Prompt learning is an efficient approach to adapt transformers by inserting learnable set of parameters into the input and intermediate representations of a pre-trained model. In this work, we present Expressive Prompts with Residuals (EXPRES) which modifies the prompt learning paradigm specifically for effective adaptation of vision transformers (ViT). Our method constructs downstream representations via learnable “output” tokens (shal-low prompts), that are akin to the learned class tokens of the ViT. Further for better steering of the downstream representation processed by the frozen transformer, we introduce residual learnable tokens that are added to the output of various computations. We apply EXPRES for image classification and few-shot semantic segmentation, and show our method is capable of achieving state of the art prompt tuning on 3/3 categories of the VTAB benchmark. In addition to strong performance, we observe that our approach is an order of magnitude more prompt efficient than existing visual prompting baselines. We analytically show the computational benefits of our approach over weight space adaptation techniques like finetuning. Lastly we systematically corroborate the architectural design of our method via a series of ablation experiments.

We propose an approach to estimate the number of samples required for a model to reach a target performance. We find that the power law, the de facto principle to estimate model performance, leads to a large error when using a small dataset (e.g., 5 samples per class) for extrapolation. This is because the log-performance error against the log-dataset size follows a nonlinear progression in the few-shot regime followed by a linear progression in the high-shot regime. We introduce a novel piecewise power law (PPL) that handles the two data regimes differently. To estimate the parameters of the PPL, we introduce a random forest regressor trained via meta learning that generalizes across classification/detection tasks, ResNet/ViT based architectures, and random/pre-trained initializations. The PPL improves the performance estimation on average by 37% across 16 classification and 33% across 10 detection datasets, compared to the power law. We further extend the PPL to provide a confidence bound and use it to limit the prediction horizon that reduces over-estimation of data by 76% on classification and 91% on detection datasets.

Visual anomaly classification and segmentation are vital for automating industrial quality inspection. The focus of prior research in the field has been on training custom models for each quality inspection task, which requires task-specific images and annotation. In this paper we move away from this regime, addressing zero-shot and few-normal-shot anomaly classification and segmentation. Recently CLIP, a vision-language model, has shown revolutionary generality with competitive zero-/few-shot performance in comparison to full-supervision. But CLIP falls short on anomaly classification and segmentation tasks. Hence, we propose window-based CLIP (WinCLIP) with (1) a compositional ensemble on state words and prompt templates and (2) efficient extraction and aggregation of window/patch/image-level features aligned with text. We also propose its few-normal-shot extension Win-CLIP+, which uses complementary information from normal images. In MVTec-AD (and VisA), without further tuning, WinCLIP achieves 91.8%/85.1% (78.1%/79.6%) AU-ROC in zero-shot anomaly classification and segmentation while WinCLIP + does 93.1%/95.2% (83.8%/96.4%) in 1-normal-shot, surpassing state-of-the-art by large margins.

We present a framework for transfer learning that efficiently adapts a large base-model by learning lightweight cross-attention modules attached to its intermediate activations. We name our approach InCA (Introspective-Cross-Attention) and show that it can efficiently survey a network’s representations and identify strong performing adapter models for a downstream task. During training, InCA enables training numerous adapters efficiently and in parallel, isolated from the frozen base model. On the ViT-L/16 architecture, our experiments show that a single adapter, 1.3% of the full model, is able to reach full fine-tuning accuracy on average across 11 challenging downstream classification tasks. Compared with other forms of parameter-efficient adaptation, the isolated nature of the InCA adaptation is computationally desirable for large-scale models. For instance, we adapt ViT-G/14 (1.8B+ parameters) quickly with 20+ adapters in parallel on a single V100 GPU (76% GPU memory reduction) and exhaustively identify its most useful representations. We further demonstrate how the adapters learned by InCA can be incrementally modified or combined for flexible learning scenarios and our approach achieves state of the art performance on the ImageNet-to-Sketch multi-task benchmark.

