Transformers use the dense self-attention mechanism which gives a lot of flexibility for long-range connectivity. Over multiple layers of a deep transformer, the number of possible connectivity patterns increases exponentially. However, very few of these contribute to the performance of the network, and even fewer are essential. We hypothesize that there are sparsely connected sub-networks within a transformer, called information pathways which can be trained independently. However, the dynamic (i.e., input-dependent) nature of these pathways makes it difficult to prune dense self-attention during training. But the overall distribution of these pathways is often predictable. We take advantage of this fact to propose Stochastically Subsampled self-Attention (SSA) - a general-purpose training strategy for transformers that can reduce both the memory and computational cost of self-attention by 4 to 8 times during training while also serving as a regularization method - improving generalization over dense training. We show that an ensemble of sub-models can be formed from the subsampled pathways within a network, which can achieve better performance than its densely attended counterpart. We perform experiments on a variety of NLP, computer vision and graph learning tasks in both generative and discriminative settings to provide empirical evidence for our claims and show the effectiveness of the proposed method.
Large Language Models (LLMs) have emerged as powerful tools capable of accomplishing a broad spectrum of tasks. Their abilities span numerous areas, and one area where they have made a significant impact is in the domain of code generation. Here, we propose using the coding abilities of LLMs to introduce meaningful variations to code defining neural networks. Meanwhile, Quality-Diversity (QD) algorithms are known to discover diverse and robust solutions. By merging the code-generating abilities of LLMs with the diversity and robustness of QD solutions, we introduce \texttt{LLMatic}, a Neural Architecture Search (NAS) algorithm. While LLMs struggle to conduct NAS directly through prompts, \texttt{LLMatic} uses a procedural approach, leveraging QD for prompts and network architecture to create diverse and high-performing networks. We test \texttt{LLMatic} on the CIFAR-10 and NAS-bench-201 benchmarks, demonstrating that it can produce competitive networks while evaluating just $2,000$ candidates, even without prior knowledge of the benchmark domain or exposure to any previous top-performing models for the benchmark. The open-sourced code is available in \url{https://github.com/umair-nasir14/LLMatic}.
Chest radiography or called Chest X-ray(CXR) is common and one of the most cost-effective diagnostic procedures done in medical facilities. Interpretation of the chest X-ray is important in the diagnosis and detection of diseases such as pneumonia, pneumothorax, interstitial lung disease, heart failure, bone fracture, tuberculosis, pneumoconiosis, COVID-19, and even early lung cancer, as it contains a lot of information about the patient's medical condition. In practice, radiologists or consultants review these images and diagnose diseases, but this process can lead to misdiagnoses due to human error and expertise. Also, interpretation takes time. Assistive systems can be developed to reduce the impact of events that challenge health systems, such as the Covid-19 pandemic, to alleviate the workload and to reduce human error. For this purpose, chest X-ray images of a total of 336, Covid-19, Infiltrative and healthy patients collected from Derince Education and Research Hospital were collected. In this study, experiments were conducted to classify these images with Convolutional Neural Networks. Our CNN network with Residual Blocks and CNN with ResNet backbone by transfer learning were compared. As a result, the CNN model achieved 86.76% accuracy on test data, and (K=5) K-fold cross-validation 84.84% accuracy. The CNN with ResNet backbone by transfer learning achieved 89.71% accuracy on test data, and (K=5) K-fold cross-validation 84.23% accuracy.
Despite the impressive generalization capabilities of deep neural networks, they have been repeatedly shown to be overconfident when they are wrong. Fixing this issue is known as model calibration, and has consequently received much attention in the form of modified training schemes and post-training calibration procedures such as temperature scaling. While temperature scaling is frequently used because of its simplicity, it is often outperformed by modified training schemes. In this work, we identify a specific bottleneck for the performance of temperature scaling. We show that for empirical risk minimizers for a general set of distributions in which the supports of classes have overlaps, the performance of temperature scaling degrades with the amount of overlap between classes, and asymptotically becomes no better than random when there are a large number of classes. On the other hand, we prove that optimizing a modified form of the empirical risk induced by the Mixup data augmentation technique can in fact lead to reasonably good calibration performance, showing that training-time calibration may be necessary in some situations. We also verify that our theoretical results reflect practice by showing that Mixup significantly outperforms empirical risk minimization (with respect to multiple calibration metrics) on image classification benchmarks with class overlaps introduced in the form of label noise.
