About

Konrad P Kording, PhD, is a Professor of Bioengineering at the University of Pennsylvania. He collaborates with clinicians to develop and critically evaluate machine learning applications in healthcare, with a focus on causal inference methods for understanding treatment effects when randomized trials aren't feasible. His current work includes video-based health analysis systems for infant developmental assessment. Dr. Kording brings rigorous technical expertise combined with critical assessment of AI systems, ensuring that machine learning tools are validated, interpretable, and genuinely useful for patient care. He serves as Co-Director of the CIFAR Learning in Machines & Brains Program and leads research aimed at reverse engineering and simulating complete nervous systems, starting with C. elegans, as a step toward understanding more complex brains and the relationship between neural structure and function. His lab employs a transdisciplinary approach centered on causal inference in observational data and understanding intelligence through an evolutionary lens. Through initiatives like Neuromatch, he has trained over 10,000 students globally, democratizing computational neuroscience education across more than 100 countries. His Community for Rigor initiative develops platforms to address research biases and improve scientific methodology. The lab's work spans neural data analysis, brain-computer interfaces, and machine learning applications in medicine, always maintaining methodological rigor and a collaborative, data-driven approach to understanding brain computation and building better AI systems.

Research topics

• Computer Science
• Artificial Intelligence
• Machine Learning
• Psychology
• Data science
• Neuroscience
• Political Science
• Cognitive science
• Cognitive psychology
• Cartography
• Mathematics
• Geography
• Physical medicine and rehabilitation
• Management
• Developmental psychology
• Medicine
• Communication
• Mathematics education
• Engineering
• Epistemology
• Management science

Selected publications

• Compiling molecular ultrastructure into neural dynamics

  arXiv (Cornell University) · 2026-03-26

  preprintOpen access1st authorCorresponding

  High-resolution brain imaging can now capture not just synapse locations but their molecular composition, with the cost of such mapping falling exponentially. Yet such ultrastructural data has so far told us little about local neuronal physiology - specifically, the parameters (e.g., synaptic efficacies, local conductances) that govern neural dynamics. We propose to translate molecularly annotated ultrastructure into physiology, introducing the concept of an ultrastructure-to-dynamics compiler: a learned mapping from molecularly annotated ultrastructure to simulator-ready, uncertainty-aware physiological parameters. The requirement is paired training data, with jointly acquired ultrastructure from imaging, and dynamical responses to perturbations from physiological experiments. With this data we can train models that predict local physiology directly from structure. Such a compiler would support biophysical simulations by turning anatomical maps into models of circuit dynamics, shifting structure-to-function from a descriptive program to a predictive one and opening routes to understanding neural computation and forecasting intervention effects.

  PublisherDOI

• Mid-superior temporal sulcus encodes spatial context and behavioral state in freely moving macaques

  bioRxiv (Cold Spring Harbor Laboratory) · 2026-04-17

  articleOpen access

  Primate neuroscience has traditionally studied the brain under highly constrained conditions, limiting our understanding of neural function during real-world behavior. The mid-superior temporal sulcus (mSTS) is implicated in social perception, but its role during unconstrained behavior has not been tested. Here we performed wireless depth-electrode recordings from both banks of mSTS in macaques freely exploring a large three-dimensional arena, combined with 3D pose tracking and behavioral segmentation. Neural encoding models revealed mSTS firing rates were jointly modulated by spatial position, body kinematics, and geometric visual proxies, preferentially encoded in allocentric coordinates but with a shift toward body-centric encoding during vertical exploration. Neural populations carried decodable information about discrete behavioral syllables, with broad temporal generalization and neural similarity that tracked the sequential structure of behavior. Population manifold analysis revealed that the same behavior occupied different regions of population space at different spatial locations, and population dynamics showed structured organization around behavioral transitions. Together, these results suggest that mSTS populations carry joint information about spatial context and behavioral state during natural behavior.

