The growing reproducibility crisis in machine learning has brought forward a need for careful examination of research findings. This paper investigates the claims made by Lei et al. (2023) regarding their proposed method, LICO, for enhancing post-hoc interpretability techniques and improving image classification performance. LICO leverages natural language supervision from a vision-language model to enrich feature representations and guide the learning process. We conduct a comprehensive reproducibility study, employing (Wide) ResNets and established interpretability methods like Grad-CAM and RISE. We were mostly unable to reproduce the authors' results. In particular, we did not find that LICO consistently led to improved classification performance or improvements in quantitative and qualitative measures of interpretability. Thus, our findings highlight the importance of rigorous evaluation and transparent reporting in interpretability research.
Not all learnable parameters (e.g., weights) contribute equally to a neural network's decision function. In fact, entire layers' parameters can sometimes be reset to random values with little to no impact on the model's decisions. We revisit earlier studies that examined how architecture and task complexity influence this phenomenon and ask: is this phenomenon also affected by how we train the model? We conducted experimental evaluations on a diverse set of ImageNet-1k classification models to explore this, keeping the architecture and training data constant but varying the training pipeline. Our findings reveal that the training method strongly influences which layers become critical to the decision function for a given task. For example, improved training regimes and self-supervised training increase the importance of early layers while significantly under-utilizing deeper layers. In contrast, methods such as adversarial training display an opposite trend. Our preliminary results extend previous findings, offering a more nuanced understanding of the inner mechanics of neural networks. Code: https://github.com/paulgavrikov/layer_criticality
Data embeddings with CLIP and ImageBind provide powerful features for the analysis of multimedia and/or multimodal data. We assess their performance here for classification using a Gaussian Mixture models (GMMs) based layer as an alternative to the standard Softmax layer. GMMs based classifiers have recently been shown to have interesting performances as part of deep learning pipelines trained end-to-end. Our first contribution is to investigate GMM based classification performance taking advantage of the embedded spaces CLIP and ImageBind. Our second contribution is in proposing our own GMM based classifier with a lower parameters count than previously proposed. Our findings are, that in most cases, on these tested embedded spaces, one gaussian component in the GMMs is often enough for capturing each class, and we hypothesize that this may be due to the contrastive loss used for training these embedded spaces that naturally concentrates features together for each class. We also observed that ImageBind often provides better performance than CLIP for classification of image datasets even when these embedded spaces are compressed using PCA.
Contrastive Language-Image Pretraining (CLIP) performs zero-shot image classification by mapping images and textual class representation into a shared embedding space, then retrieving the class closest to the image. This work provides a new approach for interpreting CLIP models for image classification from the lens of mutual knowledge between the two modalities. Specifically, we ask: what concepts do both vision and language CLIP encoders learn in common that influence the joint embedding space, causing points to be closer or further apart? We answer this question via an approach of textual concept-based explanations, showing their effectiveness, and perform an analysis encompassing a pool of 13 CLIP models varying in architecture, size and pretraining datasets. We explore those different aspects in relation to mutual knowledge, and analyze zero-shot predictions. Our approach demonstrates an effective and human-friendly way of understanding zero-shot classification decisions with CLIP.
The rapid growth of model scale has necessitated substantial computational resources for fine-tuning. Existing approach such as Low-Rank Adaptation (LoRA) has sought to address the problem of handling the large updated parameters in full fine-tuning. However, LoRA utilize random initialization and optimization of low-rank matrices to approximate updated weights, which can result in suboptimal convergence and an accuracy gap compared to full fine-tuning. To address these issues, we propose LoLDU, a Parameter-Efficient Fine-Tuning (PEFT) approach that significantly reduces trainable parameters by 2600 times compared to regular PEFT methods while maintaining comparable performance. LoLDU leverages Lower-Diag-Upper Decomposition (LDU) to initialize low-rank matrices for faster convergence and orthogonality. We focus on optimizing the diagonal matrix for scaling transformations. To the best of our knowledge, LoLDU has the fewest parameters among all PEFT approaches. We conducted extensive experiments across 4 instruction-following datasets, 6 natural language understanding (NLU) datasets, 8 image classification datasets, and image generation datasets with multiple model types (LLaMA2, RoBERTa, ViT, and Stable Diffusion), providing a comprehensive and detailed analysis. Our open-source code can be accessed at \href{https://github.com/SKDDJ/LoLDU}{https://github.com/SKDDJ/LoLDU}.
