Jacopo Teneggi

Hi! I am Jacopo, a PhD student in the Computer Science Department at Johns Hopkins University, where I work with Jeremias Sulam and the Mathematical Institute for Data Science. I came to Baltimore from Politecnico di Torino, Italy, where I earned my Bachelor’s degree in Biomedical Engineering.

I am passionate about building AI systems we can actually trust—especially when people’s health is on the line. My research develops statistical methods that help us understand when and why AI works (or does not!).

What keeps me up at night:

Interpretability: I use tools like conditional independence testing, sequential testing, and the Shapley value to understand what drives AI decisions.
Uncertainty quantification: I work on conformal prediction and conformal risk control for generative systems.
AI for scientific discovery: I study how AI agents can accelerate molecular and protein design.

I have had the privilege of interning at Polymathic AI (2025), Profluent (2023), and nference (2021), where I got to see these challenges play out in real-world scenarios. Before going into machine learning, I worked with Rhonda Dzakpasu at Georgetown University studying the development of in-vitro neural-astrocyte networks (the real, biological kind!).

Let’s connect: jtenegg1 [at] jhu [dot] edu .

news

May, 2025	I am interning at Polymathic AI as a Research Scientist this summer. Excited to be working on AI agents for scientific discovery with Siavash Golkar, Tanya Marwah, Alberto Bietti, and collaborators!
May, 2025	Our latest preprint “Direct Preference Optimization for Adaptive Concept-based Explanations” has been posted to arXiv!
Mar, 2025	Our work “Conformal Risk Control for Semantic Uncertainty Quantification” has been accepted for publication in MICCAI!
Sep, 2024	Our paper “Testing Semantic Importance via Betting” has been accepted to NeurIPS, see you in Vancouver!
Nov, 2023	Our paper “Examination-level Supervision for Deep Learning-Based Intracranial Hemorrhage Detection on Head CT” has been accepted for publication in Radiology: Artificial Intelligence.
Nov, 2023	Our paper “SHAP-XRT: The Shapley Value Meets Conditional Independence Testing” has been accepted for publication in TMLR.
Jun, 2023	I am interning at Profluent as a ML Scientist this summer, working on uncertainty quantification for sequence generation under the supervision of Aadyot Bhatnagar.
Apr, 2023	Our paper “How to Trust Your Diffusion Model: A Convex Optimization Approach to Conformal Risk Control” has been accepted for presentation at ICML 2023!
Nov, 2022	Our paper “Weakly Supervised Learning Significantly Reduces the Number of Labels Required for Intracranial Hemorrhage Detection on Head CT” received a Trainee Research Prize at the RSNA 2022 Annual Meeting. Check out the abstract on Microsoft Research’s project InnerEye website!
Jul, 2022	Our paper “Fast Hierarchical Games for Image Explanations” has been accepted for publication in IEEE TPAMI!

publications

2025
arXiv

Direct Preference Optimization for Adaptive Concept-based Explanations

Teneggi, J., Zhenzhen, W., Yi, P., Shu, T., and Sulam, J.

Abstract

Concept-based explanation methods aim at making machine learning models more transparent by finding the most important semantic features of an input (e.g., colors, patterns, shapes) for a given prediction task. However, these methods generally ignore the communicative context of explanations, such as the preferences of a listener. For example, medical doctors understand explanations in terms of clinical markers, but patients may not, needing a different vocabulary to rationalize the same diagnosis. We address this gap with listener-adaptive explanations grounded in principles of pragmatic reasoning and the rational speech act. We introduce an iterative training procedure based on direct preference optimization where a speaker learns to compose explanations that maximize communicative utility for a listener. Our approach only needs access to pairwise preferences, which can be collected from human feedback, making it particularly relevant in real-world scenarios where a model of the listener may not be available. We demonstrate that our method is able to align speakers with the preferences of simulated listeners on image classification across three datasets, and further validate that pragmatic explanations generated with our method improve the classification accuracy of participants in a user study.
2025
MICCAI

Conformal Risk Control for Semantic Uncertainty Quantification in Computed Tomography

Teneggi, J., Stayman, W., and Sulam, J.

