Search Results for author: Thomas Hartvigsen

Found 34 papers, 19 papers with code

Low-Bit Quantization Favors Undertrained LLMs: Scaling Laws for Quantized LLMs with 100T Training Tokens

no code implementations26 Nov 2024 Xu Ouyang, Tao Ge, Thomas Hartvigsen, Zhisong Zhang, Haitao Mi, Dong Yu

To gain deeper insights into this trend, we study over 1500 quantized LLM checkpoints of various sizes and at different training levels (undertrained or fully trained) in a controlled setting, deriving scaling laws for understanding the relationship between QiD and factors such as the number of training tokens, model size and bit width.

Quantization

BendVLM: Test-Time Debiasing of Vision-Language Embeddings

1 code implementation7 Nov 2024 Walter Gerych, Haoran Zhang, Kimia Hamidieh, Eileen Pan, Maanas Sharma, Thomas Hartvigsen, Marzyeh Ghassemi

Vision-language model (VLM) embeddings have been shown to encode biases present in their training data, such as societal biases that prescribe negative characteristics to members of various racial and gender identities.

Attribute Image Generation +2

Identifying Implicit Social Biases in Vision-Language Models

no code implementations1 Nov 2024 Kimia Hamidieh, Haoran Zhang, Walter Gerych, Thomas Hartvigsen, Marzyeh Ghassemi

Finally, we conduct an analysis of the source of such biases, by showing that the same harmful stereotypes are also present in a large image-text dataset used to train CLIP models for examples of biases that we find.

Fairness

SkipSNN: Efficiently Classifying Spike Trains with Event-attention

1 code implementation29 Oct 2024 Hang Yin, Yao Su, LiPing Liu, Thomas Hartvigsen, Xin Dai, Xiangnan Kong

Spike train classification has recently become an important topic in the machine learning community, where each spike train is a binary event sequence with \emph{temporal-sparsity of signals of interest} and \emph{temporal-noise} properties.

Computational Efficiency

Offline Reinforcement Learning With Combinatorial Action Spaces

no code implementations28 Oct 2024 Matthew Landers, Taylor W. Killian, Hugo Barnes, Thomas Hartvigsen, Afsaneh Doryab

Reinforcement learning problems often involve large action spaces arising from the simultaneous execution of multiple sub-actions, resulting in combinatorial action spaces.

reinforcement-learning Reinforcement Learning

Backdoor in Seconds: Unlocking Vulnerabilities in Large Pre-trained Models via Model Editing

no code implementations23 Oct 2024 Dongliang Guo, Mengxuan Hu, Zihan Guan, Junfeng Guo, Thomas Hartvigsen, Sheng Li

Through empirical studies on the capability for performing backdoor attack in large pre-trained models ($\textit{e. g.,}$ ViT), we find the following unique challenges of attacking large pre-trained models: 1) the inability to manipulate or even access large training datasets, and 2) the substantial computational resources required for training or fine-tuning these models.

Backdoor Attack Image Captioning +3

Math Neurosurgery: Isolating Language Models' Math Reasoning Abilities Using Only Forward Passes

1 code implementation22 Oct 2024 Bryan R. Christ, Zack Gottesman, Jonathan Kropko, Thomas Hartvigsen

MathNeuro builds on existing work by using weights and activations to calculate parameter importance, but isolates math-specific parameters by removing those important for general language tasks.

GSM8K Language Modeling +3

Wait, but Tylenol is Acetaminophen... Investigating and Improving Language Models' Ability to Resist Requests for Misinformation

no code implementations30 Sep 2024 Shan Chen, Mingye Gao, Kuleen Sasse, Thomas Hartvigsen, Brian Anthony, Lizhou Fan, Hugo Aerts, Jack Gallifant, Danielle Bitterman

Background: Large language models (LLMs) are trained to follow directions, but this introduces a vulnerability to blindly comply with user requests even if they generate wrong information.

Logical Reasoning Misinformation

FedMedICL: Towards Holistic Evaluation of Distribution Shifts in Federated Medical Imaging

1 code implementation11 Jul 2024 Kumail Alhamoud, Yasir Ghunaim, Motasem Alfarra, Thomas Hartvigsen, Philip Torr, Bernard Ghanem, Adel Bibi, Marzyeh Ghassemi

In response, we introduce FedMedICL, a unified framework and benchmark to holistically evaluate federated medical imaging challenges, simultaneously capturing label, demographic, and temporal distribution shifts.