Transformer plays a central role in many fundamental deep learning models, e.g., the ViT in computer vision and the BERT and GPT in natural language processing, whose effectiveness is mainly attributed to its multi-head attention (MHA) mechanism. In this study, we propose a simple and novel channel-wise sample permutation (CSP) operator, achieving a new structured MHA with fewer parameters and lower complexity. Given an input matrix, CSP circularly shifts the samples of different channels with various steps and then sorts grouped samples of each channel. This operator is equivalent to implicitly implementing cross-channel attention maps as permutation matrices, which achieves linear complexity and suppresses the risk of rank collapse when representing data. We replace the MHA of some representative models with CSP and test the CSP-based models in several discriminative tasks, including image classification and long sequence analysis. Experiments show that the CSP-based models achieve comparable or better performance with fewer parameters and lower computational costs than the classic Transformer and its state-of-the-art variants. The code is available at https://github.com/DaShenZi721/CSP.
The edge-device environment imposes severe resource limitations, encompassing computation costs, hardware resource usage, and energy consumption for deploying deep neural network models. Ultra-low-bit quantization and hardware accelerators have been explored as promising approaches to address these challenges. Ultra-low-bit quantization significantly reduces the model size and the computational cost. Despite progress so far, many competitive ultra-low-bit models still partially rely on float or non-ultra-low-bit quantized computation such as the input and output layer. We introduce Efficiera Residual Networks (ERNs), a model optimized for low-resource edge devices. ERNs achieve full ultra-low-bit quantization, with all weights, including the initial and output layers, being binary, and activations set at 2 bits. We introduce the shared constant scaling factor technique to enable integer-valued computation in residual connections, allowing our model to operate without float values until the final convolution layer. Demonstrating competitiveness, ERNs achieve an ImageNet top-1 accuracy of 72.5pt with a ResNet50-compatible architecture and 63.6pt with a model size less than 1MB. Moreover, ERNs exhibit impressive inference times, reaching 300FPS with the smallest model and 60FPS with the largest model on a cost-efficient FPGA device.
Vision mambas have demonstrated strong performance with linear complexity to the number of vision tokens. Their efficiency results from processing image tokens sequentially. However, most existing methods employ patch-based image tokenization and then flatten them into 1D sequences for causal processing, which ignore the intrinsic 2D structural correlations of images. It is also difficult to extract global information by sequential processing of local patches. In this paper, we propose a global image serialization method to transform the image into a sequence of causal tokens, which contain global information of the 2D image. We first convert the image from the spatial domain to the frequency domain using Discrete Cosine Transform (DCT) and then arrange the pixels with corresponding frequency ranges. We further transform each set within the same frequency band back to the spatial domain to obtain a series of images before tokenization. We construct a vision mamba model, GlobalMamba, with a causal input format based on the proposed global image serialization, which can better exploit the causal relations among image sequences. Extensive experiments demonstrate the effectiveness of our GlobalMamba, including image classification on ImageNet-1K, object detection on COCO, and semantic segmentation on ADE20K.
Parameter-efficient tuning (PET) techniques calibrate the model's predictions on downstream tasks by freezing the pre-trained models and introducing a small number of learnable parameters. However, despite the numerous PET methods proposed, their robustness has not been thoroughly investigated. In this paper, we systematically explore the robustness of four classical PET techniques (e.g., VPT, Adapter, AdaptFormer, and LoRA) under both white-box attacks and information perturbations. For white-box attack scenarios, we first analyze the performance of PET techniques using FGSM and PGD attacks. Subsequently, we further explore the transferability of adversarial samples and the impact of learnable parameter quantities on the robustness of PET methods. Under information perturbation attacks, we introduce four distinct perturbation strategies, including Patch-wise Drop, Pixel-wise Drop, Patch Shuffle, and Gaussian Noise, to comprehensively assess the robustness of these PET techniques in the presence of information loss. Via these extensive studies, we enhance the understanding of the robustness of PET methods, providing valuable insights for improving their performance in computer vision applications. The code is available at https://github.com/JCruan519/PETRobustness.
In this work, we present our solution for the MICCAI 2024 CXR-LT challenge, achieving 4th place in Subtask 2 and 5th in Subtask 1. We leveraged an ensemble of ConvNeXt V2 and MaxViT models, pretrained on an external chest X-ray dataset, to address the long-tailed distribution of chest findings. The proposed method combines state-of-the-art image classification techniques, asymmetric loss for handling class imbalance, and view-based prediction aggregation to enhance classification performance. Through experiments, we demonstrate the advantages of our approach in improving both detection accuracy and the handling of the long-tailed distribution in CXR findings. The code is available at https://github.com/yamagishi0824/cxrlt24-multiview-pp.
