How Powerful Potential of Attention on Image Restoration?

Transformers have demonstrated their effectiveness in image restoration tasks. Existing Transformer architectures typically comprise two essential components: multi-head self-attention and feed-forward network (FFN). The former captures long-range pixel dependencies, while the latter enables the model to learn complex patterns and relationships in the data. Previous studies have demonstrated that FFNs are key-value memories \cite{geva2020transformer}, which are vital in modern Transformer architectures. In this paper, we conduct an empirical study to explore the potential of attention mechanisms without using FFN and provide novel structures to demonstrate that removing FFN is flexible for image restoration. Specifically, we propose Continuous Scaling Attention (\textbf{CSAttn}), a method that computes attention continuously in three stages without using FFN. To achieve competitive performance, we propose a series of key components within the attention. Our designs provide a closer look at the attention mechanism and reveal that some simple operations can significantly affect the model performance. We apply our \textbf{CSAttn} to several image restoration tasks and show that our model can outperform CNN-based and Transformer-based image restoration approaches.

SequentialAttention++ for Block Sparsification: Differentiable Pruning Meets Combinatorial Optimization

Neural network pruning is a key technique towards engineering large yet scalable, interpretable, and generalizable models. Prior work on the subject has developed largely along two orthogonal directions: (1) differentiable pruning for efficiently and accurately scoring the importance of parameters, and (2) combinatorial optimization for efficiently searching over the space of sparse models. We unite the two approaches, both theoretically and empirically, to produce a coherent framework for structured neural network pruning in which differentiable pruning guides combinatorial optimization algorithms to select the most important sparse set of parameters. Theoretically, we show how many existing differentiable pruning techniques can be understood as nonconvex regularization for group sparse optimization, and prove that for a wide class of nonconvex regularizers, the global optimum is unique, group-sparse, and provably yields an approximate solution to a sparse convex optimization problem. The resulting algorithm that we propose, SequentialAttention++, advances the state of the art in large-scale neural network block-wise pruning tasks on the ImageNet and Criteo datasets.

Learning from Aggregate responses: Instance Level versus Bag Level Loss Functions

Due to the rise of privacy concerns, in many practical applications the training data is aggregated before being shared with the learner, in order to protect privacy of users' sensitive responses. In an aggregate learning framework, the dataset is grouped into bags of samples, where each bag is available only with an aggregate response, providing a summary of individuals' responses in that bag. In this paper, we study two natural loss functions for learning from aggregate responses: bag-level loss and the instance-level loss. In the former, the model is learnt by minimizing a loss between aggregate responses and aggregate model predictions, while in the latter the model aims to fit individual predictions to the aggregate responses. In this work, we show that the instance-level loss can be perceived as a regularized form of the bag-level loss. This observation lets us compare the two approaches with respect to bias and variance of the resulting estimators, and introduce a novel interpolating estimator which combines the two approaches. For linear regression tasks, we provide a precise characterization of the risk of the interpolating estimator in an asymptotic regime where the size of the training set grows in proportion to the features dimension. Our analysis allows us to theoretically understand the effect of different factors, such as bag size on the model prediction risk. In addition, we propose a mechanism for differentially private learning from aggregate responses and derive the optimal bag size in terms of prediction risk-privacy trade-off. We also carry out thorough experiments to corroborate our theory and show the efficacy of the interpolating estimator.

Greedy PIG: Adaptive Integrated Gradients

Deep learning has become the standard approach for most machine learning tasks. While its impact is undeniable, interpreting the predictions of deep learning models from a human perspective remains a challenge. In contrast to model training, model interpretability is harder to quantify and pose as an explicit optimization problem. Inspired by the AUC softmax information curve (AUC SIC) metric for evaluating feature attribution methods, we propose a unified discrete optimization framework for feature attribution and feature selection based on subset selection. This leads to a natural adaptive generalization of the path integrated gradients (PIG) method for feature attribution, which we call Greedy PIG. We demonstrate the success of Greedy PIG on a wide variety of tasks, including image feature attribution, graph compression/explanation, and post-hoc feature selection on tabular data. Our results show that introducing adaptivity is a powerful and versatile method for making attribution methods more powerful.

Towards High-Quality Specular Highlight Removal by Leveraging Large-Scale Synthetic Data

This paper aims to remove specular highlights from a single object-level image. Although previous methods have made some progresses, their performance remains somewhat limited, particularly for real images with complex specular highlights. To this end, we propose a three-stage network to address them. Specifically, given an input image, we first decompose it into the albedo, shading, and specular residue components to estimate a coarse specular-free image. Then, we further refine the coarse result to alleviate its visual artifacts such as color distortion. Finally, we adjust the tone of the refined result to match that of the input as closely as possible. In addition, to facilitate network training and quantitative evaluation, we present a large-scale synthetic dataset of object-level images, covering diverse objects and illumination conditions. Extensive experiments illustrate that our network is able to generalize well to unseen real object-level images, and even produce good results for scene-level images with multiple background objects and complex lighting.