Diversity Federated Learning

Composable Interventions for Language Models

1 code implementation9 Jul 2024 Arinbjorn Kolbeinsson, Kyle O'Brien, Tianjin Huang, ShangHua Gao, Shiwei Liu, Jonathan Richard Schwarz, Anurag Vaidya, Faisal Mahmood, Marinka Zitnik, Tianlong Chen, Thomas Hartvigsen

Test-time interventions for language models can enhance factual accuracy, mitigate harmful outputs, and improve model efficiency without costly retraining.

knowledge editing Machine Unlearning +1

Language Models are Surprisingly Fragile to Drug Names in Biomedical Benchmarks

1 code implementation17 Jun 2024 Jack Gallifant, Shan Chen, Pedro Moreira, Nikolaj Munch, Mingye Gao, Jackson Pond, Leo Anthony Celi, Hugo Aerts, Thomas Hartvigsen, Danielle Bitterman

Medical knowledge is context-dependent and requires consistent reasoning across various natural language expressions of semantically equivalent phrases.

MedQA

Dr-LLaVA: Visual Instruction Tuning with Symbolic Clinical Grounding

no code implementations29 May 2024 Shenghuan Sun, Alexander Schubert, Gregory M. Goldgof, Zhiqing Sun, Thomas Hartvigsen, Atul J. Butte, Ahmed Alaa

For this purpose, we propose a new alignment algorithm that uses symbolic representations of clinical reasoning to ground VLMs in medical knowledge.

PolygloToxicityPrompts: Multilingual Evaluation of Neural Toxic Degeneration in Large Language Models

2 code implementations15 May 2024 Devansh Jain, Priyanshu Kumar, Samuel Gehman, Xuhui Zhou, Thomas Hartvigsen, Maarten Sap

Recent advances in large language models (LLMs) have led to their extensive global deployment, and ensuring their safety calls for comprehensive and multilingual toxicity evaluations.

Benchmarking

TAXI: Evaluating Categorical Knowledge Editing for Language Models

1 code implementation23 Apr 2024 Derek Powell, Walter Gerych, Thomas Hartvigsen

For example, in learning a korat is a type of cat, you also infer it is a mammal and has claws, ensuring your model of the world is consistent.

knowledge editing Multiple-choice

UniTS: A Unified Multi-Task Time Series Model

1 code implementation29 Feb 2024 ShangHua Gao, Teddy Koker, Owen Queen, Thomas Hartvigsen, Theodoros Tsiligkaridis, Marinka Zitnik

We introduce UniTS, a unified multi-task time series model that utilizes task tokenization to integrate predictive and generative tasks into a single framework.

Anomaly Detection Imputation +3

MATHWELL: Generating Educational Math Word Problems Using Teacher Annotations

2 code implementations24 Feb 2024 Bryan R Christ, Jonathan Kropko, Thomas Hartvigsen

To be educational, problems must be solvable, have accurate answers, and, most importantly, be educationally appropriate.

Language Modeling Language Modelling +1

Improving Black-box Robustness with In-Context Rewriting

1 code implementation13 Feb 2024 Kyle O'Brien, Nathan Ng, Isha Puri, Jorge Mendez, Hamid Palangi, Yoon Kim, Marzyeh Ghassemi, Thomas Hartvigsen

Machine learning models for text classification often excel on in-distribution (ID) data but struggle with unseen out-of-distribution (OOD) inputs.

News Classification text-classification +1

Learning from Time Series under Temporal Label Noise

no code implementations6 Feb 2024 Sujay Nagaraj, Walter Gerych, Sana Tonekaboni, Anna Goldenberg, Berk Ustun, Thomas Hartvigsen

We first demonstrate the importance of modelling the temporal nature of the label noise function and how existing methods will consistently underperform.

Time Series

Machine Learning for Health symposium 2023 -- Findings track

no code implementations1 Dec 2023 Stefan Hegselmann, Antonio Parziale, Divya Shanmugam, Shengpu Tang, Mercy Nyamewaa Asiedu, Serina Chang, Thomas Hartvigsen, Harvineet Singh

A collection of the accepted Findings papers that were presented at the 3rd Machine Learning for Health symposium (ML4H 2023), which was held on December 10, 2023, in New Orleans, Louisiana, USA.

Multi-State Brain Network Discovery

no code implementations4 Nov 2023 Hang Yin, Yao Su, Xinyue Liu, Thomas Hartvigsen, Yanhua Li, Xiangnan Kong

We refer to such brain networks as multi-state, and this mixture can help us understand human behavior.

Continuous Time Evidential Distributions for Irregular Time Series

1 code implementation25 Jul 2023 Taylor W. Killian, Haoran Zhang, Thomas Hartvigsen, Ava P. Amini

Prevalent in many real-world settings such as healthcare, irregular time series are challenging to formulate predictions from.

