A5076316802- Functional Brain Connectivity Studies
- Domain Adaptation and Few-Shot Learning
- Machine Learning and Algorithms
- Statistical Methods and Inference
- Neural dynamics and brain function
- Privacy-Preserving Technologies in Data
- Machine Learning and Data Classification
- Sparse and Compressive Sensing Techniques
- Stochastic Gradient Optimization Techniques
- Bayesian Modeling and Causal Inference
- Gaussian Processes and Bayesian Inference
- Generative Adversarial Networks and Image Synthesis
- Topic Modeling
- Imbalanced Data Classification Techniques
- Face and Expression Recognition
- Bayesian Methods and Mixture Models
- Explainable Artificial Intelligence (XAI)
- Data Stream Mining Techniques
- EEG and Brain-Computer Interfaces
- Advanced Neuroimaging Techniques and Applications
- Gene expression and cancer classification
- Neural Networks and Applications
- Adversarial Robustness in Machine Learning
- Advanced Neural Network Applications
- Bioinformatics and Genomic Networks
Stanford University
2014-2025
Google (United States)
2021-2024
DeepMind (United Kingdom)
2024
Duke University
2024
Laboratoire d'Informatique de Paris-Nord
2023
University of Illinois Urbana-Champaign
2016-2022
University of Illinois Chicago
2022
Nature Inspires Creativity Engineers Lab
2013-2020
University of Illinois System
2018
The University of Texas at Austin
2009-2014
Quality control of MRI is essential for excluding problematic acquisitions and avoiding bias in subsequent image processing analysis. Visual inspection subjective impractical large scale datasets. Although automated quality assessments have been demonstrated on single-site datasets, it unclear that solutions can generalize to unseen data acquired at new sites. Here, we introduce the Control tool (MRIQC), a extracting measures fitting binary (accept/exclude) classifier. Our be run both...
Abstract Psychiatric disorders are characterized by major fluctuations in psychological function over the course of weeks and months, but dynamic characteristics brain this timescale healthy individuals unknown. Here, as a proof concept to address question, we present MyConnectome project. An intensive phenome-wide assessment single human was performed period 18 including functional structural connectivity using magnetic resonance imaging, physical health, gene expression metabolomics. A...
Federated learning enables training on a massive number of edge devices. To improve flexibility and scalability, we propose new asynchronous federated optimization algorithm. We prove that the proposed approach has near-linear convergence to global optimum, for both strongly convex restricted family non-convex problems. Empirical results show algorithm converges quickly tolerates staleness in various applications.
The large-scale sharing of task-based functional neuroimaging data has the potential to allow novel insights into organization mental function in brain, but field lagged behind other areas bioscience development resources. This paper describes OpenFMRI project (accessible online at http://www.openfmri.org), which aims provide community with a resource support open fMRI studies. We describe motivation project, focusing particularly on how this addresses some well-known challenges data....
Little is currently known about the coordination of neural activity over longitudinal timescales and how these changes relate to behavior. To investigate this issue, we used resting-state fMRI data from a single individual identify presence two distinct temporal states that fluctuated course 18 mo. These were associated with patterns time-resolved blood oxygen level dependent (BOLD) connectivity within scanning sessions also related significant alterations in global efficiency brain as well...
We propose three new robust aggregation rules for distributed synchronous Stochastic Gradient Descent~(SGD) under a general Byzantine failure model. The attackers can arbitrarily manipulate the data transferred between servers and workers in parameter server~(PS) architecture. prove resilience properties of these rules. Empirical analysis shows that proposed techniques outperform current approaches realistic use cases attack scenarios.
PurposeTo recognize and address various sources of bias essential for algorithmic fairness trustworthiness to contribute a just equitable deployment AI in medical imaging, there is an increasing interest developing imaging-based machine learning methods, also known as imaging artificial intelligence (AI), the detection, diagnosis, prognosis, risk assessment disease with goal clinical implementation. These tools are intended help improve traditional human decision-making imaging. However,...
Large Language models are among the most exciting technologies developed in last few years. While model's capabilities continue to improve, researchers, practitioners, and general public increasingly aware of some its shortcomings. What will it take build trustworthy large language models?
A central goal of cognitive neuroscience is to decode human brain activity-that is, infer mental processes from observed patterns whole-brain activation. Previous decoding efforts have focused on classifying activity into a small set discrete states. To attain maximal utility, framework must be open-ended, systematic, and context-sensitive-that capable interpreting numerous states, presented in arbitrary combinations, light prior information. Here we take steps towards this objective by...
Machine learning based decision making systems are increasingly affecting humans. An individual can suffer an undesirable outcome under such (e.g. denied credit) irrespective of whether the is fair or accurate. Individual recourse pertains to problem providing actionable set changes a person undertake in order improve their outcome. We propose algorithm that models underlying data distribution manifold. then provide mechanism generate smallest will individual's This be easily used for any...
Abstract Deep learning (DL) models can harness electronic health records (EHRs) to predict diseases and extract radiologic findings for diagnosis. With ambulatory chest radiographs (CXRs) frequently ordered, we investigated detecting type 2 diabetes (T2D) by combining radiographic EHR data using a DL model. Our model, developed from 271,065 CXRs 160,244 patients, was tested on prospective dataset of 9,943 CXRs. Here show the model effectively detected T2D with ROC AUC 0.84 16% prevalence....
Recently, new defense techniques have been developed to tolerate Byzantine failures for distributed machine learning. The model captures workers that behave arbitrarily, including malicious and compromised workers. In this paper, we break two prevailing Byzantine-tolerant techniques. Specifically show robust aggregation methods synchronous SGD -- coordinate-wise median Krum can be broken using attack strategies based on inner product manipulation. We prove our results theoretically, as well...
We present Zeno, a technique to make distributed machine learning, particularly Stochastic Gradient Descent (SGD), tolerant an arbitrary number of faulty workers. Zeno generalizes previous results that assumed majority non-faulty nodes; we need assume only one worker. Our key idea is suspect workers are potentially defective. Since this likely lead false positives, use ranking-based preference mechanism. prove the convergence SGD for non-convex problems under these scenarios. Experimental...
Self-attention performs well in long context but has quadratic complexity. Existing RNN layers have linear complexity, their performance is limited by the expressive power of hidden state. We propose a new class sequence modeling with complexity and an The key idea to make state machine learning model itself, update rule step self-supervised learning. Since updated training even on test sequences, our are called Test-Time Training (TTT) layers. consider two instantiations: TTT-Linear...
Recent interest in building foundation models for KGs has highlighted a fundamental challenge: knowledge-graph data is relatively scarce. The best-known are primarily human-labeled, created by pattern-matching, or extracted using early NLP techniques. While human-generated short supply, automatically of questionable quality. We present solution to this scarcity problem the form text-to-KG generator (KGGen), package that uses language create high-quality graphs from plaintext. Unlike other KG...