Modern deep learning approaches usually utilize modality-specific processing. For example, the most common deep learning approach to image classification involves decoding image file bytes into an RGB tensor which is passed into a neural network. Instead, we investigate modality-independent representation learning by performing classification directly on file bytes, without the need for decoding files at inference time. This enables models to operate on various modalities without any hand-designed, modality-specific processing. Our model, ByteFormer, improves ImageNet Top-1 classification accuracy by $5\%$ (from $72.2\%$ to $77.33\%$) relative to DeIT models of similar size. Compared to Perceiver IO, our model requires absolutely no modality-specific processing at inference time, and uses an order of magnitude fewer parameters at equivalent accuracy on ImageNet. We demonstrate that the same ByteFormer architecture can perform audio classification without modifications or modality-specific preprocessing. We achieve $95.42\%$ classification accuracy on the Speech Commands V2 dataset (comparable to the state-of-the-art accuracy of $98.7\%$). Additionally, we demonstrate that ByteFormer can operate jointly on images and audio, handling joint classification without explicit knowledge of the input modality. We release our code at https://github.com/apple/corenet/tree/main/projects/byteformer.
Modern hierarchical vision transformers have added several vision-specific components in the pursuit of supervised classification performance. While these components lead to effective accuracies and attractive FLOP counts, the added complexity actually makes these transformers slower than their vanilla ViT counterparts. In this paper, we argue that this additional bulk is unnecessary. By pretraining with a strong visual pretext task (MAE), we can strip out all the bells-and-whistles from a state-of-the-art multi-stage vision transformer without losing accuracy. In the process, we create Hiera, an extremely simple hierarchical vision transformer that is more accurate than previous models while being significantly faster both at inference and during training. We evaluate Hiera on a variety of tasks for image and video recognition. Our code and models are available at https://github.com/facebookresearch/hiera.
The development of federated learning (FL) methods, which aim to learn from distributed databases (i.e., clients) without accessing data on clients, has recently attracted great attention. Most of these methods assume that the clients are associated with the same data modality. However, remote sensing (RS) images in different clients can be associated with different data modalities that can improve the classification performance when jointly used. To address this problem, in this paper we introduce a novel multi-modal FL framework that aims to learn from decentralized multi-modal RS image archives for RS image classification problems. The proposed framework is made up of three modules: 1) multi-modal fusion (MF); 2) feature whitening (FW); and 3) mutual information maximization (MIM). The MF module performs iterative model averaging to learn without accessing data on clients in the case that clients are associated with different data modalities. The FW module aligns the representations learned among the different clients. The MIM module maximizes the similarity of images from different modalities. Experimental results show the effectiveness of the proposed framework compared to iterative model averaging, which is a widely used algorithm in FL. The code of the proposed framework is publicly available at https://git.tu-berlin.de/rsim/MM-FL.
The recent upsurge in pre-trained large models (e.g. GPT-4) has swept across the entire deep learning community. Such powerful large language models (LLMs) demonstrate advanced generative ability and multimodal understanding capability, which quickly achieve new state-of-the-art performances on a variety of benchmarks. The pre-trained LLM usually plays the role as a universal AI model that can conduct various tasks, including context reasoning, article analysis and image content comprehension. However, considering the prohibitively high memory and computational cost for implementing such a large model, the conventional models (such as CNN and ViT), are still essential for many visual perception tasks. In this paper, we propose to enhance the representation ability of ordinary vision models for perception tasks (e.g. image classification) by taking advantage of large pre-trained models. We present a new learning paradigm in which the knowledge extracted from large pre-trained models are utilized to help models like CNN and ViT learn enhanced representations and achieve better performance. Firstly, we curate a high quality description set by prompting a multimodal LLM to generate descriptive text for all training images. Furthermore, we feed these detailed descriptions into a pre-trained encoder to extract text embeddings with rich semantic information that encodes the content of images. During training, text embeddings will serve as extra supervising signals and be aligned with image representations learned by vision models. The alignment process helps vision models learn better and achieve higher accuracy with the assistance of pre-trained LLMs. We conduct extensive experiments to verify that the proposed algorithm consistently improves the performance for various vision models with heterogeneous architectures.