  PublisherDOI

• Compiling to Linear Neurons

  Proceedings of the ACM on Programming Languages · 2026-01-08

  articleOpen access

  We don’t program neural networks directly. Instead, we rely on an indirect style where learning algorithms, like gradient descent, determine a neural network’s function by learning from data. This indirect style is often a virtue; it empowers us to solve problems that were previously impossible. But it lacks discrete structure. We can’t compile most algorithms into a neural network—even if these algorithms could help the network learn. This limitation occurs because discrete algorithms are not obviously differentiable, making them incompatible with the gradient-based learning algorithms that determine a neural network’s function. To address this, we introduce Cajal: a typed, higher-order and linear programming language intended to be a minimal vehicle for exploring a direct style of programming neural networks. We prove Cajal programs compile to linear neurons, allowing discrete algorithms to be expressed in a differentiable form compatible with gradient-based learning. With our implementation of Cajal, we conduct several experiments where we link these linear neurons against other neural networks to determine part of their function prior to learning. Linking with these neurons allows networks to learn faster, with greater data-efficiency, and in a way that’s easier to debug. A key lesson is that linear programming languages provide a path towards directly programming neural networks, enabling a rich interplay between learning and the discrete structures of ordinary programming.

  PublisherDOI

• Compiling molecular ultrastructure into neural dynamics

  ArXiv.org · 2026-03-26

  articleOpen access1st authorCorresponding

  High-resolution brain imaging can now capture not just synapse locations but their molecular composition, with the cost of such mapping falling exponentially. Yet such ultrastructural data has so far told us little about local neuronal physiology - specifically, the parameters (e.g., synaptic efficacies, local conductances) that govern neural dynamics. We propose to translate molecularly annotated ultrastructure into physiology, introducing the concept of an ultrastructure-to-dynamics compiler: a learned mapping from molecularly annotated ultrastructure to simulator-ready, uncertainty-aware physiological parameters. The requirement is paired training data, with jointly acquired ultrastructure from imaging, and dynamical responses to perturbations from physiological experiments. With this data we can train models that predict local physiology directly from structure. Such a compiler would support biophysical simulations by turning anatomical maps into models of circuit dynamics, shifting structure-to-function from a descriptive program to a predictive one and opening routes to understanding neural computation and forecasting intervention effects.

  PublisherOA PDF

• Widespread use of invalid statistical tests in biomedical machine learning

  bioRxiv (Cold Spring Harbor Laboratory) · 2026-05-20

  articleOpen access

  Abstract Machine learning is accelerating biomedical research. Cross-validation is widely used to compare predictive performance – not only to benchmark algorithms, but also to inform scientific applications, such as ranking biomarkers. However, prediction performance estimates across cross-validation folds are not independent. Standard tests for comparing prediction performance (e.g., paired t-test) assume independence and can therefore inflate false positive rates. In a PRISMA-guided meta-analysis of 210 studies (impact factor ≥15, 1 June 2020 – 1 June 2025), we find that 97% ignored fold dependence when comparing prediction performance. This problem is ubiquitous across scientific fields and unaffected by impact factor, rigor-promoting policies, or open science practices. Simulations across 420 scenarios spanning four diverse datasets show that ignoring fold dependence leads to invalid false positive control in most settings. Repeated cross-validation further compounds this problem, with false positive rates rising toward 100% as the number of repetitions grows. Existing fold-dependence-aware tests rely on strong assumptions because the variance of fold-level statistics and the between-fold correlation cannot be disentangled under standard cross-validation. We therefore propose the SHARP (Split-HAlf RePeated) test, a simple modification to standard cross-validation that enables direct estimation of variance and correlation. Benchmarked against 12 tests, SHARP provides the best overall balance of false-positive control, statistical power, and confidence-interval calibration across simulation schemes. We conclude by providing best practices and reporting guidelines for valid model comparison inference in biomedical machine learning and beyond.

  PublisherOA PDFDOI

• A preregistered, open pipeline for early cerebral palsy risk assessment from infant videos