Adapting pre-trained models with broad capabilities has become standard practice for learning a wide range of downstream tasks. The typical approach of fine-tuning different models for each task is performant, but incurs a substantial memory cost. To efficiently learn multiple down-stream tasks we introduce Task Adaptive Parameter Sharing (TAPS), a simple method for tuning a base model to a new task by adaptively modifying a small, task-specific subset of layers. This enables multi-task learning while minimizing the resources used and avoids catastrophic forgetting and competition between tasks. TAPS solves a joint optimization problem which determines both the layers that are shared with the base model and the value of the task-specific weights. Further, a sparsity penalty on the number of active layers promotes weight sharing with the base model. Compared to other methods, TAPS retains a high accuracy on the target tasks while still introducing only a small number of task-specific parameters. Moreover, TAPS is agnostic to the particular architecture used and requires only minor changes to the training scheme. We evaluate our method on a suite of fine-tuning tasks and architectures (ResNet, DenseNet, ViT) and show that it achieves state-of-the-art performance while being simple to implement.

Class-incremental learning (CIL) has been widely stud-ied under the setting of starting from a small number of classes (base classes). Instead, we explore an understud-ied real-world setting of CIL that starts with a strong model pre-trained on a large number of base classes. We hypoth-esize that a strong base model can provide a good repre-sentation for novel classes and incremental learning can be done with small adaptations. We propose a 2-stage training scheme, i) feature augmentation - cloning part of the backbone and fine-tuning it on the novel data, and ii) fusion - combining the base and novel classifiers into a unified classifier. Experiments show that the proposed method sig-nificantly outperforms state-of-the-art CIL methods on the large-scale ImageNet dataset (e.g. + 10% overall accuracy than the best). We also propose and analyze understudied practical CIL scenarios, such as base-novel overlap with distribution shift. Our proposed method is robust and gen-eralizes to all analyzed CIL settings.

We study semi-supervised learning (SSL) for vision transformers (ViT), an under-explored topic despite the wide adoption of the ViT architectures to different tasks. To tackle this problem, we use a SSL pipeline, consisting of first un/self-supervised pre-training, followed by supervised fine-tuning, and finally semi-supervised fine-tuning. At the semi-supervised fine-tuning stage, we adopt an exponential moving average (EMA)-Teacher framework instead of the popular FixMatch, since the former is more stable and delivers higher accuracy for semi-supervised vision transformers. In addition, we propose a probabilistic pseudo mixup mechanism to interpolate unlabeled samples and their pseudo labels for improved regularization, which is important for training ViTs with weak inductive bias. Our proposed method, dubbed Semi-ViT, achieves comparable or better performance than the CNN counterparts in the semi-supervised classification setting. Semi-ViT also enjoys the scalability benefits of ViTs that can be readily scaled up to large-size models with increasing accuracy. For example, Semi-ViT-Huge achieves an impressive 80\% top-1 accuracy on ImageNet using only 1\% labels, which is comparable with Inception-v4 using 100\% ImageNet labels. The code is available at https://github.com/amazon-science/semi-vit.

Classifiers that are linear in their parameters, and trained by optimizing a convex loss function, have predictable behavior with respect to changes in the training data, initial conditions, and optimization. Such desirable properties are absent in deep neural networks (DNNs), typically trained by non-linear fine-tuning of a pre-trained model. Previous attempts to linearize DNNs have led to interesting theoretical insights, but have not impacted the practice due to the substantial performance gap compared to standard non-linear optimization. We present the first method for linearizing a pre-trained model that achieves comparable performance to non-linear fine-tuning on most of real-world image classification tasks tested, thus enjoying the interpretability of linear models without incurring punishing losses in performance. LQF consists of simple modifications to the architecture, loss function and optimization typically used for classification: Leaky-ReLU instead of ReLU, mean squared loss instead of cross-entropy, and pre-conditioning using Kronecker factorization. None of these changes in isolation is sufficient to approach the performance of non-linear fine-tuning. When used in combination, they allow us to reach comparable performance, and even superior in the low-data regime, while enjoying the simplicity, robustness and interpretability of linear-quadratic optimization.

We present a plug-in replacement for batch normalization (BN) called exponential moving average normalization (EMAN), which improves the performance of existing student-teacher based self- and semi-supervised learning techniques. Unlike the standard BN, where the statistics are computed within each batch, EMAN, used in the teacher, updates its statistics by exponential moving average from the BN statistics of the student. This design reduces the intrinsic cross-sample dependency of BN and enhances the generalization of the teacher. EMAN improves strong baselines for self-supervised learning by 4-6/1-2 points and semi-supervised learning by about 7/2 points, when 1%/10% supervised labels are available on ImageNet. These improvements are consistent across methods, network architectures, training duration, and datasets, demonstrating the general effectiveness of this technique. The code will be made available online.

