\textit{Relative overgeneralization} (RO) occurs in cooperative multi-agent learning tasks when agents converge towards a suboptimal joint policy due to overfitting to suboptimal behavior of other agents. In early work, optimism has been shown to mitigate the \textit{RO} problem when using tabular Q-learning. However, with function approximation optimism can amplify overestimation and thus fail on complex tasks. On the other hand, recent deep multi-agent policy gradient (MAPG) methods have succeeded in many complex tasks but may fail with severe \textit{RO}. We propose a general, yet simple, framework to enable optimistic updates in MAPG methods and alleviate the RO problem. Specifically, we employ a \textit{Leaky ReLU} function where a single hyperparameter selects the degree of optimism to reshape the advantages when updating the policy. Intuitively, our method remains optimistic toward individual actions with lower returns which are potentially caused by other agents' sub-optimal behavior during learning. The optimism prevents the individual agents from quickly converging to a local optimum. We also provide a formal analysis from an operator view to understand the proposed advantage transformation. In extensive evaluations on diverse sets of tasks, including illustrative matrix games, complex \textit{Multi-agent MuJoCo} and \textit{Overcooked} benchmarks, the proposed method\footnote{Code can be found at \url{https://github.com/wenshuaizhao/optimappo}.} outperforms strong baselines on 13 out of 19 tested tasks and matches the performance on the rest.

Given the massive cost of language model pre-training, a non-trivial improvement of the optimization algorithm would lead to a material reduction on the time and cost of training. Adam and its variants have been state-of-the-art for years, and more sophisticated second-order (Hessian-based) optimizers often incur too much per-step overhead. In this paper, we propose Sophia, Second-order Clipped Stochastic Optimization, a simple scalable second-order optimizer that uses a light-weight estimate of the diagonal Hessian as the pre-conditioner. The update is the moving average of the gradients divided by the moving average of the estimated Hessian, followed by element-wise clipping. The clipping controls the worst-case update size and tames the negative impact of non-convexity and rapid change of Hessian along the trajectory. Sophia only estimates the diagonal Hessian every handful of iterations, which has negligible average per-step time and memory overhead. On language modeling with GPT models of sizes ranging from 125M to 1.5B, Sophia achieves a 2x speed-up compared to Adam in the number of steps, total compute, and wall-clock time, achieving the same perplexity with 50% fewer steps, less total compute, and reduced wall-clock time. Theoretically, we show that Sophia, in a much simplified setting, adapts to the heterogeneous curvatures in different parameter dimensions, and thus has a run-time bound that does not depend on the condition number of the loss.

Training an image captioner without annotated image-sentence pairs has gained traction in recent years. Previous approaches can be categorized into two strategies: crawling sentences from mismatching corpora and aligning them with the given images as pseudo annotations, or pre-training the captioner using external image-text pairs. However, the aligning setting seems to reach its performance limit due to the quality problem of pairs, and pre-training requires significant computational resources. To address these challenges, we propose a new strategy ``LPM + retrieval-augmented learning" where the prior knowledge from large pre-trained models (LPMs) is leveraged as supervision, and a retrieval process is integrated to further reinforce its effectiveness. Specifically, we introduce Retrieval-augmented Pseudo Sentence Generation (RaPSG), which adopts an efficient approach to retrieve highly relevant short region descriptions from the mismatching corpora and use them to generate a variety of pseudo sentences with distinct representations as well as high quality via LPMs. In addition, a fluency filter and a CLIP-guided training objective are further introduced to facilitate model optimization. Experimental results demonstrate that our method surpasses the SOTA pre-training model (Flamingo3B) by achieving a CIDEr score of 78.1 (+5.1) while utilizing only 0.3% of its trainable parameters (1.3B VS 33M). Importantly, our approach eliminates the need of computationally expensive pre-training processes on external datasets (e.g., the requirement of 312M image-text pairs for Flamingo3B). We further show that with a simple extension, the generated pseudo sentences can be deployed as weak supervision to boost the 1% semi-supervised image caption benchmark up to 93.4 CIDEr score (+8.9) which showcases the versatility and effectiveness of our approach.

Momentum is known to accelerate the convergence of gradient descent in strongly convex settings without stochastic gradient noise. In stochastic optimization, such as training neural networks, folklore suggests that momentum may help deep learning optimization by reducing the variance of the stochastic gradient update, but previous theoretical analyses do not find momentum to offer any provable acceleration. Theoretical results in this paper clarify the role of momentum in stochastic settings where the learning rate is small and gradient noise is the dominant source of instability, suggesting that SGD with and without momentum behave similarly in the short and long time horizons. Experiments show that momentum indeed has limited benefits for both optimization and generalization in practical training regimes where the optimal learning rate is not very large, including small- to medium-batch training from scratch on ImageNet and fine-tuning language models on downstream tasks.

