Chandra Sripada

Ruixuan Deng, Xiaoyang Hu, Miles Gilberti et al.

We identify semantically coherent, context-consistent network components in large language models (LLMs) using coactivation of sparse autoencoder (SAE) features collected from just a handful of prompts. Focusing on concept-relation prediction tasks, we show that ablating these components for concepts (e.g., countries and words) and relations (e.g., capital city and translation language) changes model outputs in predictable ways, while amplifying these components induces counterfactual responses. Notably, composing relation and concept components yields compound counterfactual outputs. Further analysis reveals that while most concept components emerge from the very first layer, more abstract relation components are concentrated in later layers. Lastly, we show that extracted components more comprehensively capture concepts and relations than individual features while maintaining specificity. Overall, our findings suggest a modular organization of knowledge accessed through compositional operations, and advance methods for efficient, targeted LLM manipulation.

Gino Carmona-Díaz, William Jiménez-Leal, María Alejandra Grisales et al.

Analyzing texts such as open-ended responses, headlines, or social media posts is a time- and labor-intensive process highly susceptible to bias. LLMs are promising tools for text analysis, using either a predefined (top-down) or a data-driven (bottom-up) taxonomy, without sacrificing quality. Here we present a step-by-step tutorial to efficiently develop, test, and apply taxonomies for analyzing unstructured data through an iterative and collaborative process between researchers and LLMs. Using personal goals provided by participants as an example, we demonstrate how to write prompts to review datasets and generate a taxonomy of life domains, evaluate and refine the taxonomy through prompt and direct modifications, test the taxonomy and assess intercoder agreements, and apply the taxonomy to categorize an entire dataset with high intercoder reliability. We discuss the possibilities and limitations of using LLMs for text analysis.

Daiwei Zhang, Lexin Li, Chandra Sripada et al.

Delineating the associations between images and a vector of covariates is of central interest in medical imaging studies. To tackle this problem of image response regression, we propose a novel nonparametric approach in the framework of spatially varying coefficient models, where the spatially varying functions are estimated through deep neural networks. Compared to existing solutions, the proposed method explicitly accounts for spatial smoothness and subject heterogeneity, has straightforward interpretations, and is highly flexible and accurate in capturing complex association patterns. A key idea in our approach is to treat the image voxels as the effective samples, which not only alleviates the limited sample size issue that haunts the majority of medical imaging studies, but also leads to more robust and reproducible results. Focusing on a broad family of piecewise smooth functions, we establish the estimation and selection consistency, and derive the asymptotic error bounds. We demonstrate the efficacy of the method through intensive simulations, and further illustrate its advantages with analyses of two functional magnetic resonance imaging datasets.

Pascal Sturmfels, Saige Rutherford, Mike Angstadt et al.

Given the wide success of convolutional neural networks (CNNs) applied to natural images, researchers have begun to apply them to neuroimaging data. To date, however, exploration of novel CNN architectures tailored to neuroimaging data has been limited. Several recent works fail to leverage the 3D structure of the brain, instead treating the brain as a set of independent 2D slices. Approaches that do utilize 3D convolutions rely on architectures developed for object recognition tasks in natural 2D images. Such architectures make assumptions about the input that may not hold for neuroimaging. For example, existing architectures assume that patterns in the brain exhibit translation invariance. However, a pattern in the brain may have different meaning depending on where in the brain it is located. There is a need to explore novel architectures that are tailored to brain images. We present two simple modifications to existing CNN architectures based on brain image structure. Applied to the task of brain age prediction, our network achieves a mean absolute error (MAE) of 1.4 years and trains 30% faster than a CNN baseline that achieves a MAE of 1.6 years. Our results suggest that lessons learned from developing models on natural images may not directly transfer to neuroimaging tasks. Instead, there remains a large space of unexplored questions regarding model development in this area, whose answers may differ from conventional wisdom.

Takanori Watanabe, Daniel Kessler, Clayton Scott et al.

Substantial evidence indicates that major psychiatric disorders are associated with distributed neural dysconnectivity, leading to strong interest in using neuroimaging methods to accurately predict disorder status. In this work, we are specifically interested in a multivariate approach that uses features derived from whole-brain resting state functional connectomes. However, functional connectomes reside in a high dimensional space, which complicates model interpretation and introduces numerous statistical and computational challenges. Traditional feature selection techniques are used to reduce data dimensionality, but are blind to the spatial structure of the connectomes. We propose a regularization framework where the 6-D structure of the functional connectome is explicitly taken into account via the fused Lasso or the GraphNet regularizer. Our method only restricts the loss function to be convex and margin-based, allowing non-differentiable loss functions such as the hinge-loss to be used. Using the fused Lasso or GraphNet regularizer with the hinge-loss leads to a structured sparse support vector machine (SVM) with embedded feature selection. We introduce a novel efficient optimization algorithm based on the augmented Lagrangian and the classical alternating direction method, which can solve both fused Lasso and GraphNet regularized SVM with very little modification. We also demonstrate that the inner subproblems of the algorithm can be solved efficiently in analytic form by coupling the variable splitting strategy with a data augmentation scheme. Experiments on simulated data and resting state scans from a large schizophrenia dataset show that our proposed approach can identify predictive regions that are spatially contiguous in the 6-D "connectome space," offering an additional layer of interpretability that could provide new insights about various disease processes.