In quality control, microstructures are investigated rigorously to ensure structural integrity, exclude the presence of critical volume defects, and validate the formation of the target microstructure. For quenched, hierarchically-structured steels, the morphology of the bainitic and martensitic microstructures are of major concern to guarantee the reliability of the material under service conditions. Therefore, industries conduct small sample-size inspections of materials cross-sections through metallographers to validate the needle morphology of such microstructures. We demonstrate round-robin test results revealing that this visual grading is afflicted by pronounced subjectivity despite the thorough training of personnel. Instead, we propose a deep learning image classification approach that distinguishes steels based on their microstructure type and classifies their needle length alluding to the ISO 643 grain size assessment standard. This classification approach facilitates the reliable, objective, and automated classification of hierarchically structured steels. Specifically, an accuracy of 96% and roughly 91% is attained for the distinction of martensite/bainite subtypes and needle length, respectively. This is achieved on an image dataset that contains significant variance and labeling noise as it is acquired over more than ten years from multiple plants, alloys, etchant applications, and light optical microscopes by many metallographers (raters). Interpretability analysis gives insights into the decision-making of these models and allows for estimating their generalization capability.
Continual learning (CL) is an important technique to allow artificial neural networks to work in open environments. CL enables a system to learn new tasks without severe interference to its performance on old tasks, i.e., overcome the problems of catastrophic forgetting. In joint learning, it is well known that the out-of-distribution (OOD) problem caused by intentional attacks or environmental perturbations will severely impair the ability of networks to generalize. In this work, we reported a special form of catastrophic forgetting raised by the OOD problem in continual learning settings, and we named it out-of-distribution forgetting (OODF). In continual image classification tasks, we found that for a given category, introducing an intra-class distribution shift significantly impaired the recognition accuracy of CL methods for that category during subsequent learning. Interestingly, this phenomenon is special for CL as the same level of distribution shift had only negligible effects in the joint learning scenario. We verified that CL methods without dedicating subnetworks for individual tasks are all vulnerable to OODF. Moreover, OODF does not depend on any specific way of shifting the distribution, suggesting it is a risk for CL in a wide range of circumstances. Taken together, our work identified an under-attended risk during CL, highlighting the importance of developing approaches that can overcome OODF. Code available: \url{https://github.com/Hiroid/OODF}
Post-training quantization (PTQ) has been gaining popularity for the deployment of deep neural networks on resource-limited devices since unlike quantization-aware training, neither a full training dataset nor end-to-end training is required at all. As PTQ schemes based on reconstructing each layer or block output turn out to be effective to enhance quantized model performance, recent works have developed algorithms to devise and learn a new weight-rounding scheme so as to better reconstruct each layer or block output. In this work, we propose a simple yet effective new weight-rounding mechanism for PTQ, coined \emph{FlexRound}, based on element-wise division instead of typical element-wise addition such that FlexRound enables jointly learning a common quantization grid size as well as a different scale for each pre-trained weight. Thanks to the reciprocal rule of derivatives induced by element-wise division, FlexRound is inherently able to exploit pre-trained weights when updating their corresponding scales, and thus, flexibly quantize pre-trained weights depending on their magnitudes. We empirically validate the efficacy of FlexRound on a wide range of models and tasks. To the best of our knowledge, our work is the first to carry out comprehensive experiments on not only image classification and natural language understanding but also natural language generation. Moreover, we demonstrate, for the first time, that large language models can be efficiently quantized, with only a negligible impact on performance compared to half-precision baselines, achieved by reconstructing the output in a block-by-block manner. Our code is available at \url{https://github.com/onliwad101/FlexRound_LRQ}.