We show that the influence of a subset of the training samples can be removed – or "forgotten" – from the weights of a network trained on large-scale image classification tasks, and we provide strong computable bounds on the amount of remaining information after forgetting. Inspired by real-world applications of forgetting techniques, we introduce a novel notion of forgetting in mixed-privacy setting, where we know that a "core" subset of the training samples does not need to be forgotten. While this variation of the problem is conceptually simple, we show that working in this setting significantly improves the accuracy and guarantees of forgetting methods applied to vision classification tasks. Moreover, our method allows efficient removal of all information contained in non-core data by simply setting to zero a subset of the weights with minimal loss in performance. We achieve these results by replacing a standard deep network with a suitable linear approximation. With opportune changes to the network architecture and training procedure, we show that such linear approximation achieves comparable performance to the original network and that the forgetting problem becomes quadratic and can be solved efficiently even for large models. Unlike previous forgetting methods on deep networks, ours can achieve close to the state-of-the-art accuracy on large scale vision tasks. In particular, we show that our method allows forgetting without having to trade off the model accuracy.

Episodic training is a core ingredient of few-shot learning to train models on tasks with limited labelled data. Despite its success, episodic training remains largely understudied, prompting us to ask the question: what is the best way to sample episodes? In this paper, we first propose a method to approximate episode sampling distributions based on their difficulty. Building on this method, we perform an extensive analysis and find that sampling uniformly over episode difficulty outperforms other sampling schemes, including curriculum and easy-/hard-mining. As the proposed sampling method is algorithm agnostic, we can leverage these insights to improve few-shot learning accuracies across many episodic training algorithms. We demonstrate the efficacy of our method across popular few-shot learning datasets, algorithms, network architectures, and protocols.

We tackle the problem of predicting the number of optimization steps that a pre-trained deep network needs to converge to a given value of the loss function. To do so, we leverage the fact that the training dynamics of a deep network during fine-tuning are well approximated by those of a linearized model. This allows us to approximate the training loss and accuracy at any point during training by solving a low-dimensional Stochastic Differential Equation (SDE) in function space. Using this result, we are able to predict the time it takes for Stochastic Gradient Descent (SGD) to fine-tune a model to a given loss without having to perform any training. In our experiments, we are able to predict training time of a ResNet within a 20\% error margin on a variety of datasets and hyper-parameters, at a 30 to 45-fold reduction in cost compared to actual training. We also discuss how to further reduce the computational and memory cost of our method, and in particular we show that by exploiting the spectral properties of the gradients' matrix it is possible to predict training time on a large dataset while processing only a subset of the samples.

Many meta-learning approaches for few-shot learning rely on simple base learners such as nearest-neighbor classifiers. However, even in the few-shot regime, discriminatively trained linear predictors can offer better generalization. We propose to use these predictors as base learners to learn representations for few-shot learning and show they offer better tradeoffs between feature size and performance across a range of few-shot recognition benchmarks. Our objective is to learn feature embeddings that generalize well under a linear classification rule for novel categories. To efficiently solve the objective, we exploit two properties of linear classifiers: implicit differentiation of the optimality conditions of the convex problem and the dual formulation of the optimization problem. This allows us to use high-dimensional embeddings with improved generalization at a modest increase in computational overhead. Our approach, named MetaOptNet, achieves state-of-the-art performance on miniImageNet, tieredImageNet, CIFAR-FS, and FC100 few-shot learning benchmarks.

We introduce a method to generate vectorial representations of visual classification tasks which can be used to reason about the nature of those tasks and their relations. Given a dataset with ground-truth labels and a loss function, we process images through a "probe network" and compute an embedding based on estimates of the Fisher information matrix associated with the probe network parameters. This provides a fixed-dimensional embedding of the task that is independent of details such as the number of classes and requires no understanding of the class label semantics. We demonstrate that this embedding is capable of predicting task similarities that match our intuition about semantic and taxonomic relations between different visual tasks. We demonstrate the practical value of this framework for the meta-task of selecting a pre-trained feature extractor for a novel task. We present a simple meta-learning framework for learning a metric on embeddings that is capable of predicting which feature extractors will perform well on which task. Selecting a feature extractor with task embedding yields performance close to the best available feature extractor, with substantially less computational effort than exhaustively training and evaluating all available models.

