While deep learning and deep reinforcement learning (RL) systems have demonstrated impressive results in domains such as image classification, game playing, and robotic control, data efficiency remains a major challenge. Multi-task learning has emerged as a promising approach for sharing structure across multiple tasks to enable more efficient learning. However, the multi-task setting presents a number of optimization challenges, making it difficult to realize large efficiency gains compared to learning tasks independently. The reasons why multi-task learning is so challenging compared to single-task learning are not fully understood. In this work, we identify a set of three conditions of the multi-task optimization landscape that cause detrimental gradient interference, and develop a simple yet general approach for avoiding such interference between task gradients. We propose a form of gradient surgery that projects a task's gradient onto the normal plane of the gradient of any other task that has a conflicting gradient. On a series of challenging multi-task supervised and multi-task RL problems, this approach leads to substantial gains in efficiency and performance. Further, it is model-agnostic and can be combined with previously-proposed multi-task architectures for enhanced performance.
In self-supervised learning, a system is tasked with achieving a surrogate objective by defining alternative targets on a set of unlabeled data. The aim is to build useful representations that can be used in downstream tasks, without costly manual annotation. In this work, we propose a novel self-supervised formulation of relational reasoning that allows a learner to bootstrap a signal from information implicit in unlabeled data. Training a relation head to discriminate how entities relate to themselves (intra-reasoning) and other entities (inter-reasoning), results in rich and descriptive representations in the underlying neural network backbone, which can be used in downstream tasks such as classification and image retrieval. We evaluate the proposed method following a rigorous experimental procedure, using standard datasets, protocols, and backbones. Self-supervised relational reasoning outperforms the best competitor in all conditions by an average 14% in accuracy, and the most recent state-of-the-art model by 3%. We link the effectiveness of the method to the maximization of a Bernoulli log-likelihood, which can be considered as a proxy for maximizing the mutual information, resulting in a more efficient objective with respect to the commonly used contrastive losses.
The amount of manually labeled data is limited in medical applications, so semi-supervised learning and automatic labeling strategies can be an asset for training deep neural networks. However, the quality of the automatically generated labels can be uneven and inferior to manual labels. In this paper, we suggest a semi-supervised training strategy for leveraging both manually labeled data and extra unlabeled data. In contrast to the existing approaches, we apply robust loss for the automated labeled data to automatically compensate for the uneven data quality using a teacher-student framework. First, we generate pseudo-labels for unlabeled data using a teacher model pre-trained on labeled data. These pseudo-labels are noisy, and using them along with labeled data for training a deep neural network can severely degrade learned feature representations and the generalization of the network. Here we mitigate the effect of these pseudo-labels by using robust loss functions. Specifically, we use three robust loss functions, namely beta cross-entropy, symmetric cross-entropy, and generalized cross-entropy. We show that our proposed strategy improves the model performance by compensating for the uneven quality of labels in image classification as well as segmentation applications.
The correct interpretation and understanding of deep learning models are essential in many applications. Explanatory visual interpretation approaches for image, and natural language processing allow domain experts to validate and understand almost any deep learning model. However, they fall short when generalizing to arbitrary time series, which is inherently less intuitive and more diverse. Whether a visualization explains valid reasoning or captures the actual features is difficult to judge. Hence, instead of blind trust, we need an objective evaluation to obtain trustworthy quality metrics. We propose a framework of six orthogonal metrics for gradient-, propagation- or perturbation-based post-hoc visual interpretation methods for time series classification and segmentation tasks. An experimental study includes popular neural network architectures for time series and nine visual interpretation methods. We evaluate the visual interpretation methods with diverse datasets from the UCR repository and a complex, real-world dataset and study the influence of standard regularization techniques during training. We show that none of the methods consistently outperforms others on all metrics, while some are sometimes ahead. Our insights and recommendations allow experts to choose suitable visualization techniques for the model and task.