Deep learning models have been used in creating various effective image classification applications. However, they are vulnerable to adversarial attacks that seek to misguide the models into predicting incorrect classes. Our study of major adversarial attack models shows that they all specifically target and exploit the neural networking structures in their designs. This understanding makes us develop a hypothesis that most classical machine learning models, such as Random Forest (RF), are immune to adversarial attack models because they do not rely on neural network design at all. Our experimental study of classical machine learning models against popular adversarial attacks supports this hypothesis. Based on this hypothesis, we propose a new adversarial-aware deep learning system by using a classical machine learning model as the secondary verification system to complement the primary deep learning model in image classification. Although the secondary classical machine learning model has less accurate output, it is only used for verification purposes, which does not impact the output accuracy of the primary deep learning model, and at the same time, can effectively detect an adversarial attack when a clear mismatch occurs. Our experiments based on CIFAR-100 dataset show that our proposed approach outperforms current state-of-the-art adversarial defense systems.
Neural architecture search (NAS) has allowed for the automatic creation of new and effective neural network architectures, offering an alternative to the laborious process of manually designing complex architectures. However, traditional NAS algorithms are slow and require immense amounts of computing power. Recent research has investigated training-free NAS metrics for image classification architectures, drastically speeding up search algorithms. In this paper, we investigate training-free NAS metrics for recurrent neural network (RNN) and BERT-based transformer architectures, targeted towards language modeling tasks. First, we develop a new training-free metric, named hidden covariance, that predicts the trained performance of an RNN architecture and significantly outperforms existing training-free metrics. We experimentally evaluate the effectiveness of the hidden covariance metric on the NAS-Bench-NLP benchmark. Second, we find that the current search space paradigm for transformer architectures is not optimized for training-free neural architecture search. Instead, a simple qualitative analysis can effectively shrink the search space to the best performing architectures. This conclusion is based on our investigation of existing training-free metrics and new metrics developed from recent transformer pruning literature, evaluated on our own benchmark of trained BERT architectures. Ultimately, our analysis shows that the architecture search space and the training-free metric must be developed together in order to achieve effective results.
Self-training is an important technique for solving semi-supervised learning problems. It leverages unlabeled data by generating pseudo-labels and combining them with a limited labeled dataset for training. The effectiveness of self-training heavily relies on the accuracy of these pseudo-labels. In this paper, we introduce doubly robust self-training, a novel semi-supervised algorithm that provably balances between two extremes. When the pseudo-labels are entirely incorrect, our method reduces to a training process solely using labeled data. Conversely, when the pseudo-labels are completely accurate, our method transforms into a training process utilizing all pseudo-labeled data and labeled data, thus increasing the effective sample size. Through empirical evaluations on both the ImageNet dataset for image classification and the nuScenes autonomous driving dataset for 3D object detection, we demonstrate the superiority of the doubly robust loss over the standard self-training baseline.
Spiking neural networks (SNNs) have shown advantages in computation and energy efficiency over traditional artificial neural networks (ANNs) thanks to their event-driven representations. SNNs also replace weight multiplications in ANNs with additions, which are more energy-efficient and less computationally intensive. However, it remains a challenge to train deep SNNs due to the discrete spike function. A popular approach to circumvent this challenge is ANN-to-SNN conversion. However, due to the quantization error and accumulating error, it often requires lots of time steps (high inference latency) to achieve high performance, which negates SNN's advantages. To this end, this paper proposes Fast-SNN that achieves high performance with low latency. We demonstrate the equivalent mapping between temporal quantization in SNNs and spatial quantization in ANNs, based on which the minimization of the quantization error is transferred to quantized ANN training. With the minimization of the quantization error, we show that the sequential error is the primary cause of the accumulating error, which is addressed by introducing a signed IF neuron model and a layer-wise fine-tuning mechanism. Our method achieves state-of-the-art performance and low latency on various computer vision tasks, including image classification, object detection, and semantic segmentation. Codes are available at: https://github.com/yangfan-hu/Fast-SNN.