Irregular Time Series Time Series +1

Interpretable Unified Language Checking

1 code implementation7 Apr 2023 Tianhua Zhang, Hongyin Luo, Yung-Sung Chuang, Wei Fang, Luc Gaitskell, Thomas Hartvigsen, Xixin Wu, Danny Fox, Helen Meng, James Glass

Despite recent concerns about undesirable behaviors generated by large language models (LLMs), including non-factual, biased, and hateful language, we find LLMs are inherent multi-task language checkers based on their latent representations of natural and social knowledge.

Fact Checking Fairness +2

Finding Short Signals in Long Irregular Time Series with Continuous-Time Attention Policy Networks

no code implementations8 Feb 2023 Thomas Hartvigsen, Jidapa Thadajarassiri, Xiangnan Kong, Elke Rundensteiner

Using this insight, we then propose CAT, a model that classifies multivariate ITS by explicitly seeking highly-relevant portions of an input series' timeline.

Imputation Irregular Time Series +2

Aging with GRACE: Lifelong Model Editing with Discrete Key-Value Adaptors

1 code implementation NeurIPS 2023 Thomas Hartvigsen, Swami Sankaranarayanan, Hamid Palangi, Yoon Kim, Marzyeh Ghassemi

We propose GRACE, a lifelong model editing method, which implements spot-fixes on streaming errors of a deployed model, ensuring minimal impact on unrelated inputs.

Model Editing World Knowledge

Class-Specific Explainability for Deep Time Series Classifiers

1 code implementation11 Oct 2022 Ramesh Doddaiah, Prathyush Parvatharaju, Elke Rundensteiner, Thomas Hartvigsen

Instead, when a classifier is choosing between many classes, an effective explanation must show what sets the chosen class apart from the rest.

Time Series Time Series Analysis +1

Stop&Hop: Early Classification of Irregular Time Series

1 code implementation21 Aug 2022 Thomas Hartvigsen, Walter Gerych, Jidapa Thadajarassiri, Xiangnan Kong, Elke Rundensteiner

We bridge this gap and study early classification of irregular time series, a new setting for early classifiers that opens doors to more real-world problems.

Early Classification General Classification +3

TWEET-FID: An Annotated Dataset for Multiple Foodborne Illness Detection Tasks

no code implementations LREC 2022 Ruofan Hu, Dongyu Zhang, Dandan Tao, Thomas Hartvigsen, Hao Feng, Elke Rundensteiner

To accelerate the development of machine learning-based models for foodborne outbreak detection, we thus present TWEET-FID (TWEET-Foodborne Illness Detection), the first publicly available annotated dataset for multiple foodborne illness incident detection tasks.

slot-filling Slot Filling

The Road to Explainability is Paved with Bias: Measuring the Fairness of Explanations

no code implementations6 May 2022 Aparna Balagopalan, Haoran Zhang, Kimia Hamidieh, Thomas Hartvigsen, Frank Rudzicz, Marzyeh Ghassemi

Across two different blackbox model architectures and four popular explainability methods, we find that the approximation quality of explanation models, also known as the fidelity, differs significantly between subgroups.

BIG-bench Machine Learning Fairness

ToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detection

1 code implementation ACL 2022 Thomas Hartvigsen, Saadia Gabriel, Hamid Palangi, Maarten Sap, Dipankar Ray, Ece Kamar

To help mitigate these issues, we create ToxiGen, a new large-scale and machine-generated dataset of 274k toxic and benign statements about 13 minority groups.

Hate Speech Detection Language Modelling

Human Attention Maps for Text Classification: Do Humans and Neural Networks Focus on the Same Words?

no code implementations ACL 2020 Cansu Sen, Thomas Hartvigsen, Biao Yin, Xiangnan Kong, Elke Rundensteiner

Motivated by human attention, computational attention mechanisms have been designed to help neural networks adjust their focus on specific parts of the input data.

General Classification text-classification +1

Reducing Computation in Recurrent Networks by Selectively Updating State Neurons

no code implementations25 Sep 2019 Thomas Hartvigsen, Cansu Sen, Xiangnan Kong, Elke Rundensteiner

As a result, even for high-dimensional hidden states, all dimensions are updated at each timestep regardless of the recurrent memory cell.

Adaptive-Halting Policy Network for Early Classification

1 code implementation KDD 2019 Thomas Hartvigsen, Cansu Sen, Xiangnan Kong, Elke Rundensteiner

Early classification of time series is the prediction of the class label of a time series before it is observed in its entirety.

Classification Early Classification +3

Cannot find the paper you are looking for? You can Submit a new open access paper.