Manifold learning (ML) aims to seek low-dimensional embedding from high-dimensional data. The problem is challenging on real-world datasets, especially with under-sampling data, and we find that previous methods perform poorly in this case. Generally, ML methods first transform input data into a low-dimensional embedding space to maintain the data's geometric structure and subsequently perform downstream tasks therein. The poor local connectivity of under-sampling data in the former step and inappropriate optimization objectives in the latter step leads to two problems: structural distortion and underconstrained embedding. This paper proposes a novel ML framework named Deep Local-flatness Manifold Embedding (DLME) to solve these problems. The proposed DLME constructs semantic manifolds by data augmentation and overcomes the structural distortion problem using a smoothness constrained based on a local flatness assumption about the manifold. To overcome the underconstrained embedding problem, we design a loss and theoretically demonstrate that it leads to a more suitable embedding based on the local flatness. Experiments on three types of datasets (toy, biological, and image) for various downstream tasks (classification, clustering, and visualization) show that our proposed DLME outperforms state-of-the-art ML and contrastive learning methods.
The performance of existing single-view 3D reconstruction methods heavily relies on large-scale 3D annotations. However, such annotations are tedious and expensive to collect. Semi-supervised learning serves as an alternative way to mitigate the need for manual labels, but remains unexplored in 3D reconstruction. Inspired by the recent success of semi-supervised image classification tasks, we propose SSP3D, a semi-supervised framework for 3D reconstruction. In particular, we introduce an attention-guided prototype shape prior module for guiding realistic object reconstruction. We further introduce a discriminator-guided module to incentivize better shape generation, as well as a regularizer to tolerate noisy training samples. On the ShapeNet benchmark, the proposed approach outperforms previous supervised methods by clear margins under various labeling ratios, (i.e., 1%, 5% , 10% and 20%). Moreover, our approach also performs well when transferring to real-world Pix3D datasets under labeling ratios of 10%. We also demonstrate our method could transfer to novel categories with few novel supervised data. Experiments on the popular ShapeNet dataset show that our method outperforms the zero-shot baseline by over 12% and we also perform rigorous ablations and analysis to validate our approach.
Computationally expensive training strategies make self-supervised learning (SSL) impractical for resource constrained industrial settings. Techniques like knowledge distillation (KD), dynamic computation (DC), and pruning are often used to obtain a lightweightmodel, which usually involves multiple epochs of fine-tuning (or distilling steps) of a large pre-trained model, making it more computationally challenging. In this work we present a novel perspective on the interplay between SSL and DC paradigms. In particular, we show that it is feasible to simultaneously learn a dense and gated sub-network from scratch in a SSL setting without any additional fine-tuning or pruning steps. The co-evolution during pre-training of both dense and gated encoder offers a good accuracy-efficiency trade-off and therefore yields a generic and multi-purpose architecture for application specific industrial settings. Extensive experiments on several image classification benchmarks including CIFAR-10/100, STL-10 and ImageNet-100, demonstrate that the proposed training strategy provides a dense and corresponding gated sub-network that achieves on-par performance compared with the vanilla self-supervised setting, but at a significant reduction in computation in terms of FLOPs, under a range of target budgets (td ).
Online continual learning aims to get closer to a live learning experience by learning directly on a stream of data with temporally shifting distribution and by storing a minimum amount of data from that stream. In this empirical evaluation, we evaluate various methods from the literature that tackle online continual learning. More specifically, we focus on the class-incremental setting in the context of image classification, where the learner must learn new classes incrementally from a stream of data. We compare these methods on the Split-CIFAR100 and Split-TinyImagenet benchmarks, and measure their average accuracy, forgetting, stability, and quality of the representations, to evaluate various aspects of the algorithm at the end but also during the whole training period. We find that most methods suffer from stability and underfitting issues. However, the learned representations are comparable to i.i.d. training under the same computational budget. No clear winner emerges from the results and basic experience replay, when properly tuned and implemented, is a very strong baseline. We release our modular and extensible codebase at https://github.com/AlbinSou/ocl_survey based on the avalanche framework to reproduce our results and encourage future research.
