Search for: All records

Large language models (LLMs) have revolution- ized machine learning due to their ability to cap- ture complex interactions between input features. Popular post-hoc explanation methods like SHAP provide marginal feature attributions, while their extensions to interaction importances only scale to small input lengths (≈20). We propose Spectral Ex- plainer (SPEX), a model-agnostic interaction attri- bution algorithm that efficiently scales to large input lengths (≈1000). SPEX exploits underlying nat- ural sparsity among interactions—common in real- world data—and applies a sparse Fourier transform using a channel decoding algorithm to efficiently identify important interactions. We perform exper- iments across three difficult long-context datasets that require LLMs to utilize interactions between inputs to complete the task. For large inputs, SPEX outperforms marginal attribution methods by up to 20% in terms of faithfully reconstructing LLM out- puts. Further, SPEX successfully identifies key fea- tures and interactions that strongly influence model output. For one of our datasets, HotpotQA, SPEX provides interactions that align with human annota- tions. Finally, we use our model-agnostic approach to generate explanations to demonstrate abstract rea- soning in closed-source LLMs (GPT-4o mini) and compositional reasoning in vision-language models. 

Background:The limited diagnostic accuracy of prostate-specific antigen screening for prostate cancer (PCa) has prompted innovative solutions, such as the state-of-the-art 18-gene urine test for clinically-significant PCa (MyProstateScore2.0 (MPS2)).Objective:We aim to develop a non-invasive biomarker test, the simplified MPS2 (sMPS2), which achieves similar state-of-the-art accuracy as MPS2 for predicting high-grade PCa but requires substantially fewer genes than the 18-gene MPS2 to improve its accessibility for routine clinical care.Methods:We grounded the development of sMPS2 in the Predictability, Computability, and Stability (PCS) framework for veridical data science. Under this framework, we stress-tested the development of sMPS2 across various data preprocessing and modeling choices and developed a stability-driven PCS ranking procedure for selecting the most predictive and robust genes for use in sMPS2.Results:The final sMPS2 model consisted of 7 genes and achieved a 0.784 AUROC (95% confidence interval, 0.742–0.825) for predicting high-grade PCa on a blinded external validation cohort. This is only 2.3% lower than the 18-gene MPS2, which is similar in magnitude to the 1–2% in uncertainty induced by different data preprocessing choices.Conclusions:The 7-gene sMPS2 provides a unique opportunity to expand the reach and adoption of non-invasive PCa screening. 

Concept bottleneck models (CBM) aim to improve model interpretability by predicting human level “concepts” in a bottleneck within a deep learning model architecture. However, how the predicted concepts are used in predicting the target still either remains black-box or is simplified to maintain interpretability at the cost of prediction performance. We propose to use Fast Interpretable Greedy Sum- Trees (FIGS) to obtain Binary Distillation (BD). This new method, called FIGSBD, distills a binary-augmented concept-to-target portion of the CBM into an interpretable tree-based model, while maintaining the competitive prediction performance of the CBM teacher. FIGS-BD can be used in downstream tasks to explain and decompose CBM predictions into interpretable binary-concept-interaction attributions and guide adaptive test-time intervention. Across 4 datasets, we demonstrate that our adaptive test-time intervention identifies key concepts that significantly improve performance for realistic human-in-the-loop settings that only allow for limited concept interventions. All code is made available on Github (https://github.com/mattyshen/adaptiveTTI). 

Automated mechanistic interpretation research has attracted great interest due to its potential to scale explanations of neural network internals to large models. Existing automated circuit discovery work relies on activation patching or its ap- proximations to identify subgraphs in models for specific tasks (circuits). They often suffer from slow runtime, approximation errors, and specific requirements of metrics, such as non-zero gradients. In this work, we introduce contextual decomposition for transformers (CD-T) to build interpretable circuits in large lan- guage models. CD-T can produce circuits at any level of abstraction and is the first to efficiently produce circuits as fine-grained as attention heads at specific sequence positions. CD-T is compatible to all transformer types, and requires no training or manually-crafted examples. CD-T consists of a set of mathematical equations to isolate contribution of model features. Through recursively comput- ing contribution of all nodes in a computational graph of a model using CD-T followed by pruning, we are able to reduce circuit discovery runtime from hours to seconds compared to state-of-the-art baselines. On three standard circuit eval- uation datasets (indirect object identification, greater-than comparisons, and doc- string completion), we demonstrate that CD-T outperforms ACDC and EAP by better recovering the manual circuits with an average of 97% ROC AUC under low runtimes. In addition, we provide evidence that faithfulness of CD-T circuits is not due to random chance by showing our circuits are 80% more faithful than random circuits of up to 60% of the original model size. Finally, we show CD-T circuits are able to perfectly replicate original models’ behavior (faithfulness = 1) using fewer nodes than the baselines for all tasks. Our results underscore the great promise of CD-T for efficient automated mechanistic interpretability, paving the way for new insights into the workings of large language models. All code for using CD-T and reproducing results is made available on Github (https://github.com/adelaidehsu/CD_Circuit).