POROver: Improving Safety and Reducing Overrefusal in Large Language Models with Overgeneration and Preference Optimization

Balancing safety and usefulness in large language models has become a critical challenge in recent years. Models often exhibit unsafe behavior or adopt an overly cautious approach, leading to frequent overrefusal of benign prompts, which reduces their usefulness. Addressing these issues requires methods that maintain safety while avoiding overrefusal. In this work, we examine how the overgeneration of training data using advanced teacher models (e.g., GPT-4o), including responses to both general-purpose and toxic prompts, influences the safety and overrefusal balance of instruction-following language models. Additionally, we present POROver, a strategy to use preference optimization methods in order to reduce overrefusal, via employing a superior teacher model's completions. Our results show that overgenerating completions for general-purpose prompts significantly improves the balance between safety and usefulness. Specifically, the F1 score calculated between safety and usefulness increases from 70.8% to 88.3%. Moreover, overgeneration for toxic prompts substantially reduces overrefusal, decreasing it from 94.4% to 45.2%. Furthermore, preference optimization algorithms, when applied with carefully curated preference data, can effectively reduce a model's overrefusal from 45.2% to 15.0% while maintaining comparable safety levels. Our code and data are available at https://github.com/batuhankmkaraman/POROver.

Adapting to Shifting Correlations with Unlabeled Data Calibration

Distribution shifts between sites can seriously degrade model performance since models are prone to exploiting unstable correlations. Thus, many methods try to find features that are stable across sites and discard unstable features. However, unstable features might have complementary information that, if used appropriately, could increase accuracy. More recent methods try to adapt to unstable features at the new sites to achieve higher accuracy. However, they make unrealistic assumptions or fail to scale to multiple confounding features. We propose Generalized Prevalence Adjustment (GPA for short), a flexible method that adjusts model predictions to the shifting correlations between prediction target and confounders to safely exploit unstable features. GPA can infer the interaction between target and confounders in new sites using unlabeled samples from those sites. We evaluate GPA on several real and synthetic datasets, and show that it outperforms competitive baselines.

Knockout: A simple way to handle missing inputs

Deep learning models can extract predictive and actionable information from complex inputs. The richer the inputs, the better these models usually perform. However, models that leverage rich inputs (e.g., multi-modality) can be difficult to deploy widely, because some inputs may be missing at inference. Current popular solutions to this problem include marginalization, imputation, and training multiple models. Marginalization can obtain calibrated predictions but it is computationally costly and therefore only feasible for low dimensional inputs. Imputation may result in inaccurate predictions because it employs point estimates for missing variables and does not work well for high dimensional inputs (e.g., images). Training multiple models whereby each model takes different subsets of inputs can work well but requires knowing missing input patterns in advance. Furthermore, training and retaining multiple models can be costly. We propose an efficient way to learn both the conditional distribution using full inputs and the marginal distributions. Our method, Knockout, randomly replaces input features with appropriate placeholder values during training. We provide a theoretical justification of Knockout and show that it can be viewed as an implicit marginalization strategy. We evaluate Knockout in a wide range of simulations and real-world datasets and show that it can offer strong empirical performance.

The 2024 Brain Tumor Segmentation (BraTS) Challenge: Glioma Segmentation on Post-treatment MRI

Gliomas are the most common malignant primary brain tumors in adults and one of the deadliest types of cancer. There are many challenges in treatment and monitoring due to the genetic diversity and high intrinsic heterogeneity in appearance, shape, histology, and treatment response. Treatments include surgery, radiation, and systemic therapies, with magnetic resonance imaging (MRI) playing a key role in treatment planning and post-treatment longitudinal assessment. The 2024 Brain Tumor Segmentation (BraTS) challenge on post-treatment glioma MRI will provide a community standard and benchmark for state-of-the-art automated segmentation models based on the largest expert-annotated post-treatment glioma MRI dataset. Challenge competitors will develop automated segmentation models to predict four distinct tumor sub-regions consisting of enhancing tissue (ET), surrounding non-enhancing T2/fluid-attenuated inversion recovery (FLAIR) hyperintensity (SNFH), non-enhancing tumor core (NETC), and resection cavity (RC). Models will be evaluated on separate validation and test datasets using standardized performance metrics utilized across the BraTS 2024 cluster of challenges, including lesion-wise Dice Similarity Coefficient and Hausdorff Distance. Models developed during this challenge will advance the field of automated MRI segmentation and contribute to their integration into clinical practice, ultimately enhancing patient care.

Assessing the significance of longitudinal data in Alzheimer's Disease forecasting

In this study, we employ a transformer encoder model to characterize the significance of longitudinal patient data for forecasting the progression of Alzheimer's Disease (AD). Our model, Longitudinal Forecasting Model for Alzheimer's Disease (LongForMAD), harnesses the comprehensive temporal information embedded in sequences of patient visits that incorporate multimodal data, providing a deeper understanding of disease progression than can be drawn from single-visit data alone. We present an empirical analysis across two patient groups-Cognitively Normal (CN) and Mild Cognitive Impairment (MCI)-over a span of five follow-up years. Our findings reveal that models incorporating more extended patient histories can outperform those relying solely on present information, suggesting a deeper historical context is critical in enhancing predictive accuracy for future AD progression. Our results support the incorporation of longitudinal data in clinical settings to enhance the early detection and monitoring of AD. Our code is available at \url{https://github.com/batuhankmkaraman/LongForMAD}.