When evaluating the performance of a pre-trained model transferred to a downstream task, it is imperative to assess not only the in-distribution (ID) accuracy of the downstream model but also its capacity to generalize and identify out-of-distribution (OOD) samples. In this paper, we unveil the hidden costs associated with intrusive fine-tuning techniques. Specifically, we demonstrate that commonly used fine-tuning methods not only distort the representations necessary for generalizing to covariate-shifted OOD samples (OOD generalization) but also distort the representations necessary for detecting semantically-shifted OOD samples (OOD detection). To address these challenges, we introduce a new model reprogramming approach for fine-tuning, which we name Reprogrammer. Reprogrammer aims to improve the holistic performance of the downstream model across ID, OOD generalization, and OOD detection tasks. Our empirical evidence reveals that Reprogrammer is less intrusive and yields superior downstream models. Furthermore, we demonstrate that by appending an additional representation residual connection to Reprogrammer, we can further preserve pre-training representations, resulting in an even more safe and robust downstream model capable of excelling in many ID classification, OOD generalization, and OOD detection settings.
Deep neural networks (DNNs) are powerful for cognitive tasks such as image classification, object detection, and scene segmentation. One drawback however is the significant high computational complexity and memory consumption, which makes them unfeasible to run real-time on embedded platforms because of the limited hardware resources. Block floating point (BFP) quantization is one of the representative compression approaches for reducing the memory and computational burden owing to their capability to effectively capture the broad data distribution of DNN models. Unfortunately, prior works on BFP-based quantization empirically choose the block size and the precision that preserve accuracy. In this paper, we develop a BFP-based bitwidth-aware analytical modeling framework (called ``BitQ'') for the best BFP implementation of DNN inference on embedded platforms. We formulate and resolve an optimization problem to identify the optimal BFP block size and bitwidth distribution by the trade-off of both accuracy and performance loss. Experimental results show that compared with an equal bitwidth setting, the BFP DNNs with optimized bitwidth allocation provide efficient computation, preserving accuracy on famous benchmarks. The source code and data are available at https://github.com/Cheliosoops/BitQ.
The platonic representation hypothesis suggests that vision and language embeddings become more homogeneous as model and dataset sizes increase. In particular, pairwise distances within each modality become more similar. This suggests that as foundation models mature, it may become possible to match vision and language embeddings in a fully unsupervised fashion, i.e. without parallel data. We present the first feasibility study, and investigate conformity of existing vision and language foundation models in the context of unsupervised, or "blind", matching. First, we formulate unsupervised matching as a quadratic assignment problem and introduce a novel heuristic that outperforms previous solvers. We also develop a technique to find optimal matching problems, for which a non-trivial match is very likely. Second, we conduct an extensive study deploying a range of vision and language models on four datasets. Our analysis reveals that for many problem instances, vision and language representations can be indeed matched without supervision. This finding opens up the exciting possibility of embedding semantic knowledge into other modalities virtually annotation-free. As a proof of concept, we showcase an unsupervised classifier, which achieves non-trivial classification accuracy without any image-text annotation.
We introduce a new function-preserving transformation for efficient neural architecture search. This network transformation allows reusing previously trained networks and existing successful architectures that improves sample efficiency. We aim to address the limitation of current network transformation operations that can only perform layer-level architecture modifications, such as adding (pruning) filters or inserting (removing) a layer, which fails to change the topology of connection paths. Our proposed path-level transformation operations enable the meta-controller to modify the path topology of the given network while keeping the merits of reusing weights, and thus allow efficiently designing effective structures with complex path topologies like Inception models. We further propose a bidirectional tree-structured reinforcement learning meta-controller to explore a simple yet highly expressive tree-structured architecture space that can be viewed as a generalization of multi-branch architectures. We experimented on the image classification datasets with limited computational resources (about 200 GPU-hours), where we observed improved parameter efficiency and better test results (97.70% test accuracy on CIFAR-10 with 14.3M parameters and 74.6% top-1 accuracy on ImageNet in the mobile setting), demonstrating the effectiveness and transferability of our designed architectures.