3D point cloud classification requires distinct models from 2D image classification due to the divergent characteristics of the respective input data. While 3D point clouds are unstructured and sparse, 2D images are structured and dense. Bridging the domain gap between these two data types is a non-trivial challenge to enable model interchangeability. Recent research using Lattice Point Classifier (LPC) highlights the feasibility of cross-domain applicability. However, the lattice projection operation in LPC generates 2D images with disconnected projected pixels. In this paper, we explore three distinct algorithms for mapping 3D point clouds into 2D images. Through extensive experiments, we thoroughly examine and analyze their performance and defense mechanisms. Leveraging current large foundation models, we scrutinize the feature disparities between regular 2D images and projected 2D images. The proposed approaches demonstrate superior accuracy and robustness against adversarial attacks. The generative model-based mapping algorithms yield regular 2D images, further minimizing the domain gap from regular 2D classification tasks. The source code is available at https://github.com/KaidongLi/pytorch-LatticePointClassifier.git.
A mechanistic understanding of how MLPs do computation in deep neural networks remains elusive. Current interpretability work can extract features from hidden activations over an input dataset but generally cannot explain how MLP weights construct features. One challenge is that element-wise nonlinearities introduce higher-order interactions and make it difficult to trace computations through the MLP layer. In this paper, we analyze bilinear MLPs, a type of Gated Linear Unit (GLU) without any element-wise nonlinearity that nevertheless achieves competitive performance. Bilinear MLPs can be fully expressed in terms of linear operations using a third-order tensor, allowing flexible analysis of the weights. Analyzing the spectra of bilinear MLP weights using eigendecomposition reveals interpretable low-rank structure across toy tasks, image classification, and language modeling. We use this understanding to craft adversarial examples, uncover overfitting, and identify small language model circuits directly from the weights alone. Our results demonstrate that bilinear layers serve as an interpretable drop-in replacement for current activation functions and that weight-based interpretability is viable for understanding deep-learning models.
Vision Transformers with various attention modules have demonstrated superior performance on vision tasks. While using sparsity-adaptive attention, such as in DAT, has yielded strong results in image classification, the key-value pairs selected by deformable points lack semantic relevance when fine-tuning for semantic segmentation tasks. The query-aware sparsity attention in BiFormer seeks to focus each query on top-k routed regions. However, during attention calculation, the selected key-value pairs are influenced by too many irrelevant queries, reducing attention on the more important ones. To address these issues, we propose the Deformable Bi-level Routing Attention (DBRA) module, which optimizes the selection of key-value pairs using agent queries and enhances the interpretability of queries in attention maps. Based on this, we introduce the Deformable Bi-level Routing Attention Transformer (DeBiFormer), a novel general-purpose vision transformer built with the DBRA module. DeBiFormer has been validated on various computer vision tasks, including image classification, object detection, and semantic segmentation, providing strong evidence of its effectiveness.Code is available at {https://github.com/maclong01/DeBiFormer}
In many critical Machine Learning applications, such as autonomous driving and medical image diagnosis, the detection of out-of-distribution (OOD) samples is as crucial as accurately classifying in-distribution (ID) inputs. Recently Outlier Exposure (OE) based methods have shown promising results in detecting OOD inputs via model fine-tuning with auxiliary outlier data. However, most of the previous OE-based approaches emphasize more on synthesizing extra outlier samples or introducing regularization to diversify OOD sample space, which is rather unquantifiable in practice. In this work, we propose a novel and straightforward method called Margin bounded Confidence Scores (MaCS) to address the nontrivial OOD detection problem by enlarging the disparity between ID and OOD scores, which in turn makes the decision boundary more compact facilitating effective segregation with a simple threshold. Specifically, we augment the learning objective of an OE regularized classifier with a supplementary constraint, which penalizes high confidence scores for OOD inputs compared to that of ID and significantly enhances the OOD detection performance while maintaining the ID classification accuracy. Extensive experiments on various benchmark datasets for image classification tasks demonstrate the effectiveness of the proposed method by significantly outperforming state-of-the-art (S.O.T.A) methods on various benchmarking metrics. The code is publicly available at https://github.com/lakpa-tamang9/margin_ood
The evolution of biological neural systems has led to both modularity and sparse coding, which enables energy efficiency and robustness across the diversity of tasks in the lifespan. In contrast, standard neural networks rely on dense, non-specialized architectures, where all model parameters are simultaneously updated to learn multiple tasks, leading to interference. Current sparse neural network approaches aim to alleviate this issue but are hindered by limitations such as 1) trainable gating functions that cause representation collapse, 2) disjoint experts that result in redundant computation and slow learning, and 3) reliance on explicit input or task IDs that limit flexibility and scalability. In this paper we propose Conditionally Overlapping Mixture of ExperTs (COMET), a general deep learning method that addresses these challenges by inducing a modular, sparse architecture with an exponential number of overlapping experts. COMET replaces the trainable gating function used in Sparse Mixture of Experts with a fixed, biologically inspired random projection applied to individual input representations. This design causes the degree of expert overlap to depend on input similarity, so that similar inputs tend to share more parameters. This results in faster learning \rev{per update step} and improved \rev{out-of-sample} generalization. We demonstrate the effectiveness of COMET on a range of tasks, including image classification, language modeling, and regression, using several popular deep learning architectures.