Learning from Aggregated Data: Curated Bags versus Random Bags

Protecting user privacy is a major concern for many machine learning systems that are deployed at scale and collect from a diverse set of population. One way to address this concern is by collecting and releasing data labels in an aggregated manner so that the information about a single user is potentially combined with others. In this paper, we explore the possibility of training machine learning models with aggregated data labels, rather than individual labels. Specifically, we consider two natural aggregation procedures suggested by practitioners: curated bags where the data points are grouped based on common features and random bags where the data points are grouped randomly in bag of similar sizes. For the curated bag setting and for a broad range of loss functions, we show that we can perform gradient-based learning without any degradation in performance that may result from aggregating data. Our method is based on the observation that the sum of the gradients of the loss function on individual data examples in a curated bag can be computed from the aggregate label without the need for individual labels. For the random bag setting, we provide a generalization risk bound based on the Rademacher complexity of the hypothesis class and show how empirical risk minimization can be regularized to achieve the smallest risk bound. In fact, in the random bag setting, there is a trade-off between size of the bag and the achievable error rate as our bound indicates. Finally, we conduct a careful empirical study to confirm our theoretical findings. In particular, our results suggest that aggregate learning can be an effective method for preserving user privacy while maintaining model accuracy.

Approximately Optimal Core Shapes for Tensor Decompositions

This work studies the combinatorial optimization problem of finding an optimal core tensor shape, also called multilinear rank, for a size-constrained Tucker decomposition. We give an algorithm with provable approximation guarantees for its reconstruction error via connections to higher-order singular values. Specifically, we introduce a novel Tucker packing problem, which we prove is NP-hard, and give a polynomial-time approximation scheme based on a reduction to the 2-dimensional knapsack problem with a matroid constraint. We also generalize our techniques to tree tensor network decompositions. We implement our algorithm using an integer programming solver, and show that its solution quality is competitive with (and sometimes better than) the greedy algorithm that uses the true Tucker decomposition loss at each step, while also running up to 1000x faster.

Sequential Attention for Feature Selection

Feature selection is the problem of selecting a subset of features for a machine learning model that maximizes model quality subject to a resource budget constraint. For neural networks, prior methods, including those based on $\ell_1$ regularization, attention, and stochastic gates, typically select all of the features in one evaluation round, ignoring the residual value of the features during selection (i.e., the marginal contribution of a feature conditioned on the previously selected features). We propose a feature selection algorithm called Sequential Attention that achieves state-of-the-art empirical results for neural networks. This algorithm is based on an efficient implementation of greedy forward selection and uses attention weights at each step as a proxy for marginal feature importance. We provide theoretical insights into our Sequential Attention algorithm for linear regression models by showing that an adaptation to this setting is equivalent to the classical Orthogonal Matching Pursuit algorithm [PRK1993], and thus inherits all of its provable guarantees. Lastly, our theoretical and empirical analyses provide new explanations towards the effectiveness of attention and its connections to overparameterization, which might be of independent interest.

Scene Inference for Object Illumination Editing

The seamless illumination integration between a foreground object and a background scene is an important but challenging task in computer vision and augmented reality community. However, to our knowledge, there is no publicly available high-quality dataset that meets the illumination seamless integration task, which greatly hinders the development of this research direction. To this end, we apply a physically-based rendering method to create a large-scale, high-quality dataset, named IH dataset, which provides rich illumination information for seamless illumination integration task. In addition, we propose a deep learning-based SI-GAN method, a multi-task collaborative network, which makes full use of the multi-scale attention mechanism and adversarial learning strategy to directly infer mapping relationship between the inserted foreground object and corresponding background environment, and edit object illumination according to the proposed illumination exchange mechanism in parallel network. By this means, we can achieve the seamless illumination integration without explicit estimation of 3D geometric information. Comprehensive experiments on both our dataset and real-world images collected from the Internet show that our proposed SI-GAN provides a practical and effective solution for image-based object illumination editing, and validate the superiority of our method against state-of-the-art methods.

Feature Cross Search via Submodular Optimization

In this paper, we study feature cross search as a fundamental primitive in feature engineering. The importance of feature cross search especially for the linear model has been known for a while, with well-known textbook examples. In this problem, the goal is to select a small subset of features, combine them to form a new feature (called the crossed feature) by considering their Cartesian product, and find feature crosses to learn an \emph{accurate} model. In particular, we study the problem of maximizing a normalized Area Under the Curve (AUC) of the linear model trained on the crossed feature column. First, we show that it is not possible to provide an $n^{1/\log\log n}$-approximation algorithm for this problem unless the exponential time hypothesis fails. This result also rules out the possibility of solving this problem in polynomial time unless $\mathsf{P}=\mathsf{NP}$. On the positive side, by assuming the \naive\ assumption, we show that there exists a simple greedy $(1-1/e)$-approximation algorithm for this problem. This result is established by relating the AUC to the total variation of the commutator of two probability measures and showing that the total variation of the commutator is monotone and submodular. To show this, we relate the submodularity of this function to the positive semi-definiteness of a corresponding kernel matrix. Then, we use Bochner's theorem to prove the positive semi-definiteness by showing that its inverse Fourier transform is non-negative everywhere. Our techniques and structural results might be of independent interest.