What Do VLMs NOTICE? A Mechanistic Interpretability Pipeline for Noise-free Text-Image Corruption and Evaluation

Vision-Language Models (VLMs) have gained community-spanning prominence due to their ability to integrate visual and textual inputs to perform complex tasks. Despite their success, the internal decision-making processes of these models remain opaque, posing challenges in high-stakes applications. To address this, we introduce NOTICE, the first Noise-free Text-Image Corruption and Evaluation pipeline for mechanistic interpretability in VLMs. NOTICE incorporates a Semantic Minimal Pairs (SMP) framework for image corruption and Symmetric Token Replacement (STR) for text. This approach enables semantically meaningful causal mediation analysis for both modalities, providing a robust method for analyzing multimodal integration within models like BLIP. Our experiments on the SVO-Probes, MIT-States, and Facial Expression Recognition datasets reveal crucial insights into VLM decision-making, identifying the significant role of middle-layer cross-attention heads. Further, we uncover a set of ``universal cross-attention heads'' that consistently contribute across tasks and modalities, each performing distinct functions such as implicit image segmentation, object inhibition, and outlier inhibition. This work paves the way for more transparent and interpretable multimodal systems.

Retrieval Augmented Zero-Shot Text Classification

Zero-shot text learning enables text classifiers to handle unseen classes efficiently, alleviating the need for task-specific training data. A simple approach often relies on comparing embeddings of query (text) to those of potential classes. However, the embeddings of a simple query sometimes lack rich contextual information, which hinders the classification performance. Traditionally, this has been addressed by improving the embedding model with expensive training. We introduce QZero, a novel training-free knowledge augmentation approach that reformulates queries by retrieving supporting categories from Wikipedia to improve zero-shot text classification performance. Our experiments across six diverse datasets demonstrate that QZero enhances performance for state-of-the-art static and contextual embedding models without the need for retraining. Notably, in News and medical topic classification tasks, QZero improves the performance of even the largest OpenAI embedding model by at least 5% and 3%, respectively. Acting as a knowledge amplifier, QZero enables small word embedding models to achieve performance levels comparable to those of larger contextual models, offering the potential for significant computational savings. Additionally, QZero offers meaningful insights that illuminate query context and verify topic relevance, aiding in understanding model predictions. Overall, QZero improves embedding-based zero-shot classifiers while maintaining their simplicity. This makes it particularly valuable for resource-constrained environments and domains with constantly evolving information.

On the Effect of Image Resolution on Semantic Segmentation

High-resolution semantic segmentation requires substantial computational resources. Traditional approaches in the field typically downscale the input images before processing and then upscale the low-resolution outputs back to their original dimensions. While this strategy effectively identifies broad regions, it often misses finer details. In this study, we demonstrate that a streamlined model capable of directly producing high-resolution segmentations can match the performance of more complex systems that generate lower-resolution results. By simplifying the network architecture, we enable the processing of images at their native resolution. Our approach leverages a bottom-up information propagation technique across various scales, which we have empirically shown to enhance segmentation accuracy. We have rigorously tested our method using leading-edge semantic segmentation datasets. Specifically, for the Cityscapes dataset, we further boost accuracy by applying the Noisy Student Training technique.

Revisiting invariances and introducing priors in Gromov-Wasserstein distances

Gromov-Wasserstein distance has found many applications in machine learning due to its ability to compare measures across metric spaces and its invariance to isometric transformations. However, in certain applications, this invariance property can be too flexible, thus undesirable. Moreover, the Gromov-Wasserstein distance solely considers pairwise sample similarities in input datasets, disregarding the raw feature representations. We propose a new optimal transport-based distance, called Augmented Gromov-Wasserstein, that allows for some control over the level of rigidity to transformations. It also incorporates feature alignments, enabling us to better leverage prior knowledge on the input data for improved performance. We present theoretical insights into the proposed metric. We then demonstrate its usefulness for single-cell multi-omic alignment tasks and a transfer learning scenario in machine learning.

One-Versus-Others Attention: Scalable Multimodal Integration

Multimodal learning models have become increasingly important as they surpass single-modality approaches on diverse tasks ranging from question-answering to autonomous driving. Despite the importance of multimodal learning, existing efforts focus on NLP applications, where the number of modalities is typically less than four (audio, video, text, images). However, data inputs in other domains, such as the medical field, may include X-rays, PET scans, MRIs, genetic screening, clinical notes, and more, creating a need for both efficient and accurate information fusion. Many state-of-the-art models rely on pairwise cross-modal attention, which does not scale well for applications with more than three modalities. For $n$ modalities, computing attention will result in $n \choose 2$ operations, potentially requiring considerable amounts of computational resources. To address this, we propose a new domain-neutral attention mechanism, One-Versus-Others (OvO) attention, that scales linearly with the number of modalities and requires only $n$ attention operations, thus offering a significant reduction in computational complexity compared to existing cross-modal attention algorithms. Using three diverse real-world datasets as well as an additional simulation experiment, we show that our method improves performance compared to popular fusion techniques while decreasing computation costs.

