Screening for a Reweighted Penalized Conditional Gradient Method

The conditional gradient method (CGM) is widely used in large-scale sparse convex optimization, having a low per iteration computational cost for structured sparse regularizers and a greedy approach to collecting nonzeros. We explore the sparsity acquiring properties of a general penalized CGM (P-CGM) for convex regularizers and a reweighted penalized CGM (RP-CGM) for nonconvex regularizers, replacing the usual convex constraints with gauge-inspired penalties. This generalization does not increase the per-iteration complexity noticeably. Without assuming bounded iterates or using line search, we show $O(1/t)$ convergence of the gap of each subproblem, which measures distance to a stationary point. We couple this with a screening rule which is safe in the convex case, converging to the true support at a rate $O(1/(\delta^2))$ where $\delta \geq 0$ measures how close the problem is to degeneracy. In the nonconvex case the screening rule converges to the true support in a finite number of iterations, but is not necessarily safe in the intermediate iterates. In our experiments, we verify the consistency of the method and adjust the aggressiveness of the screening rule by tuning the concavity of the regularizer.

Approximate Frank-Wolfe Algorithms over Graph-structured Support Sets

In this paper, we propose approximate Frank-Wolfe (FW) algorithms to solve convex optimization problems over graph-structured support sets where the \textit{linear minimization oracle} (LMO) cannot be efficiently obtained in general. We first demonstrate that two popular approximation assumptions (\textit{additive} and \textit{multiplicative gap errors)}, are not valid for our problem, in that no cheap gap-approximate LMO oracle exists in general. Instead, a new \textit{approximate dual maximization oracle} (DMO) is proposed, which approximates the inner product rather than the gap. When the objective is $L$-smooth, we prove that the standard FW method using a $\delta$-approximate DMO converges as $\mathcal{O}(L / \delta t + (1-\delta)(\delta^{-1} + \delta^{-2}))$ in general, and as $\mathcal{O}(L/(\delta^2(t+2)))$ over a $\delta$-relaxation of the constraint set. Additionally, when the objective is $\mu$-strongly convex and the solution is unique, a variant of FW converges to $\mathcal{O}(L^2\log(t)/(\mu \delta^6 t^2))$ with the same per-iteration complexity. Our empirical results suggest that even these improved bounds are pessimistic, with significant improvement in recovering real-world images with graph-structured sparsity.

Dynamic Metric Learning: Towards a Scalable Metric Space to Accommodate Multiple Semantic Scales

This paper introduces a new fundamental characteristic, \ie, the dynamic range, from real-world metric tools to deep visual recognition. In metrology, the dynamic range is a basic quality of a metric tool, indicating its flexibility to accommodate various scales. Larger dynamic range offers higher flexibility. In visual recognition, the multiple scale problem also exist. Different visual concepts may have different semantic scales. For example, ``Animal'' and ``Plants'' have a large semantic scale while ``Elk'' has a much smaller one. Under a small semantic scale, two different elks may look quite \emph{different} to each other . However, under a large semantic scale (\eg, animals and plants), these two elks should be measured as being \emph{similar}. %We argue that such flexibility is also important for deep metric learning, because different visual concepts indeed correspond to different semantic scales. Introducing the dynamic range to deep metric learning, we get a novel computer vision task, \ie, the Dynamic Metric Learning. It aims to learn a scalable metric space to accommodate visual concepts across multiple semantic scales. Based on three types of images, \emph{i.e.}, vehicle, animal and online products, we construct three datasets for Dynamic Metric Learning. We benchmark these datasets with popular deep metric learning methods and find Dynamic Metric Learning to be very challenging. The major difficulty lies in a conflict between different scales: the discriminative ability under a small scale usually compromises the discriminative ability under a large one, and vice versa. As a minor contribution, we propose Cross-Scale Learning (CSL) to alleviate such conflict. We show that CSL consistently improves the baseline on all the three datasets. The datasets and the code will be publicly available at https://github.com/SupetZYK/DynamicMetricLearning.

Latent Group Structured Multi-task Learning

In multi-task learning (MTL), we improve the performance of key machine learning algorithms by training various tasks jointly. When the number of tasks is large, modeling task structure can further refine the task relationship model. For example, often tasks can be grouped based on metadata, or via simple preprocessing steps like K-means. In this paper, we present our group structured latent-space multi-task learning model, which encourages group structured tasks defined by prior information. We use an alternating minimization method to learn the model parameters. Experiments are conducted on both synthetic and real-world datasets, showing competitive performance over single-task learning (where each group is trained separately) and other MTL baselines.