Abstract

Uncertainty quantification is necessary for developers, physicians, and regulatory agencies to build trust in machine learning predictors and improve patient care. Beyond measuring uncertainty, it is crucial to express it in clinically meaningful terms that provide actionable insights. This work introduces a conformal risk control (CRC) procedure for organ-dependent uncertainty estimation, ensuring high-probability coverage of the ground-truth image. We first present a high-dimensional CRC procedure that leverages recent ideas of length minimization. We make this procedure semantically adaptive to each patient’s anatomy and positioning of organs. Our method, sem-CRC, provides tighter uncertainty intervals with valid coverage on real-world computed tomography (CT) data while communicating uncertainty with clinically relevant features.
2024
NeurIPS

I Bet You Did Not Mean That: Testing Semantic Importance via Betting

Teneggi, J., and Sulam, J.

Abstract

Recent works have extended notions of feature importance to semantic concepts that are inherently interpretable to the users interacting with a black-box predictive model. Yet, precise statistical guarantees, such as false positive rate control, are needed to communicate findings transparently and to avoid unintended consequences in real-world scenarios. In this paper, we formalize the global (i.e., over a population) and local (i.e., for a sample) statistical importance of semantic concepts for the predictions of opaque models, by means of conditional independence, which allows for rigorous testing. We use recent ideas of sequential kernelized testing (SKIT) to induce a rank of importance across concepts, and showcase the effectiveness and flexibility of our framework on synthetic datasets as well as on image classification tasks using vision-language models such as CLIP.
2023
Rad AI

Examination-level Supervision for Deep Learning-Based Intracranial Hemorrhage Detection on Head CT

Teneggi, J., Yi, P., and Sulam, J.

Abstract

Purpose. To compare the effectiveness of weak supervision (i.e., with examination-level labels only) and strong supervision (i.e., with image-level labels) in training deep learning (DL) models for detection of intracranial hemorrhage (ICH) on head CT. Materials and Methods. In this retrospective study, an attention-based convolutional neural network was trained with either local (i.e., image-level) or global (i.e., examination-level) binary labels on the Radiological Society of North America (RSNA) 2019 Brain CT Hemorrhage Challenge dataset of 21,784 exams (41% ICH) and 752,803 images (14% ICH). The CQ500 (436 exams, 49% ICH) and CT-ICH (75 exams, 48% ICH) datasets were employed for external testing. Performance in detecting ICH was compared between weak (examination-level labels) and strong (image-level labels) learners as a function of the number of labels available during training. Results. On examination-level binary classification, weak and strong learners did not have different area under the receiver operating characteristic curve values on the internal validation split (0.96 vs. 0.96, p=0.64) and the CQ500 dataset (0.95 vs. 0.92, p=0.15). Weak learners outperformed strong ones on the CT-ICH dataset (0.92 vs. 0.90, p=0.03). Weak learners had better slice-level ICH detection performance when more than 10,000 labels were available for training (average , 0.73 vs 0.65, 𝑓1 p<0.001). Weaklysupervised models trained on the entire RSNA dataset required 35 times fewer labels than equivalent strong learners. Conclusion. Strongly-supervised models did not achieve better performance than weakly-supervised ones, which could dramatically reduce radiologist labor requirements for prospective dataset curation.
2023
TMLR

SHAP-XRT: The Shapley Value Meets Conditional Independence Testing

Teneggi, J.*, Bharti, B.*, Romano, Y., and Sulam, J.

Abstract

The complex nature of artificial neural networks raises concerns on their reliability, trustworthiness, and fairness in real-world scenarios. The Shapley value—a solution concept from game theory—is one of the most popular explanation methods for machine learning models. More traditionally, from a statistical perspective, feature importance is defined in terms of conditional independence. So far, these two approaches to interpretability and feature importance have been considered separate and distinct. In this work, we show that Shapley-based explanation methods and conditional independence testing are closely related. We introduce the SHAPley EXplanation Randomization Test (SHAP-XRT), a testing procedure inspired by the Conditional Randomization Test (CRT) for a specific notion of local (i.e., on a sample) conditional independence. With it, we prove that for binary classification problems, the marginal contributions in the Shapley value provide lower and upper bounds to the expected p-values of their respective tests. Furthermore, we show that the Shapley value itself provides an upper bound to the expected p-value of a global (i.e., overall) null hypothesis. As a result, we further our understanding of Shapley-based explanation methods from a novel perspective and characterize the conditions under which one can make statistically valid claims about feature importance via the Shapley value.
2023
ICML

How to Trust Your Diffusion model: A Convex Optimization Approach to Conformal Risk Control

Teneggi, J., Tivnan, M., Stayman, W., and Sulam, J.