Machine learning (ML) is becoming a commodity. Numerous ML frameworks and services are available to data holders who are not ML experts but want to train predictive models on their data. It is important that ML models trained on sensitive inputs (e.g., personal images or documents) not leak too much information about the training data. We consider a malicious ML provider who supplies model-training code to the data holder, does not observe the training, but then obtains white- or black-box access to the resulting model. In this setting, we design and implement practical algorithms, some of them very similar to standard ML techniques such as regularization and data augmentation, that "memorize" information about the training dataset in the model yet the model is as accurate and predictive as a conventionally trained model. We then explain how the adversary can extract memorized information from the model. We evaluate our techniques on standard ML tasks for image classification (CIFAR10), face recognition (LFW and FaceScrub), and text analysis (20 Newsgroups and IMDB). In all cases, we show how our algorithms create models that have high predictive power yet allow accurate extraction of subsets of their training data.
We revisit the long-standing problem of learning from a small sample, to which end we propose a novel method called GLICO (Generative Latent Implicit Conditional Optimization). GLICO learns a mapping from the training examples to a latent space and a generator that generates images from vectors in the latent space. Unlike most recent works, which rely on access to large amounts of unlabeled data, GLICO does not require access to any additional data other than the small set of labeled points. In fact, GLICO learns to synthesize completely new samples for every class using as little as 5 or 10 examples per class, with as few as 10 such classes without imposing any prior. GLICO is then used to augment the small training set while training a classifier on the small sample. To this end, our proposed method samples the learned latent space using spherical interpolation, and generates new examples using the trained generator. Empirical results show that the new sampled set is diverse enough, leading to improvement in image classification in comparison with the state of the art, when trained on small samples obtained from CIFAR-10, CIFAR-100, and CUB-200.
Transformers with powerful global relation modeling abilities have been introduced to fundamental computer vision tasks recently. As a typical example, the Vision Transformer (ViT) directly applies a pure transformer architecture on image classification, by simply splitting images into tokens with a fixed length, and employing transformers to learn relations between these tokens. However, such naive tokenization could destruct object structures, assign grids to uninterested regions such as background, and introduce interference signals. To mitigate the above issues, in this paper, we propose an iterative and progressive sampling strategy to locate discriminative regions. At each iteration, embeddings of the current sampling step are fed into a transformer encoder layer, and a group of sampling offsets is predicted to update the sampling locations for the next step. The progressive sampling is differentiable. When combined with the Vision Transformer, the obtained PS-ViT network can adaptively learn where to look. The proposed PS-ViT is both effective and efficient. When trained from scratch on ImageNet, PS-ViT performs 3.8% higher than the vanilla ViT in terms of top-1 accuracy with about $4\times$ fewer parameters and $10\times$ fewer FLOPs. Code is available at https://github.com/yuexy/PS-ViT.
The vulnerability of the Lottery Ticket Hypothesis has not been studied from the purview of Membership Inference Attacks. Through this work, we are the first to empirically show that the lottery ticket networks are equally vulnerable to membership inference attacks. A Membership Inference Attack (MIA) is the process of determining whether a data sample belongs to a training set of a trained model or not. Membership Inference Attacks could leak critical information about the training data that can be used for targeted attacks. Recent deep learning models often have very large memory footprints and a high computational cost associated with training and drawing inferences. Lottery Ticket Hypothesis is used to prune the networks to find smaller sub-networks that at least match the performance of the original model in terms of test accuracy in a similar number of iterations. We used CIFAR-10, CIFAR-100, and ImageNet datasets to perform image classification tasks and observe that the attack accuracies are similar. We also see that the attack accuracy varies directly according to the number of classes in the dataset and the sparsity of the network. We demonstrate that these attacks are transferable across models with high accuracy.
