Batuhan K. Karaman

Alan Q. Wang, Batuhan K. Karaman, Heejong Kim et al.

Interpretability for machine learning models in medical imaging (MLMI) is an important direction of research. However, there is a general sense of murkiness in what interpretability means. Why does the need for interpretability in MLMI arise? What goals does one actually seek to address when interpretability is needed? To answer these questions, we identify a need to formalize the goals and elements of interpretability in MLMI. By reasoning about real-world tasks and goals common in both medical image analysis and its intersection with machine learning, we identify five core elements of interpretability: localization, visual recognizability, physical attribution, model transparency, and actionability. From this, we arrive at a framework for interpretability in MLMI, which serves as a step-by-step guide to approaching interpretability in this context. Overall, this paper formalizes interpretability needs in the context of medical imaging, and our applied perspective clarifies concrete MLMI-specific goals and considerations in order to guide method design and improve real-world usage. Our goal is to provide practical and didactic information for model designers and practitioners, inspire developers of models in the medical imaging field to reason more deeply about what interpretability is achieving, and suggest future directions of interpretability research.

Batuhan K. Karaman, Katerina Dodelzon, Gozde B. Akar et al.

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.

Batuhan K. Karaman, Ishmam Zabir, Alon Benhaim et al.

Achieving both high safety and high usefulness simultaneously 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. A major factor underlying these behaviors is how the models are finetuned and aligned, particularly the nature and extent of the data used.In this work, we examine how overgenerating finetuning data with advanced teacher models (e.g., GPT-4o)-covering both general-purpose and toxic prompts-affects safety and usefulness in instruction-following language models.Additionally, we present POROver, an alignment strategy designed for models that are highly safe but prone to overrefusal. POROver employs preference optimization algorithms and leverages completions from an advanced teacher model to reduce overrefusals while maintaining safety.Our results show that overgenerating completions for general-purpose prompts significantly boosts safety with only a minimal impact on usefulness. Specifically, the F1 score calculated between safety and usefulness increases from 74.4% to 91.8% because of a substantial rise in safety. Moreover, overgeneration for toxic prompts raises usefulness from 11.1% to 57.6% while preserving safety. Finally, applying POROVer increases usefulness further-from 57.6% to 82.1%-while keeping safety at comparable levels. Our data and code are available at https://github.com/batuhankmkaraman/POROver.

Batuhan K. Karaman, Aditya Rawal, Suhaila Shakiah et al.

Reinforcement learning with verifiable rewards has emerged as a promising paradigm for enhancing the reasoning capabilities of large language models particularly in mathematics. Current approaches in this domain present a clear trade-off: PPO-style methods (e.g., GRPO/DAPO) offer training stability but exhibit slow learning trajectories due to their trust-region constraints on policy updates, while REINFORCE-style approaches (e.g., CISPO) demonstrate improved learning efficiency but suffer from performance instability as they clip importance sampling weights while still permitting non-zero gradients outside the trust-region. To address these limitations, we introduce DISPO, a simple yet effective REINFORCE-style algorithm that decouples the up-clipping and down-clipping of importance sampling weights for correct and incorrect responses, yielding four controllable policy update regimes. Through targeted ablations, we uncover how each regime impacts training: for correct responses, weights >1 increase the average token entropy (i.e., exploration) while weights <1 decrease it (i.e., distillation) -- both beneficial but causing gradual performance degradation when excessive. For incorrect responses, overly restrictive clipping triggers sudden performance collapse through repetitive outputs (when weights >1) or vanishing response lengths (when weights <1). By separately tuning these four clipping parameters, DISPO maintains the exploration-distillation balance while preventing catastrophic failures, achieving 61.04% on AIME'24 (vs. 55.42% CISPO and 50.21% DAPO) with similar gains across various benchmarks and models.