghorbani*, Michael P. Kim*, James Zou (*equal contribution)

International Conference in Machine Learning (ICML)

Jul 2020

[Paper] [Code] [Slides][Presentation]

We propose a the distributional Shapley framework where the value of a data point is defined in the context of an underlying data distribution. We demonstrate the utility of this approach in a data market setting.

EchoNet-Dynamic:

Video-based AI for beat-to-beat assessment of cardiac function

David Ouyang, Bryan He, Amirata Ghorbani, et al.

Nature

March 2020

[Paper] [Code]

We present a video-based deep learning algorithm that surpasses the performance of human experts in the critical tasks of segmenting the left ventricle, estimating ejection fraction and assessing cardiomyopathy.

EchoNet:

Deep Learning Interpretation of Echocardiograms

Amirata Ghorbani*, David Ouyang* , et al. (*equal contribution)

Nature Digital Medicine

Jan 2020

[Paper] [Slides] [Presentation]

Using CNNs with a large new dataset, we show that deep learning applied to echocardiography can identify local cardiac structures, estimate cardiac function, and predict systemic phenotypes.

2019:

DermGan:

Synthetic Generation of Clinical Skin Images with Pathology

Amirata Ghorbani, Vivek Natarajan, David Coz, Yuan Liu

Machine Learning for Health (ML4H) at NeurIPS 2019

Dec 2019

[Paper] [Slides] [Poster]

We use Generative Adverserial Networks (GAN) to synthesize realistic clinical images with skin condition.

ACE:

Towards Automatic Concept-based Explanations

Amirata Ghorbani, James Wexler, James Zou, Been Kim

Neural Information Processing Systems (NeurIPS)

Dec 2019 (Poster Presentation)

[Paper] [Code] [Slides] [Poster]

We develop a new algorithm, ACE, to interpret deep convolutional neural networks. ACE is able to automatically extract visual concepts that are human-meaningful, coherent and important and use them to explain the neural network’s predictions.

Data Shapley:

Equitable Valuation of Data for Machine Learning

Amirata Ghorbani, James Zou

International Conference in Machine Learning (ICML)

Jun 2019 (Long Oral Presentation)

[Paper] [Code] [Slides] [Presentation] [Poster]

We propose efficient algorithms for estimating the value of data points or data sources for the supervised machine learning problem.

Multiaccuracy:

Black-box post-processing for fairness in classification

Michael P Kim*, Amirata Ghorbani*, James Zou (* equal contribution)

AAAI/ACM Conference on AI, Ethics, and Society (AIES)

Feb 2019 (Spotlight Presentation)

[Paper] [Code] [Slides] [Poster]

We develop a rigorous framework of multiaccuracy auditing and post-processing to improve predictor accuracy across identifiable subgroups.

Knockoffs for the mass: New feature importance statistics with false discovery guarantees

Jaime Roquero Gimenez, Amirata Ghorbani, James Zou

International Conference on Artificial Intelligence and Statistics (AISTATS)

Apr 2019 (Poster Presentation)

[Paper]

We develop an efficient algorithm to generate valid knockoffs from Bayesian Networks. Then we systematically evaluate knockoff test statistics and develop new statistics with improved power.

Interpretation of Neural Network is Fragile

Amirata Ghorbani*, Abubakar Abid*, James Zou (* equal contribution)

AAAI Conference on Artificial Intelligence (AAAI)

Feb 2019 (Oral Presentation)

[Paper] [Code] [Slides] [Presentation]

We demonstrate how to generate adversarial perturbations that produce perceptively indistinguishable inputs that are assigned the same predicted label, yet have very different interpretations


2018:

Embedding for Informative Missingness:

Deep Learning With Incomplete Data

Amirata Ghorbani, James Zou

Allerton Conference on Communication, Control, and Computing (Allerton)

Oct 2018 (Oral Presentation)

[Paper] [Slides]

We propose a general embedding approach to learn representations for missingness.

2017:

Mutation-Convolution-Max Layers Enhance Deep Learning of DNA Motifs

Abubakar Abid*, Amirata Ghorbani*, James Zou (* equal contribution)

Conference on Neural Information Processing Systems, Workshop on Computational Biology

Dec 2017 (Spotlight Presentation)

[Paper] [Slides]


Why Are Deep Networks Fragile:

Deformation of Intertwined Data

Amirata Ghorbani, James Zou

International Conference on Machine Learning, Reliable Machine Learning in the Wild Workshop

Aug 2017 (Oral Presentation)

[Paper] [Presentation] [Poster]

We quantify how fitting an intertwined dataset requires the model to deform the original space of the datasets in a way that small perturbations can result in big changes in the model’s output.

Blind Iterative Nonlinear Distortion Compensation Based on Thresholding

Masoumeh Azghani, Amirata Ghorbani, and Farokh Marvasti

IEEE Transactions on Circuits and Systems II

[Paper]

The sampling process in electrical devices includes nonlinear distortion that needs to be compensated to boost up the system efficiency. In this brief, a blind method is suggested for nonlinear distortion compensation.