COVID-19 quickly became a global pandemic after only four months of its first detection. It is crucial to detect this disease as soon as possible to decrease its spread. The use of chest X-ray (CXR) images became an effective screening strategy, complementary to the reverse transcription-polymerase chain reaction (RT-PCR). Convolutional neural networks (CNNs) are often used for automatic image classification and they can be very useful in CXR diagnostics. In this paper, 21 different CNN architectures are tested and compared in the task of identifying COVID-19 in CXR images. They were applied to the COVIDx8B dataset, a large COVID-19 dataset with 16,352 CXR images coming from patients of at least 51 countries. Ensembles of CNNs were also employed and they showed better efficacy than individual instances. The best individual CNN instance results were achieved by DenseNet169, with an accuracy of 98.15% and an F1 score of 98.12%. These were further increased to 99.25% and 99.24%, respectively, through an ensemble with five instances of DenseNet169. These results are higher than those obtained in recent works using the same dataset.
The release of tabular benchmarks, such as NAS-Bench-101 and NAS-Bench-201, has significantly lowered the computational overhead for conducting scientific research in neural architecture search (NAS). Although they have been widely adopted and used to tune real-world NAS algorithms, these benchmarks are limited to small search spaces and focus solely on image classification. Recently, several new NAS benchmarks have been introduced that cover significantly larger search spaces over a wide range of tasks, including object detection, speech recognition, and natural language processing. However, substantial differences among these NAS benchmarks have so far prevented their widespread adoption, limiting researchers to using just a few benchmarks. In this work, we present an in-depth analysis of popular NAS algorithms and performance prediction methods across 25 different combinations of search spaces and datasets, finding that many conclusions drawn from a few NAS benchmarks do not generalize to other benchmarks. To help remedy this problem, we introduce NAS-Bench-Suite, a comprehensive and extensible collection of NAS benchmarks, accessible through a unified interface, created with the aim to facilitate reproducible, generalizable, and rapid NAS research. Our code is available at https://github.com/automl/naslib.
To deploy deep learning algorithms on resource-limited scenarios, an emerging device-resistive random access memory (ReRAM) has been regarded as promising via analog computing. However, the practicability of ReRAM is primarily limited due to the weight drifting of ReRAM neural networks due to multi-factor reasons, including manufacturing, thermal noises, and etc. In this paper, we propose a novel Bayesian optimization method for fault tolerant neural network architecture (BayesFT). For neural architecture search space design, instead of conducting neural architecture search on the whole feasible neural architecture search space, we first systematically explore the weight drifting tolerance of different neural network components, such as dropout, normalization, number of layers, and activation functions in which dropout is found to be able to improve the neural network robustness to weight drifting. Based on our analysis, we propose an efficient search space by only searching for dropout rates for each layer. Then, we use Bayesian optimization to search for the optimal neural architecture robust to weight drifting. Empirical experiments demonstrate that our algorithmic framework has outperformed the state-of-the-art methods by up to 10 times on various tasks, such as image classification and object detection.
Masked Autoencoders (MAEs) learn generalizable representations for image, text, audio, video, etc., by reconstructing masked input data from tokens of the visible data. Current MAE approaches for videos rely on random patch, tube, or frame-based masking strategies to select these tokens. This paper proposes AdaMAE, an adaptive masking strategy for MAEs that is end-to-end trainable. Our adaptive masking strategy samples visible tokens based on the semantic context using an auxiliary sampling network. This network estimates a categorical distribution over spacetime-patch tokens. The tokens that increase the expected reconstruction error are rewarded and selected as visible tokens, motivated by the policy gradient algorithm in reinforcement learning. We show that AdaMAE samples more tokens from the high spatiotemporal information regions, thereby allowing us to mask 95% of tokens, resulting in lower memory requirements and faster pre-training. We conduct ablation studies on the Something-Something v2 (SSv2) dataset to demonstrate the efficacy of our adaptive sampling approach and report state-of-the-art results of 70.0% and 81.7% in top-1 accuracy on SSv2 and Kinetics-400 action classification datasets with a ViT-Base backbone and 800 pre-training epochs.