Adam with decoupled weight decay, also known as AdamW, is widely acclaimed for its superior performance in language modeling tasks, surpassing Adam with $\ell_2$ regularization in terms of generalization and optimization. However, this advantage is not theoretically well-understood. One challenge here is that though intuitively Adam with $\ell_2$ regularization optimizes the $\ell_2$ regularized loss, it is not clear if AdamW optimizes a specific objective. In this work, we make progress toward understanding the benefit of AdamW by showing that it implicitly performs constrained optimization. More concretely, we show in the full-batch setting, if AdamW converges with any non-increasing learning rate schedule whose partial sum diverges, it must converge to a KKT point of the original loss under the constraint that the $\ell_\infty$ norm of the parameter is bounded by the inverse of the weight decay factor. This result is built on the observation that Adam can be viewed as a smoothed version of SignGD, which is the normalized steepest descent with respect to $\ell_\infty$ norm, and a surprising connection between normalized steepest descent with weight decay and Frank-Wolfe.

In the rapidly evolving domain of artificial intelligence, Large Language Models (LLMs) play a crucial role due to their advanced text processing and generation abilities. This study introduces a new strategy aimed at harnessing on-device LLMs in invoking software APIs. We meticulously compile a dataset derived from software API documentation and apply fine-tuning to LLMs with capacities of 2B, 3B and 7B parameters, specifically to enhance their proficiency in software API interactions. Our approach concentrates on refining the models' grasp of API structures and syntax, significantly enhancing the accuracy of API function calls. Additionally, we propose \textit{conditional masking} techniques to ensure outputs in the desired formats and reduce error rates while maintaining inference speeds. We also propose a novel benchmark designed to evaluate the effectiveness of LLMs in API interactions, establishing a foundation for subsequent research. Octopus, the fine-tuned model, is proved to have better performance than GPT-4 for the software APIs calling. This research aims to advance automated software development and API integration, representing substantial progress in aligning LLM capabilities with the demands of practical software engineering applications.

Instructing the model to generate a sequence of intermediate steps, a.k.a., a chain of thought (CoT), is a highly effective method to improve the accuracy of large language models (LLMs) on arithmetics and symbolic reasoning tasks. However, the mechanism behind CoT remains unclear. This work provides a theoretical understanding of the power of CoT for decoder-only transformers through the lens of expressiveness. Conceptually, CoT empowers the model with the ability to perform inherently serial computation, which is otherwise lacking in transformers, especially when depth is low. Given input length $n$, previous works have shown that constant-depth transformers with finite precision $\mathsf{poly}(n)$ embedding size can only solve problems in $\mathsf{TC}^0$ without CoT. We first show an even tighter expressiveness upper bound for constant-depth transformers with constant-bit precision, which can only solve problems in $\mathsf{AC}^0$, a proper subset of $ \mathsf{TC}^0$. However, with $T$ steps of CoT, constant-depth transformers using constant-bit precision and $O(\log n)$ embedding size can solve any problem solvable by boolean circuits of size $T$. Empirically, enabling CoT dramatically improves the accuracy for tasks that are hard for parallel computation, including the composition of permutation groups, iterated squaring, and circuit value problems, especially for low-depth transformers.

Curriculum reinforcement learning (CRL) aims to speed up learning by gradually increasing the difficulty of a task, usually quantified by the achievable expected return. Inspired by the success of CRL in single-agent settings, a few works have attempted to apply CRL to multi-agent reinforcement learning (MARL) using the number of agents to control task difficulty. However, existing works typically use manually defined curricula such as a linear scheme. In this paper, we first apply state-of-the-art single-agent self-paced CRL to sparse reward MARL. Although with satisfying performance, we identify two potential flaws of the curriculum generated by existing reward-based CRL methods: (1) tasks with high returns may not provide informative learning signals and (2) the exacerbated credit assignment difficulty in tasks where more agents yield higher returns. Thereby, we further propose self-paced MARL (SPMARL) to prioritize tasks based on \textit{learning progress} instead of the episode return. Our method not only outperforms baselines in three challenging sparse-reward benchmarks but also converges faster than self-paced CRL.

A fundamental challenge in multi-agent reinforcement learning (MARL) is to learn the joint policy in an extremely large search space, which grows exponentially with the number of agents. Moreover, fully decentralized policy factorization significantly restricts the search space, which may lead to sub-optimal policies. In contrast, the auto-regressive joint policy can represent a much richer class of joint policies by factorizing the joint policy into the product of a series of conditional individual policies. While such factorization introduces the action dependency among agents explicitly in sequential execution, it does not take full advantage of the dependency during learning. In particular, the subsequent agents do not give the preceding agents feedback about their decisions. In this paper, we propose a new framework Back-Propagation Through Agents (BPTA) that directly accounts for both agents' own policy updates and the learning of their dependent counterparts. This is achieved by propagating the feedback through action chains. With the proposed framework, our Bidirectional Proximal Policy Optimisation (BPPO) outperforms the state-of-the-art methods. Extensive experiments on matrix games, StarCraftII v2, Multi-agent MuJoCo, and Google Research Football demonstrate the effectiveness of the proposed method.