Multimodal Attention-based Deep Learning for Alzheimer's Disease Diagnosis

Alzheimer's Disease (AD) is the most common neurodegenerative disorder with one of the most complex pathogeneses, making effective and clinically actionable decision support difficult. The objective of this study was to develop a novel multimodal deep learning framework to aid medical professionals in AD diagnosis. We present a Multimodal Alzheimer's Disease Diagnosis framework (MADDi) to accurately detect the presence of AD and mild cognitive impairment (MCI) from imaging, genetic, and clinical data. MADDi is novel in that we use cross-modal attention, which captures interactions between modalities - a method not previously explored in this domain. We perform multi-class classification, a challenging task considering the strong similarities between MCI and AD. We compare with previous state-of-the-art models, evaluate the importance of attention, and examine the contribution of each modality to the model's performance. MADDi classifies MCI, AD, and controls with 96.88% accuracy on a held-out test set. When examining the contribution of different attention schemes, we found that the combination of cross-modal attention with self-attention performed the best, and no attention layers in the model performed the worst, with a 7.9% difference in F1-Scores. Our experiments underlined the importance of structured clinical data to help machine learning models contextualize and interpret the remaining modalities. Extensive ablation studies showed that any multimodal mixture of input features without access to structured clinical information suffered marked performance losses. This study demonstrates the merit of combining multiple input modalities via cross-modal attention to deliver highly accurate AD diagnostic decision support.

Unbalanced CO-Optimal Transport

Optimal transport (OT) compares probability distributions by computing a meaningful alignment between their samples. CO-optimal transport (COOT) takes this comparison further by inferring an alignment between features as well. While this approach leads to better alignments and generalizes both OT and Gromov-Wasserstein distances, we provide a theoretical result showing that it is sensitive to outliers that are omnipresent in real-world data. This prompts us to propose unbalanced COOT for which we provably show its robustness to noise in the compared datasets. To the best of our knowledge, this is the first such result for OT methods in incomparable spaces. With this result in hand, we provide empirical evidence of this robustness for the challenging tasks of heterogeneous domain adaptation with and without varying proportions of classes and simultaneous alignment of samples and features across single-cell measurements.

A Joint Graph and Image Convolution Network for Automatic Brain Tumor Segmentation

We present a joint graph convolution-image convolution neural network as our submission to the Brain Tumor Segmentation (BraTS) 2021 challenge. We model each brain as a graph composed of distinct image regions, which is initially segmented by a graph neural network (GNN). Subsequently, the tumorous volume identified by the GNN is further refined by a simple (voxel) convolutional neural network (CNN), which produces the final segmentation. This approach captures both global brain feature interactions via the graphical representation and local image details through the use of convolutional filters. We find that the GNN component by itself can effectively identify and segment the brain tumors. The addition of the CNN further improves the median performance of the model by 2 percent across all metrics evaluated. On the validation set, our joint GNN-CNN model achieves mean Dice scores of 0.89, 0.81, 0.73 and mean Hausdorff distances (95th percentile) of 6.8, 12.6, 28.2mm on the whole tumor, core tumor, and enhancing tumor, respectively.

DeepDiff: Deep-learning for predicting Differential gene expression from histone modifications

Computational methods that predict differential gene expression from histone modification signals are highly desirable for understanding how histone modifications control the functional heterogeneity of cells through influencing differential gene regulation. Recent studies either failed to capture combinatorial effects on differential prediction or primarily only focused on cell type-specific analysis. In this paper, we develop a novel attention-based deep learning architecture, DeepDiff, that provides a unified and end-to-end solution to model and to interpret how dependencies among histone modifications control the differential patterns of gene regulation. DeepDiff uses a hierarchy of multiple Long short-term memory (LSTM) modules to encode the spatial structure of input signals and to model how various histone modifications cooperate automatically. We introduce and train two levels of attention jointly with the target prediction, enabling DeepDiff to attend differentially to relevant modifications and to locate important genome positions for each modification. Additionally, DeepDiff introduces a novel deep-learning based multi-task formulation to use the cell-type-specific gene expression predictions as auxiliary tasks, encouraging richer feature embeddings in our primary task of differential expression prediction. Using data from Roadmap Epigenomics Project (REMC) for ten different pairs of cell types, we show that DeepDiff significantly outperforms the state-of-the-art baselines for differential gene expression prediction. The learned attention weights are validated by observations from previous studies about how epigenetic mechanisms connect to differential gene expression. Codes and results are available at \url{deepchrome.org}

Prototype Matching Networks for Large-Scale Multi-label Genomic Sequence Classification

One of the fundamental tasks in understanding genomics is the problem of predicting Transcription Factor Binding Sites (TFBSs). With more than hundreds of Transcription Factors (TFs) as labels, genomic-sequence based TFBS prediction is a challenging multi-label classification task. There are two major biological mechanisms for TF binding: (1) sequence-specific binding patterns on genomes known as "motifs" and (2) interactions among TFs known as co-binding effects. In this paper, we propose a novel deep architecture, the Prototype Matching Network (PMN) to mimic the TF binding mechanisms. Our PMN model automatically extracts prototypes ("motif"-like features) for each TF through a novel prototype-matching loss. Borrowing ideas from few-shot matching models, we use the notion of support set of prototypes and an LSTM to learn how TFs interact and bind to genomic sequences. On a reference TFBS dataset with $2.1$ $million$ genomic sequences, PMN significantly outperforms baselines and validates our design choices empirically. To our knowledge, this is the first deep learning architecture that introduces prototype learning and considers TF-TF interactions for large-scale TFBS prediction. Not only is the proposed architecture accurate, but it also models the underlying biology.