We propose a method for learning embeddings for few-shot learning that is suitable for use with any number of shots (shot-free). Rather than fixing the class prototypes to be the Euclidean average of sample embeddings, we allow them to live in a higher-dimensional space (embedded class models) and learn the prototypes along with the model parameters. The class representation function is defined implicitly, which allows us to deal with a variable number of shots per class with a simple constant-size architecture. The class embedding encompasses metric learning, that facilitates adding new classes without crowding the class representation space. Despite being general and not tuned to the benchmark, our approach achieves state-of-the-art performance on the standard few-shot benchmark datasets.

We describe an information-driven active selection approach to determine which detectors to deploy at which location in which frame of a video to minimize semantic class label uncertainty at every pixel, with the smallest computational cost that ensures a given uncertainty bound. We show minimal performance reduction compared to a "paragon" algorithm running all detectors at all locations in all frames, at a small fraction of the computational cost. Our method can handle uncertainty in the labeling mechanism, so it can handle both "oracles" (manual annotation) or noisy detectors (automated annotation).

We consider the problem of categorizing video sequences of dynamic textures, i.e., nonrigid dynamical objects such as fire, water, steam, flags, etc. This problem is extremely challenging because the shape and appearance of a dynamic texture continuously change as a function of time. State-of-the-art dynamic texture categorization methods have been successful at classifying videos taken from the same viewpoint and scale by using a Linear Dynamical System (LDS) to model each video, and using distances or kernels in the space of LDSs to classify the videos. However, these methods perform poorly when the video sequences are taken under a different viewpoint or scale. In this paper, we propose a novel dynamic texture categorization framework that can handle such changes. We model each video sequence with a collection of LDSs, each one describing a small spatiotemporal patch extracted from the video. This Bag-of-Systems (BoS) representation is analogous to the Bag-of-Features (BoF) representation for object recognition, except that we use LDSs as feature descriptors. This choice poses several technical challenges in adopting the traditional BoF approach. Most notably, the space of LDSs is not euclidean; hence, novel methods for clustering LDSs and computing codewords of LDSs need to be developed. We propose a framework that makes use of nonlinear dimensionality reduction and clustering techniques combined with the Martin distance for LDSs to tackle these issues. Our experiments compare the proposed BoS approach to existing dynamic texture categorization methods and show that it can be used for recognizing dynamic textures in challenging scenarios which could not be handled by existing methods.

We consider the problem of non-blind deconvolution of images corrupted by a blur that is not accurately known. We propose a method that exploits dictionary-based image priors and non Gaussian noise models to improve deblurring accuracy in the presence of an inexact blur. The proposed image priors express each image patch as a linear combination of atoms from a dictionary learned from patches extracted from the same image or from an image database. When applied to blurred images, this model imposes that patches that are similar in the blurred image retain the same similarity when deblurred. We perform image deblurring by imposing this prior model in an energy minimization scheme that also deals with outliers. Experimental results on publicly available databases show that our approach is able to remove artifacts such as oscillations, which are often introduced during the deblurring process when the correct blur is not known.

We introduce a framework for defining a distance on the (non-Euclidean) space of Linear Dynamical Systems (LDSs). The proposed distance is induced by the action of the group of orthogonal matrices on the space of statespace realizations of LDSs. This distance can be efficiently computed for large-scale problems, hence it is suitable for applications in the analysis of dynamic visual scenes and other high dimensional time series. Based on this distance we devise a simple LDS averaging algorithm, which can be used for classification and clustering of time-series data. We test the validity as well as the performance of our group-action based distance on synthetic as well as real data and provide comparison with state-of-the-art methods.

We consider the problem of fitting one or more subspaces to a collection of data points drawn from the subspaces and corrupted by noise/outliers. We pose this problem as a rank minimization problem, where the goal is to decompose the corrupted data matrix as the sum of a clean, self-expressive, low-rank dictionary plus a matrix of noise/outliers. Our key contribution is to show that, for noisy data, this non-convex problem can be solved very efficiently and in closed form from the SVD of the noisy data matrix. Remarkably, this is true for both one or more subspaces. An important difference with respect to existing methods is that our framework results in a polynomial thresholding of the singular values with minimal shrinkage. Indeed, a particular case of our framework in the case of a single subspace leads to classical PCA, which requires no shrinkage. In the case of multiple subspaces, our framework provides an affinity matrix that can be used to cluster the data according to the sub-spaces. In the case of data corrupted by outliers, a closed-form solution appears elusive. We thus use an augmented Lagrangian optimization framework, which requires a combination of our proposed polynomial thresholding operator with the more traditional shrinkage-thresholding operator.