Medical image classification is crucial for supporting healthcare professionals in decision-making and training. While Convolutional Neural Networks (CNNs) have traditionally dominated this field, Transformer-based models are gaining attention. In this study, we apply explainable AI (XAI) techniques to assess the performance of various models on real-world medical data and identify areas for improvement. We compare CNN models such as VGG-16, ResNet-50, and EfficientNetV2L with a Transformer model: ViT-Base-16. Our results show that data augmentation has little impact, but hyperparameter tuning and advanced modeling improve performance. CNNs, particularly VGG-16 and ResNet-50, outperform ViT-Base-16 and EfficientNetV2L, likely due to underfitting from limited data. XAI methods like LIME and SHAP further reveal that better-performing models visualize tumors more effectively. These findings suggest that CNNs with shallower architectures are more effective for small datasets and can support medical decision-making.
Machine unlearning aims to selectively remove specific knowledge from a model. Current methods, such as task arithmetic, rely on fine-tuning models on the forget set, generating a task vector, and subtracting it from the original model. However, we argue the effectiveness of this approach is highly sensitive to hyperparameter selection, necessitating careful validation to identify the best model among many fine-tuned candidates. In this paper, we propose a novel method that leverages all given fine-tuned models rather than selecting a single one. By constructing task vectors from models trained with varied hyperparameters and merging only the components of the task vectors with consistent signs, we perform unlearning by negating the merged task vector from the original model. Given that existing methods also utilize multiple fine-tuned models, our approach delivers more effective unlearning without incurring additional computational costs. We demonstrate the effectiveness of our method on both vision-language models and standard image classification models, showing improved unlearning performance with minimal degradation on the retain set, outperforming state-of-the-art techniques.
Conformal prediction has recently emerged as a promising strategy for quantifying the uncertainty of a predictive model; these algorithms modify the model to output sets of labels that are guaranteed to contain the true label with high probability. However, existing conformal prediction algorithms have largely targeted classification and regression settings, where the structure of the prediction set has a simple form as a level set of the scoring function. However, for complex structured outputs such as text generation, these prediction sets might include a large number of labels and therefore be hard for users to interpret. In this paper, we propose a general framework for conformal prediction in the structured prediction setting, that modifies existing conformal prediction algorithms to output structured prediction sets that implicitly represent sets of labels. In addition, we demonstrate how our approach can be applied in domains where the prediction sets can be represented as a set of nodes in a directed acyclic graph; for instance, for hierarchical labels such as image classification, a prediction set might be a small subset of coarse labels implicitly representing the prediction set of all their more fine-descendants. We demonstrate how our algorithm can be used to construct prediction sets that satisfy a desired coverage guarantee in several domains.
Recent work developed convolutional deep kernel machines, achieving 92.7% test accuracy on CIFAR-10 using a ResNet-inspired architecture, which is SOTA for kernel methods. However, this still lags behind neural networks, which easily achieve over 94% test accuracy with similar architectures. In this work we introduce several modifications to improve the convolutional deep kernel machine's generalisation, including stochastic kernel regularisation, which adds noise to the learned Gram matrices during training. The resulting model achieves 94.5% test accuracy on CIFAR-10. This finding has important theoretical and practical implications, as it demonstrates that the ability to perform well on complex tasks like image classification is not unique to neural networks. Instead, other approaches including deep kernel methods can achieve excellent performance on such tasks, as long as they have the capacity to learn representations from data.
In this empirical article, we introduce INNAprop, an optimization algorithm that combines the INNA method with the RMSprop adaptive gradient scaling. It leverages second-order information and rescaling while keeping the memory requirements of standard DL methods as AdamW or SGD with momentum. After giving geometrical insights, we evaluate INNAprop on CIFAR-10, Food101, and ImageNet with ResNets, VGG, DenseNet, and ViT, and on GPT-2 (OpenWebText) train from scratch and with LoRA fine-tuning (E2E). INNAprop consistently matches or outperforms AdamW both in training speed and accuracy, with minimal hyperparameter tuning in large-scale settings. Our code is publicly available at \url{https://github.com/innaprop/innaprop}.