BrainMorph: A Foundational Keypoint Model for Robust and Flexible Brain MRI Registration

We present a keypoint-based foundation model for general purpose brain MRI registration, based on the recently-proposed KeyMorph framework. Our model, called BrainMorph, serves as a tool that supports multi-modal, pairwise, and scalable groupwise registration. BrainMorph is trained on a massive dataset of over 100,000 3D volumes, skull-stripped and non-skull-stripped, from nearly 16,000 unique healthy and diseased subjects. BrainMorph is robust to large misalignments, interpretable via interrogating automatically-extracted keypoints, and enables rapid and controllable generation of many plausible transformations with different alignment types and different degrees of nonlinearity at test-time. We demonstrate the superiority of BrainMorph in solving 3D rigid, affine, and nonlinear registration on a variety of multi-modal brain MRI scans of healthy and diseased subjects, in both the pairwise and groupwise setting. In particular, we show registration accuracy and speeds that surpass current state-of-the-art methods, especially in the context of large initial misalignments and large group settings. All code and models are available at https://github.com/alanqrwang/brainmorph.

Longitudinal Mammogram Risk Prediction

Breast cancer is one of the leading causes of mortality among women worldwide. Early detection and risk assessment play a crucial role in improving survival rates. Therefore, annual or biennial mammograms are often recommended for screening in high-risk groups. Mammograms are typically interpreted by expert radiologists based on the Breast Imaging Reporting and Data System (BI-RADS), which provides a uniform way to describe findings and categorizes them to indicate the level of concern for breast cancer. Recently, machine learning (ML) and computational approaches have been developed to automate and improve the interpretation of mammograms. However, both BI-RADS and the ML-based methods focus on the analysis of data from the present and sometimes the most recent prior visit. While it is clear that temporal changes in image features of the longitudinal scans should carry value for quantifying breast cancer risk, no prior work has conducted a systematic study of this. In this paper, we extend a state-of-the-art ML model to ingest an arbitrary number of longitudinal mammograms and predict future breast cancer risk. On a large-scale dataset, we demonstrate that our model, LoMaR, achieves state-of-the-art performance when presented with only the present mammogram. Furthermore, we use LoMaR to characterize the predictive value of prior visits. Our results show that longer histories (e.g., up to four prior annual mammograms) can significantly boost the accuracy of predicting future breast cancer risk, particularly beyond the short-term. Our code and model weights are available at https://github.com/batuhankmkaraman/LoMaR.

Robust Learning via Conditional Prevalence Adjustment

Healthcare data often come from multiple sites in which the correlations between confounding variables can vary widely. If deep learning models exploit these unstable correlations, they might fail catastrophically in unseen sites. Although many methods have been proposed to tackle unstable correlations, each has its limitations. For example, adversarial training forces models to completely ignore unstable correlations, but doing so may lead to poor predictive performance. Other methods (e.g. Invariant risk minimization [4]) try to learn domain-invariant representations that rely only on stable associations by assuming a causal data-generating process (input X causes class label Y ). Thus, they may be ineffective for anti-causal tasks (Y causes X), which are common in computer vision. We propose a method called CoPA (Conditional Prevalence-Adjustment) for anti-causal tasks. CoPA assumes that (1) generation mechanism is stable, i.e. label Y and confounding variable(s) Z generate X, and (2) the unstable conditional prevalence in each site E fully accounts for the unstable correlations between X and Y . Our crucial observation is that confounding variables are routinely recorded in healthcare settings and the prevalence can be readily estimated, for example, from a set of (Y, Z) samples (no need for corresponding samples of X). CoPA can work even if there is a single training site, a scenario which is often overlooked by existing methods. Our experiments on synthetic and real data show CoPA beating competitive baselines.

Zero-shot Learning of Individualized Task Contrast Prediction from Resting-state Functional Connectomes

Given sufficient pairs of resting-state and task-evoked fMRI scans from subjects, it is possible to train ML models to predict subject-specific task-evoked activity using resting-state functional MRI (rsfMRI) scans. However, while rsfMRI scans are relatively easy to collect, obtaining sufficient task fMRI scans is much harder as it involves more complex experimental designs and procedures. Thus, the reliance on scarce paired data limits the application of current techniques to only tasks seen during training. We show that this reliance can be reduced by leveraging group-average contrasts, enabling zero-shot predictions for novel tasks. Our approach, named OPIC (short for Omni-Task Prediction of Individual Contrasts), takes as input a subject's rsfMRI-derived connectome and a group-average contrast, to produce a prediction of the subject-specific contrast. Similar to zero-shot learning in large language models using special inputs to obtain answers for novel natural language processing tasks, inputting group-average contrasts guides the OPIC model to generalize to novel tasks unseen in training. Experimental results show that OPIC's predictions for novel tasks are not only better than simple group-averages, but are also competitive with a state-of-the-art model's in-domain predictions that was trained using in-domain tasks' data.

Characterizing the Features of Mitotic Figures Using a Conditional Diffusion Probabilistic Model

Mitotic figure detection in histology images is a hard-to-define, yet clinically significant task, where labels are generated with pathologist interpretations and where there is no ``gold-standard'' independent ground-truth. However, it is well-established that these interpretation based labels are often unreliable, in part, due to differences in expertise levels and human subjectivity. In this paper, our goal is to shed light on the inherent uncertainty of mitosis labels and characterize the mitotic figure classification task in a human interpretable manner. We train a probabilistic diffusion model to synthesize patches of cell nuclei for a given mitosis label condition. Using this model, we can then generate a sequence of synthetic images that correspond to the same nucleus transitioning into the mitotic state. This allows us to identify different image features associated with mitosis, such as cytoplasm granularity, nuclear density, nuclear irregularity and high contrast between the nucleus and the cell body. Our approach offers a new tool for pathologists to interpret and communicate the features driving the decision to recognize a mitotic figure.