Centralized training with decentralized execution (CTDE) is widely employed to stabilize partially observable multi-agent reinforcement learning (MARL) by utilizing a centralized value function during training. However, existing methods typically assume that agents make decisions based on their local observations independently, which may not lead to a correlated joint policy with sufficient coordination. Inspired by the concept of correlated equilibrium, we propose to introduce a \textit{strategy modification} to provide a mechanism for agents to correlate their policies. Specifically, we present a novel framework, AgentMixer, which constructs the joint fully observable policy as a non-linear combination of individual partially observable policies. To enable decentralized execution, one can derive individual policies by imitating the joint policy. Unfortunately, such imitation learning can lead to \textit{asymmetric learning failure} caused by the mismatch between joint policy and individual policy information. To mitigate this issue, we jointly train the joint policy and individual policies and introduce \textit{Individual-Global-Consistency} to guarantee mode consistency between the centralized and decentralized policies. We then theoretically prove that AgentMixer converges to an $\epsilon$-approximate Correlated Equilibrium. The strong experimental performance on three MARL benchmarks demonstrates the effectiveness of our method.

The integration of different imaging modalities, such as structural, diffusion tensor, and functional magnetic resonance imaging, with deep learning models has yielded promising outcomes in discerning phenotypic characteristics and enhancing disease diagnosis. The development of such a technique hinges on the efficient fusion of heterogeneous multimodal features, which initially reside within distinct representation spaces. Naively fusing the multimodal features does not adequately capture the complementary information and could even produce redundancy. In this work, we present a novel joint self-supervised and supervised contrastive learning method to learn the robust latent feature representation from multimodal MRI data, allowing the projection of heterogeneous features into a shared common space, and thereby amalgamating both complementary and analogous information across various modalities and among similar subjects. We performed a comparative analysis between our proposed method and alternative deep multimodal learning approaches. Through extensive experiments on two independent datasets, the results demonstrated that our method is significantly superior to several other deep multimodal learning methods in predicting abnormal neurodevelopment. Our method has the capability to facilitate computer-aided diagnosis within clinical practice, harnessing the power of multimodal data.

Recent work by Power et al. (2022) highlighted a surprising "grokking" phenomenon in learning arithmetic tasks: a neural net first "memorizes" the training set, resulting in perfect training accuracy but near-random test accuracy, and after training for sufficiently longer, it suddenly transitions to perfect test accuracy. This paper studies the grokking phenomenon in theoretical setups and shows that it can be induced by a dichotomy of early and late phase implicit biases. Specifically, when training homogeneous neural nets with large initialization and small weight decay on both classification and regression tasks, we prove that the training process gets trapped at a solution corresponding to a kernel predictor for a long time, and then a very sharp transition to min-norm/max-margin predictors occurs, leading to a dramatic change in test accuracy. The generalization behavior of modern over-parameterized neural nets has been puzzling: these nets have the capacity to overfit the training set, and yet they frequently exhibit a small gap between training and test performance when trained by popular gradient-based optimizers. A common view now is that the network architectures and training pipelines can automatically induce regularization effects to avoid or mitigate overfitting throughout the training trajectory. Recently, Power et al. (2022) discovered an even more perplexing generalization phenomenon called grokking: when training a neural net to learn modular arithmetic operations, it first "memorizes" the training set with zero training error and near-random test error, and then training for much longer leads to a sharp transition from no generalization to perfect generalization. See Section 2 for our reproduction of this phenomenon. Beyond modular arithmetic, grokking has been reported in learning group operations (Chughtai et al., 2023), learning sparse parity (Barak et al., 2022; Bhattamishra et al., 2022), learning greatest common divisor (Charton, 2023), and image classification (Liu et al., 2023; Radhakrishnan et al., 2022).

Stochastic Variance Reduced Gradient (SVRG), introduced by Johnson & Zhang (2013), is a theoretically compelling optimization method. However, as Defazio & Bottou (2019) highlights, its effectiveness in deep learning is yet to be proven. In this work, we demonstrate the potential of SVRG in optimizing real-world neural networks. Our analysis finds that, for deeper networks, the strength of the variance reduction term in SVRG should be smaller and decrease as training progresses. Inspired by this, we introduce a multiplicative coefficient $\alpha$ to control the strength and adjust it through a linear decay schedule. We name our method $\alpha$-SVRG. Our results show $\alpha$-SVRG better optimizes neural networks, consistently reducing training loss compared to both baseline and the standard SVRG across various architectures and image classification datasets. We hope our findings encourage further exploration into variance reduction techniques in deep learning. Code is available at https://github.com/davidyyd/alpha-SVRG.

The goal of automatic report generation is to generate a clinically accurate and coherent phrase from a single given X-ray image, which could alleviate the workload of traditional radiology reporting. However, in a real-world scenario, radiologists frequently face the challenge of producing extensive reports derived from numerous medical images, thereby medical report generation from multi-image perspective is needed. In this paper, we propose the Complex Organ Mask Guided (termed as COMG) report generation model, which incorporates masks from multiple organs (e.g., bones, lungs, heart, and mediastinum), to provide more detailed information and guide the model's attention to these crucial body regions. Specifically, we leverage prior knowledge of the disease corresponding to each organ in the fusion process to enhance the disease identification phase during the report generation process. Additionally, cosine similarity loss is introduced as target function to ensure the convergence of cross-modal consistency and facilitate model optimization.Experimental results on two public datasets show that COMG achieves a 11.4% and 9.7% improvement in terms of BLEU@4 scores over the SOTA model KiUT on IU-Xray and MIMIC, respectively. The code is publicly available at https://github.com/GaryGuTC/COMG_model.