We consider the problem of spatially and temporally registering multiple video sequences of dynamical scenes which contain, but are not limited to, nonrigid objects such as fireworks, flags fluttering in the wind, etc., taken from different vantage points. This problem is extremely challenging due to the presence of complex variations in the appearance of such dynamic scenes. In this paper, we propose a simple algorithm for matching such complex scenes. Our algorithm does not require the cameras to be synchronized, and is not based on frame-by-frame or volume-by-volume registration. Instead, we model each video as the output of a linear dynamical system and transform the task of registering the video sequences to that of registering the parameters of the corresponding dynamical models. As these parameters are not uniquely defined, one cannot directly compare them to perform registration. We resolve these ambiguities by jointly identifying the parameters from multiple video sequences, and converting the identified parameters to a canonical form. This reduces the video registration problem to a multiple image registration problem, which can be efficiently solved using existing image matching techniques. We test our algorithm on a wide variety of challenging video sequences and show that it matches the performance of significantly more computationally expensive existing methods.

System theoretic approaches to action recognition model the dynamics of a scene with linear dynamical systems (LDSs) and perform classification using metrics on the space of LDSs, e.g. Binet-Cauchy kernels. However, such approaches are only applicable to time series data living in a Euclidean space, e.g. joint trajectories extracted from motion capture data or feature point trajectories extracted from video. Much of the success of recent object recognition techniques relies on the use of more complex feature descriptors, such as SIFT descriptors or HOG descriptors, which are essentially histograms. Since histograms live in a non-Euclidean space, we can no longer model their temporal evolution with LDSs, nor can we classify them using a metric for LDSs. In this paper, we propose to represent each frame of a video using a histogram of oriented optical flow (HOOF) and to recognize human actions by classifying HOOF time-series. For this purpose, we propose a generalization of the Binet-Cauchy kernels to nonlinear dynamical systems (NLDS) whose output lives in a non-Euclidean space, e.g. the space of histograms. This can be achieved by using kernels defined on the original non-Euclidean space, leading to a well-defined metric for NLDSs. We use these kernels for the classification of actions in video sequences using (HOOF) as the output of the NLDS. We evaluate our approach to recognition of human actions in several scenarios and achieve encouraging results.

In this paper, we consider the problem of categorizing videos of dynamic textures under varying view-point. We propose to model each video with a collection of linear dynamics systems (LDSs) describing the dynamics of spatiotemporal video patches. This bag of systems (BoS) representation is analogous to the bag of features (BoF) representation, except that we use LDSs as feature descriptors. This poses several technical challenges to the BoF framework. Most notably, LDSs do not live in a Euclidean space, hence novel methods for clustering LDSs and computing codewords of LDSs need to be developed. Our framework makes use of nonlinear dimensionality reduction and clustering techniques combined with the Martin distance for LDSs for tackling these issues. Our experiments show that our BoS approach can be used for recognizing dynamic textures in challenging scenarios, which could not be handled by existing dynamic texture recognition methods.

We consider the problem of modeling a scene containing multiple dynamic textures undergoing multiple rigid-body motions, e.g., a video sequence of water taken by a rigidly moving camera. We propose to model each moving dynamic texture with a time varying linear dynamical system (LDS) plus a 2D translational motion model. We first consider a scene with a single moving dynamic texture and show how to simultaneously learn the parameters of the time varying LDS as well as the optical flow of the scene using the so-called dynamic texture constancy constraint (DTCC). We then consider a scene with multiple non-moving dynamic textures and show that learning the parameters of each time invariant LDS as well as their region of support is equivalent to clustering data living in multiple subspaces. We solve this problem with a combination of PCA and GPCA. Finally, we consider a scene with multiple moving dynamic textures, and show how to simultaneously learn the parameters of multiple time varying LDS and multiple 2D translational models, by clustering data living in multiple dynamically evolving subspaces. We test our approach on sequences of flowers, water, grass, and a beating heart.