Tensor networks developed in the context of condensed matter physics try to approximate order-$N$ tensors with a reduced number of degrees of freedom that is only polynomial in $N$ and arranged as a network of partially contracted smaller tensors. As we have recently demonstrated in the context of quantum many-body physics, computation costs can be further substantially reduced by imposing constraints on the canonical polyadic (CP) rank of the tensors in such networks [arXiv:2205.15296]. Here, we demonstrate how tree tensor networks (TTN) with CP rank constraints and tensor dropout can be used in machine learning. The approach is found to outperform other tensor-network-based methods in Fashion-MNIST image classification. A low-rank TTN classifier with branching ratio $b=4$ reaches a test set accuracy of 90.3\% with low computation costs. Consisting of mostly linear elements, tensor network classifiers avoid the vanishing gradient problem of deep neural networks. The CP rank constraints have additional advantages: The number of parameters can be decreased and tuned more freely to control overfitting, improve generalization properties, and reduce computation costs. They allow us to employ trees with large branching ratios, substantially improving the representation power.
Computer aided detection and diagnosis systems based on deep learning have shown promising performance in breast cancer detection. However, there are cases where the obtained results lack justification. In this study, our objective is to highlight the regions of interest used by a convolutional neural network (CNN) for classifying histological images as benign or malignant. We compare these regions with the regions identified by pathologists. To achieve this, we employed the VGG19 architecture and tested three visualization methods: Gradient, LRP Z, and LRP Epsilon. Additionally, we experimented with three pixel selection methods: Bins, K-means, and MeanShift. Based on the results obtained, the Gradient visualization method and the MeanShift selection method yielded satisfactory outcomes for visualizing the images.
Deep neural networks are widely known for their remarkable effectiveness across various tasks, with the consensus that deeper networks implicitly learn more complex data representations. This paper shows that sufficiently deep networks trained for supervised image classification split into two distinct parts that contribute to the resulting data representations differently. The initial layers create linearly-separable representations, while the subsequent layers, which we refer to as \textit{the tunnel}, compress these representations and have a minimal impact on the overall performance. We explore the tunnel's behavior through comprehensive empirical studies, highlighting that it emerges early in the training process. Its depth depends on the relation between the network's capacity and task complexity. Furthermore, we show that the tunnel degrades out-of-distribution generalization and discuss its implications for continual learning.
Deep neural networks have been demonstrated to achieve phenomenal success in many domains, and yet their inner mechanisms are not well understood. In this paper, we investigate the curvature of image manifolds, i.e., the manifold deviation from being flat in its principal directions. We find that state-of-the-art trained convolutional neural networks for image classification have a characteristic curvature profile along layers: an initial steep increase, followed by a long phase of a plateau, and followed by another increase. In contrast, this behavior does not appear in untrained networks in which the curvature flattens. We also show that the curvature gap between the last two layers has a strong correlation with the generalization capability of the network. Moreover, we find that the intrinsic dimension of latent codes is not necessarily indicative of curvature. Finally, we observe that common regularization methods such as mixup yield flatter representations when compared to other methods. Our experiments show consistent results over a variety of deep learning architectures and multiple data sets. Our code is publicly available at https://github.com/azencot-group/CRLM
Research in child development has shown that embodied experience handling physical objects contributes to many cognitive abilities, including visual learning. One characteristic of such experience is that the learner sees the same object from several different viewpoints. In this paper, we study how learning signals that equate different viewpoints -- e.g., assigning similar representations to different views of a single object -- can support robust visual learning. We use the Toybox dataset, which contains egocentric videos of humans manipulating different objects, and conduct experiments using a computer vision framework for self-supervised contrastive learning. We find that representations learned by equating different physical viewpoints of an object benefit downstream image classification accuracy. Further experiments show that this performance improvement is robust to variations in the gaps between viewpoints, and that the benefits transfer to several different image classification tasks.