Memory-based Jitter: Improving Visual Recognition on Long-tailed Data with Diversity In Memory

This paper considers deep visual recognition on long-tailed data, with the majority categories only occupying relatively few samples. The tail categories are prone to lack of within-class diversity, which compromises the representative ability of the learned visual concepts. A radical solution is to augment the tail categories with higher diversity. To this end, we introduce a simple and reliable method named Memory-based Jitter (MBJ) to gain extra diversity for the tail data. We observe that the deep model keeps on jittering from one historical edition to another, even when it already approaches convergence. The ``jitter'' means the small variations between historical models. We argue that such jitter largely originates from the within-class diversity of the overall data and thus encodes the within-class distribution pattern. To utilize such jitter for tail data augmentation, we store the jitter among historical models into a memory bank and get the so-called Memory-based Jitter. With slight modifications, MBJ is applicable for two fundamental visual recognition tasks, \emph{i.e.}, image classification and deep metric learning (on long-tailed data). On image classification, MBJ collects the historical embeddings to learn an accurate classifier. In contrast, on deep metric learning, it collects the historical prototypes of each class to learn a robust deep embedding. Under both scenarios, MBJ enforces higher concentration on tail classes, so as to compensate for their lack of diversity. Extensive experiments on three long-tailed classification benchmarks and two deep metric learning benchmarks (person re-identification, in particular) demonstrate the significant improvement. Moreover, the achieved performance are on par with the state-of-the-art on both tasks.

CycAs: Self-supervised Cycle Association for Learning Re-identifiable Descriptions

This paper proposes a self-supervised learning method for the person re-identification (re-ID) problem, where existing unsupervised methods usually rely on pseudo labels, such as those from video tracklets or clustering. A potential drawback of using pseudo labels is that errors may accumulate and it is challenging to estimate the number of pseudo IDs. We introduce a different unsupervised method that allows us to learn pedestrian embeddings from raw videos, without resorting to pseudo labels. The goal is to construct a self-supervised pretext task that matches the person re-ID objective. Inspired by the \emph{data association} concept in multi-object tracking, we propose the \textbf{Cyc}le \textbf{As}sociation (\textbf{CycAs}) task: after performing data association between a pair of video frames forward and then backward, a pedestrian instance is supposed to be associated to itself. To fulfill this goal, the model must learn a meaningful representation that can well describe correspondences between instances in frame pairs. We adapt the discrete association process to a differentiable form, such that end-to-end training becomes feasible. Experiments are conducted in two aspects: We first compare our method with existing unsupervised re-ID methods on seven benchmarks and demonstrate CycAs' superiority. Then, to further validate the practical value of CycAs in real-world applications, we perform training on self-collected videos and report promising performance on standard test sets.

Embodied Self-supervised Learning by Coordinated Sampling and Training

Self-supervised learning can significantly improve the performance of downstream tasks, however, the dimensions of learned representations normally lack explicit physical meanings. In this work, we propose a novel self-supervised approach to solve inverse problems by employing the corresponding physical forward process so that the learned representations can have explicit physical meanings. The proposed approach works in an analysis-by-synthesis manner to learn an inference network by iteratively sampling and training. At the sampling step, given observed data, the inference network is used to approximate the intractable posterior, from which we sample input parameters and feed them to a physical process to generate data in the observational space; At the training step, the same network is optimized with the sampled paired data. We prove the feasibility of the proposed method by tackling the acoustic-to-articulatory inversion problem to infer articulatory information from speech. Given an articulatory synthesizer, an inference model can be trained completely from scratch with random initialization. Our experiments demonstrate that the proposed method can converge steadily and the network learns to control the articulatory synthesizer to speak like a human. We also demonstrate that trained models can generalize well to unseen speakers or even new languages, and performance can be further improved through self-adaptation.

Deep Representation Learning on Long-tailed Data: A Learnable Embedding Augmentation Perspective

This paper considers learning deep features from long-tailed data. We observe that in the deep feature space, the head classes and the tail classes present different distribution patterns. The head classes have a relatively large spatial span, while the tail classes have significantly small spatial span, due to the lack of intra-class diversity. This uneven distribution between head and tail classes distorts the overall feature space, which compromises the discriminative ability of the learned features. Intuitively, we seek to expand the distribution of the tail classes by transferring from the head classes, so as to alleviate the distortion of the feature space. To this end, we propose to construct each feature into a "feature cloud". If a sample belongs to a tail class, the corresponding feature cloud will have relatively large distribution range, in compensation to its lack of diversity. It allows each tail sample to push the samples from other classes far away, recovering the intra-class diversity of tail classes. Extensive experimental evaluations on person re-identification and face recognition tasks confirm the effectiveness of our method.

Circle Loss: A Unified Perspective of Pair Similarity Optimization

This paper provides a pair similarity optimization viewpoint on deep feature learning, aiming to maximize the within-class similarity $s_p$ and minimize the between-class similarity $s_n$. We find a majority of loss functions, including the triplet loss and the softmax plus cross-entropy loss, embed $s_n$ and $s_p$ into similarity pairs and seek to reduce $(s_n-s_p)$. Such an optimization manner is inflexible, because the penalty strength on every single similarity score is restricted to be equal. Our intuition is that if a similarity score deviates far from the optimum, it should be emphasized. To this end, we simply re-weight each similarity to highlight the less-optimized similarity scores. It results in a Circle loss, which is named due to its circular decision boundary. The Circle loss has a unified formula for two elemental deep feature learning approaches, i.e. learning with class-level labels and pair-wise labels. Analytically, we show that the Circle loss offers a more flexible optimization approach towards a more definite convergence target, compared with the loss functions optimizing $(s_n-s_p)$. Experimentally, we demonstrate the superiority of the Circle loss on a variety of deep feature learning tasks. On face recognition, person re-identification, as well as several fine-grained image retrieval datasets, the achieved performance is on par with the state of the art.