  GigaScience · 2026-01-01

  articleOpen accessSenior author

  Cerebral palsy (CP), affecting approximately 1 in 500 children due to abnormal brain development, impacts movement control. Early risk assessment via the general movements assessment (GMA) at 3-4 months is highly predictive for CP but relies on trained clinicians. Machine-learning-based approaches for predicting GMA score from video have shown considerable promise, but typically rely on dataset-specific preprocessing, custom feature sets, and manually designed model pipelines, which make external benchmarking more difficult. This, combined with strict privacy constraints on sharing data, makes it challenging to train and evaluate models across datasets, which is important for assessing clinical utility. There is therefore a need to develop approaches that will work across different datasets to enable multi-site dataset aggregation and model training. To address this gap, we developed an end-to-end pipeline that uses off-the-shelf pose estimation, general-purpose feature extraction, and automated machine learning-none of which are tuned to a specific dataset. We applied this approach to a newly generated large dataset of 1053 infants (with approximately 10-12% positive class for adverse GMA outcome, drawn from a high-risk clinical cohort) within a preregistered study design. Model performance was evaluated on a strict "lock-box" test set, which remained untouched during any phase of model development or preprocessing optimization, and only used for evaluation once the final model and pipeline had been preregistered. The developed model achieved moderate predictive accuracy for clinician-assessed GMA scores (area under the receiver operating characteristic curve, ROC-AUC = 0.77; area under the precision-recall curve, PR-AUC = 0.41). The moderate accuracy is noteworthy given the 10-12% positive class prevalence, and power-law scaling of ROC-AUC as a function of increasing dataset size. By releasing de-identified feature data and open-source code, and simplifying the training pipeline using AutoML, our work establishes essential groundwork for future robust, globally relevant CP screening tools suitable for low-resource settings.

  PublisherDOI

• PrimateFace: A Machine Learning Resource for Automated Face Analysis in Human and Non-human Primates

  bioRxiv (Cold Spring Harbor Laboratory) · 2025-08-15 · 2 citations

  preprintOpen accessSenior author

  Abstract Machine learning has revolutionized human face analysis, but equivalent tools for non-human primates remain limited and species-specific, hindering progress in neuroscience, anthropology, and conservation. Here, we present PrimateFace, a comprehensive, cross-species platform for primate facial analysis comprising a systematically curated dataset of 260,000+ images spanning over 60 genera, including a genus-balanced subset of 60,000 images, annotated with bounding boxes and facial landmark configurations. Face detection and facial landmark estimation models trained on PrimateFace achieve high cross-species performance, from tarsiers to gorillas, achieving performance comparable to baseline models trained exclusively on human data (0.34 vs. 0.39 mAP for face detection; 0.061 vs. 0.053 normalized landmark error), demonstrating the generalization benefits of cross-species training. PrimateFace enables diverse downstream applications including individual recognition, gaze analysis, and automated extraction of stereotyped (e.g., lip-smacking) and subtle (e.g., soft left turn) facial movements. PrimateFace provides a standardized platform for facial phenotyping across the primate order, empowering data-driven studies that advance the health and well-being of human and non-human primates. All models, notebooks, and data can be found at github.com/KordingLab/PrimateFace .

  PublisherOA PDFDOI

• Resting-State fMRI and the Risk of Overinterpretation: Noise, Mechanisms, and a Missing Rosetta Stone

  bioRxiv (Cold Spring Harbor Laboratory) · 2025-09-19

  preprintOpen access

  Resting-state fMRI is widely used to describe spontaneous neural activity via correlation-based synchronization measures (e.g., seed-based or network "connectivity", as well as principal and independent component analyses), yet it faces two fundamental obstacles: pervasive non-neural noise and the absence of a definitive "Rosetta Stone" linking the measured BOLD signals to underlying neural events. Although "correlation does not imply causation" has become a cliche, leveraging correlations effectively--and understanding their inferential pitfalls--remains a nuanced challenge. Correlation-based analyses are typically not able to yield causal conclusions, yet they are frequently used to underpin causal narratives in neuroscience research and especially in clinical contexts, which represent a problematic case of overinterpretation. Using causal inference reasoning, simulations and analytic methods, we address three critical questions when performing resting-state fMRI: (1) How reliable are correlation estimates for capturing cross-regional synchrony? (2) What are the consequences of inaccuracies in estimated correlations? (3) To what extent do estimated correlations reflect causal neural interactions? We identify two principal pitfalls. First, correlation estimates are systematically distorted, at both the neural and hemodynamic response levels, by diverse noise sources, including variability in neurovascular coupling, producing spurious, suppressed, or even reversed effects. Second, graph-based approaches lack causal interpretability (in the standard counterfactual definition), meaning that even large samples and strong statistical evidence may obscure fundamental ambiguities in what the correlations represent. These underlying interpretational challenges extend to other noninvasive modalities that also utilize correlation-based analyses, such as EEG, MEG, and fNIRS. In light of these challenges, we advocate for three priorities: (1) cautious interpretation that avoids causal overreach, (2) multimodal validation to cross-check findings against independent measures, and (3) enhanced methodological rigor, particularly in biomarker discovery and clinical trials, to ensure that resting-state fMRI provides meaningful insights.