For multi-modal magnetic resonance (MR) brain tumor image segmentation, current methods usually directly extract the discriminative features from input images for tumor sub-region category determination and localization. However, the impact of information aliasing caused by the mutual inclusion of tumor sub-regions is often ignored. Moreover, existing methods usually do not take tailored efforts to highlight the single tumor sub-region features. To this end, a multi-modal MR brain tumor segmentation method with tumor prototype-driven and multi-expert integration is proposed. It could highlight the features of each tumor sub-region under the guidance of tumor prototypes. Specifically, to obtain the prototypes with complete information, we propose a mutual transmission mechanism to transfer different modal features to each other to address the issues raised by insufficient information on single-modal features. Furthermore, we devise a prototype-driven feature representation and fusion method with the learned prototypes, which implants the prototypes into tumor features and generates corresponding activation maps. With the activation maps, the sub-region features consistent with the prototype category can be highlighted. A key information enhancement and fusion strategy with multi-expert integration is designed to further improve the segmentation performance. The strategy can integrate the features from different layers of the extra feature extraction network and the features highlighted by the prototypes. Experimental results on three competition brain tumor segmentation datasets prove the superiority of the proposed method.

Despite extensive studies, the underlying reason as to why overparameterized neural networks can generalize remains elusive. Existing theory shows that common stochastic optimizers prefer flatter minimizers of the training loss, and thus a natural potential explanation is that flatness implies generalization. This work critically examines this explanation. Through theoretical and empirical investigation, we identify the following three scenarios for two-layer ReLU networks: (1) flatness provably implies generalization; (2) there exist non-generalizing flattest models and sharpness minimization algorithms fail to generalize, and (3) perhaps most surprisingly, there exist non-generalizing flattest models, but sharpness minimization algorithms still generalize. Our results suggest that the relationship between sharpness and generalization subtly depends on the data distributions and the model architectures and sharpness minimization algorithms do not only minimize sharpness to achieve better generalization. This calls for the search for other explanations for the generalization of over-parameterized neural networks.

Recent works on over-parameterized neural networks have shown that the stochasticity in optimizers has the implicit regularization effect of minimizing the sharpness of the loss function (in particular, the trace of its Hessian) over the family zero-loss solutions. More explicit forms of flatness regularization also empirically improve the generalization performance. However, it remains unclear why and when flatness regularization leads to better generalization. This work takes the first step toward understanding the inductive bias of the minimum trace of the Hessian solutions in an important setting: learning deep linear networks from linear measurements, also known as \emph{deep matrix factorization}. We show that for all depth greater than one, with the standard Restricted Isometry Property (RIP) on the measurements, minimizing the trace of Hessian is approximately equivalent to minimizing the Schatten 1-norm of the corresponding end-to-end matrix parameters (i.e., the product of all layer matrices), which in turn leads to better generalization. We empirically verify our theoretical findings on synthetic datasets.

Deep metric learning techniques have been used for visual representation in various supervised and unsupervised learning tasks through learning embeddings of samples with deep networks. However, classic approaches, which employ a fixed distance metric as a similarity function between two embeddings, may lead to suboptimal performance for capturing the complex data distribution. The Bregman divergence generalizes measures of various distance metrics and arises throughout many fields of deep metric learning. In this paper, we first show how deep metric learning loss can arise from the Bregman divergence. We then introduce a novel method for learning empirical Bregman divergence directly from data based on parameterizing the convex function underlying the Bregman divergence with a deep learning setting. We further experimentally show that our approach performs effectively on five popular public datasets compared to other SOTA deep metric learning methods, particularly for pattern recognition problems.

Recently, deep metric learning techniques received attention, as the learned distance representations are useful to capture the similarity relationship among samples and further improve the performance of various of supervised or unsupervised learning tasks. We propose a novel supervised metric learning method that can learn the distance metrics in both geometric and probabilistic space for image recognition. In contrast to the previous metric learning methods which usually focus on learning the distance metrics in Euclidean space, our proposed method is able to learn better distance representation in a hybrid approach. To achieve this, we proposed a Generalized Hybrid Metric Loss (GHM-Loss) to learn the general hybrid proximity features from the image data by controlling the trade-off between geometric proximity and probabilistic proximity. To evaluate the effectiveness of our method, we first provide theoretical derivations and proofs of the proposed loss function, then we perform extensive experiments on two public datasets to show the advantage of our method compared to other state-of-the-art metric learning methods.

The computer-aided disease diagnosis from radiomic data is important in many medical applications. However, developing such a technique relies on annotating radiological images, which is a time-consuming, labor-intensive, and expensive process. In this work, we present the first novel collaborative self-supervised learning method to solve the challenge of insufficient labeled radiomic data, whose characteristics are different from text and image data. To achieve this, we present two collaborative pretext tasks that explore the latent pathological or biological relationships between regions of interest and the similarity and dissimilarity information between subjects. Our method collaboratively learns the robust latent feature representations from radiomic data in a self-supervised manner to reduce human annotation efforts, which benefits the disease diagnosis. We compared our proposed method with other state-of-the-art self-supervised learning methods on a simulation study and two independent datasets. Extensive experimental results demonstrated that our method outperforms other self-supervised learning methods on both classification and regression tasks. With further refinement, our method shows the potential advantage in automatic disease diagnosis with large-scale unlabeled data available.