Abstract

Score-based generative modeling, informally referred to as diffusion models, continue to grow in popularity across several important domains and tasks. While they provide high-quality and diverse samples from empirical distributions, important questions remain on the reliability and trustworthiness of these sampling procedures for their responsible use in critical scenarios. Conformal prediction is a modern tool to construct finite-sample, distribution-free uncertainty guarantees for any black-box predictor. In this work, we focus on image-to-image regression tasks and we present a generalization of the Risk-Controlling Prediction Sets (RCPS) procedure, that we term K-RCPS, which allows to (i) provide entrywise calibrated intervals for future samples of any diffusion model, and (ii) control a certain notion of risk with respect to a ground truth image with minimal mean interval length. Differently from existing conformal risk control procedures, ours relies on a novel convex optimization approach that allows for multidimensional risk control while provably minimizing the mean interval length. We illustrate our approach on two real-world image denoising problems: on natural images of faces as well as on computed tomography (CT) scans of the abdomen, demonstrating state of the art performance.
2022
TPAMI

Fast Hierarchical Games for Image Explanations

Teneggi, J., Luster, A., and Sulam, J.

Abstract

As modern complex neural networks keep breaking records and solving harder problems, their predictions also become less and less intelligible. The current lack of interpretability often undermines the deployment of accurate machine learning tools in sensitive settings. In this work, we present a model-agnostic explanation method for image classification based on a hierarchical extension of Shapley coefficients–Hierarchical Shap (h-Shap)–that resolves some of the limitations of current approaches. Unlike other Shapley-based explanation methods, h-Shap is scalable and can be computed without the need of approximation. Under certain distributional assumptions, such as those common in multiple instance learning, h-Shap retrieves the exact Shapley coefficients with an exponential improvement in computational complexity. We compare our hierarchical approach with popular Shapley-based and non-Shapley-based methods on a synthetic dataset, a medical imaging scenario, and a general computer vision problem, showing that h-Shap outperforms the state of the art in both accuracy and runtime. Code and experiments are made publicly available.
2021
Front Neurosci

Fitting Splines to Axonal Arbors Quantifies Relationship Between Branch Order and Geometry

Athey, T., Teneggi, J., Vogelstein, J., Tward, D., Mueller, U., and Miller, M.

Abstract

Neuromorphology is crucial to identifying neuronal subtypes and understanding learning. It is also implicated in neurological disease. However, standard morphological analysis focuses on macroscopic features such as branching frequency and connectivity between regions, and often neglects the internal geometry of neurons. In this work, we treat neuron trace points as a sampling of differentiable curves and fit them with a set of branching B-splines. We designed our representation with the Frenet-Serret formulas from differential gemoetry in mind. The Frenet-Serret formulas completely characterize smooth curves, and involve two parameters, curvature and torsion. Our representation makes it possible to compute these parameters from neuron traces in closed form. These parameters are defined continuously along the curve, in contrast to other parameters like tortuosity which depend on start and end points. We applied our method to a dataset of cortical projection neurons traced in two mouse brains, and found that the parameters are distributed differently between primary, collateral, and terminal axon branches, thus quantifying geometric differences between different components of an axonal arbor. The results agreed in both brains, further validating our representation. The code used in this work can be readily applied to neuron traces in SWC format and is available in our open-source Python package brainlit: http://brainlit.neurodata.io/.
2021
Phys Rev E

Entropy estimation within in vitro neural-astrocyte networks as a measure of development instability

Teneggi, J., Chen, X., Balu, A., Barrett, C., Grisolia, G., Lucia, U., and Dzakpasu, R.

Abstract

The brain demands a significant fraction of the energy budget in an organism; in humans, it accounts for 2% of the body mass, but utilizes 20% of the total energy metabolized. This is due to the large load required for information processing; spiking demands from neurons are high but are a key component to understanding brain functioning. Astrocytic brain cells contribute to the healthy functioning of brain circuits by mediating neuronal network energy and facilitating the formation and stabilization of synaptic connectivity. During development, spontaneous activity influences synaptic formation, shaping brain circuit construction, and adverse astrocyte mutations can lead to pathological processes impacting cognitive impairment due to inefficiencies in network spiking activity. We have developed a measure that quantifies information stability within in vitro networks consisting of mixed neural-astrocyte cells. Brain cells were harvested from mice with mutations to a gene associated with the strongest known genetic risk factor for Alzheimer’s disease, APOE. We calculate energy states of the networks and using these states, we present an entropy-based measure to assess changes in information stability over time. We show that during development, stability profiles of spontaneous network activity are modified by exogenous astrocytes and that network stability, in terms of the rate of change of entropy, is allele dependent.