  PublisherOA PDFDOI

• Individual-specific strategies inform category learning

  Scientific Reports · 2025-01-23 · 1 citations

  articleOpen access

  Categorization is an essential task for sensory perception. Individuals learn category labels using a variety of strategies to ensure that sensory signals, such as sounds or images, can be assigned to proper categories. Categories are often learned on the basis of extreme examples, and the boundary between categories can differ among individuals. The trajectories for learning also differ among individuals, as different individuals rely on different strategies, such as repeating or alternating choices. However, little is understood about the relationship between individual learning trajectories and learned categorization. To study this relationship, we trained mice to categorize auditory stimuli into two categories using a two-alternative forced choice task. Because the mice took several weeks to learn the task, we were able to quantify the time course of individual strategies and how they relate to how mice categorize stimuli around the categorization boundary. Different mice exhibited different trajectories while learning the task. Mice displayed preferences for a specific category, manifested by a choice bias in their responses, but this bias drifted with learning. We found that this drift in choice bias correlated with variability in the category boundary for sounds with ambiguous category membership. Next, we asked how stimulus-independent, individual-specific strategies informed learning. We found that the tendency to repeat choices, which is a form of perseveration, contributed to long-term learning. These results indicate that long-term trends in individual strategies during category learning affect learned category boundaries.

  PublisherOA PDFDOI

• Falcon: Fractional Alternating Cut with Overcoming Minima in Unsupervised Segmentation

  ArXiv.org · 2025-04-08

  preprintOpen accessSenior author

  Today's unsupervised image segmentation algorithms often segment suboptimally. Modern graph-cut based approaches rely on high-dimensional attention maps from Transformer-based foundation models, typically employing a relaxed Normalized Cut solved recursively via the Fiedler vector (the eigenvector of the second smallest eigenvalue). Consequently, they still lag behind supervised methods in both mask generation speed and segmentation accuracy. We present a regularized fractional alternating cut (Falcon), an optimization-based K-way Normalized Cut without relying on recursive eigenvector computations, achieving substantially improved speed and accuracy. Falcon operates in two stages: (1) a fast K-way Normalized Cut solved by extending into a fractional quadratic transformation, with an alternating iterative procedure and regularization to avoid local minima; and (2) refinement of the resulting masks using complementary low-level information, producing high-quality pixel-level segmentations. Experiments show that Falcon not only surpasses existing state-of-the-art methods by an average of 2.5% across six widely recognized benchmarks (reaching up to 4.3\% improvement on Cityscapes), but also reduces runtime by around 30% compared to prior graph-based approaches. These findings demonstrate that the semantic information within foundation-model attention can be effectively harnessed by a highly parallelizable graph cut framework. Consequently, Falcon can narrow the gap between unsupervised and supervised segmentation, enhancing scalability in real-world applications and paving the way for dense prediction-based vision pre-training in various downstream tasks. The code is released in https://github.com/KordingLab/Falcon.

  PublisherOA PDFDOI

Recent grants

• The representation of uncertainty in the sensorimotor system

  NIH · $1.6M · 2011–2018

• The role of uncertainty for motor learning and adaptation

  NIH · $30k · 2008–2017

• CRCNS: Data Sharing: A Joint Database of Experiments and Models of Reaching Movement

  NSF · $425k · 2010–2014

• Massive scale electrical neural recordings in vivo using commercial ROIC chips

  NIH · $1.9M · 2017–2019

• Sub-micrometer x-ray tomography for neuroanatomy

  NIH · $2.0M · 2015–2018

Frequent coauthors

• Joshua I. Glaser

  Northwestern University

  100 shared

• Hugo L. Fernandes

  85 shared

• Pavan Ramkumar

  50 shared

• Ian H. Stevenson

  University of Connecticut

  49 shared

• Titipat Achakulvisut

  Mahidol University

  48 shared

• Edward S. Boyden

  McGovern Institute for Brain Research

  47 shared

• Lee E. Miller

  Shirley Ryan AbilityLab

  44 shared

• Iris Vilares

  Twin Cities Orthopedics

  44 shared

Education

• PhD, physics

  Eidgenössische Technische Hochschule Zürich

  2003

Awards & honors

• Catalyzing communities of research rigour champions