Efficiency for Free: Ideal Data Are Transportable Representations

Data, the seminal opportunity and challenge in modern machine learning, currently constrains the scalability of representation learning and impedes the pace of model evolution. Existing paradigms tackle the issue of learning efficiency over massive datasets from the perspective of self-supervised learning and dataset distillation independently, while neglecting the untapped potential of accelerating representation learning from an intermediate standpoint. In this work, we delve into defining the ideal data properties from both optimization and generalization perspectives. We propose that model-generated representations, despite being trained on diverse tasks and architectures, converge to a shared linear space, facilitating effective linear transport between models. Furthermore, we demonstrate that these representations exhibit properties conducive to the formation of ideal data. The theoretical/empirical insights therein inspire us to propose a Representation Learning Accelerator (ReLA), which leverages a task- and architecture-agnostic, yet publicly available, free model to form a dynamic data subset and thus accelerate (self-)supervised learning. For instance, employing a CLIP ViT B/16 as a prior model for dynamic data generation, ReLA-aided BYOL can train a ResNet-50 from scratch with 50% of ImageNet-1K, yielding performance surpassing that of training on the full dataset. Additionally, employing a ResNet-18 pre-trained on CIFAR-10 can enhance ResNet-50 training on 10% of ImageNet-1K, resulting in a 7.7% increase in accuracy.

Dynamic Mixture of Experts: An Auto-Tuning Approach for Efficient Transformer Models

The Sparse Mixture of Experts (SMoE) has been widely employed to enhance the efficiency of training and inference for Transformer-based foundational models, yielding promising results. However, the performance of SMoE heavily depends on the choice of hyper-parameters, such as the number of experts and the number of experts to be activated (referred to as top-k), resulting in significant computational overhead due to the extensive model training by searching over various hyper-parameter configurations. As a remedy, we introduce the Dynamic Mixture of Experts (DynMoE) technique. DynMoE incorporates (1) a novel gating method that enables each token to automatically determine the number of experts to activate. (2) An adaptive process automatically adjusts the number of experts during training. Extensive numerical results across Vision, Language, and Vision-Language tasks demonstrate the effectiveness of our approach to achieve competitive performance compared to GMoE for vision and language tasks, and MoE-LLaVA for vision-language tasks, while maintaining efficiency by activating fewer parameters. Our code is available at https://github.com/LINs-lab/DynMoE.

Open-Source AI-based SE Tools: Opportunities and Challenges of Collaborative Software Learning

Large Language Models (LLMs) have become instrumental in advancing software engineering (SE) tasks, showcasing their efficacy in code understanding and beyond. Like traditional SE tools, open-source collaboration is key in realising the excellent products. However, with AI models, the essential need is in data. The collaboration of these AI-based SE models hinges on maximising the sources of high-quality data. However, data especially of high quality, often holds commercial or sensitive value, making it less accessible for open-source AI-based SE projects. This reality presents a significant barrier to the development and enhancement of AI-based SE tools within the software engineering community. Therefore, researchers need to find solutions for enabling open-source AI-based SE models to tap into resources by different organisations. Addressing this challenge, our position paper investigates one solution to facilitate access to diverse organizational resources for open-source AI models, ensuring privacy and commercial sensitivities are respected. We introduce a governance framework centered on federated learning (FL), designed to foster the joint development and maintenance of open-source AI code models while safeguarding data privacy and security. Additionally, we present guidelines for developers on AI-based SE tool collaboration, covering data requirements, model architecture, updating strategies, and version control. Given the significant influence of data characteristics on FL, our research examines the effect of code data heterogeneity on FL performance.

DeFT: Flash Tree-attention with IO-Awareness for Efficient Tree-search-based LLM Inference

Decoding using tree search can greatly enhance the inference quality for transformer-based Large Language Models (LLMs). Depending on the guidance signal, it searches for the best path from root to leaf in the tree by forming LLM outputs to improve controllability, reasoning ability, alignment, et cetera. However, current tree decoding strategies and their inference systems do not suit each other well due to redundancy in computation, memory footprints, and memory access, resulting in inefficient inference. To address this issue, we propose DeFT, an IO-aware tree attention algorithm that maintains memory-efficient attention calculation with low memory footprints in two stages: (1) QKV Preparation: we propose a KV-Guided Tree Split strategy to group QKV wisely for high utilization of GPUs and reduction of memory reads/writes for the KV cache between GPU global memory and on-chip shared memory as much as possible; (2) Attention Calculation: we calculate partial attention of each QKV groups in a fused kernel then apply a Tree-topology-aware Global Reduction strategy to get final attention. Thanks to a reduction in KV cache IO by 3.6-4.5$\times$, along with an additional reduction in IO for $\mathbf{Q} \mathbf{K}^\top$ and Softmax equivalent to 25% of the total KV cache IO, DeFT can achieve a speedup of 1.7-2.4$\times$ in end-to-end latency across two practical reasoning tasks over the SOTA attention algorithms.

Persuading a Learning Agent

We study a repeated Bayesian persuasion problem (and more generally, any generalized principal-agent problem with complete information) where the principal does not have commitment power and the agent uses algorithms to learn to respond to the principal's signals. We reduce this problem to a one-shot generalized principal-agent problem with an approximately-best-responding agent. This reduction allows us to show that: if the agent uses contextual no-regret learning algorithms, then the principal can guarantee a utility that is arbitrarily close to the principal's optimal utility in the classic non-learning model with commitment; if the agent uses contextual no-swap-regret learning algorithms, then the principal cannot obtain any utility significantly more than the optimal utility in the non-learning model with commitment. The difference between the principal's obtainable utility in the learning model and the non-learning model is bounded by the agent's regret (swap-regret). If the agent uses mean-based learning algorithms (which can be no-regret but not no-swap-regret), then the principal can do significantly better than the non-learning model. These conclusions hold not only for Bayesian persuasion, but also for any generalized principal-agent problem with complete information, including Stackelberg games and contract design.

Switch EMA: A Free Lunch for Better Flatness and Sharpness

Exponential Moving Average (EMA) is a widely used weight averaging (WA) regularization to learn flat optima for better generalizations without extra cost in deep neural network (DNN) optimization. Despite achieving better flatness, existing WA methods might fall into worse final performances or require extra test-time computations. This work unveils the full potential of EMA with a single line of modification, i.e., switching the EMA parameters to the original model after each epoch, dubbed as Switch EMA (SEMA). From both theoretical and empirical aspects, we demonstrate that SEMA can help DNNs to reach generalization optima that better trade-off between flatness and sharpness. To verify the effectiveness of SEMA, we conduct comparison experiments with discriminative, generative, and regression tasks on vision and language datasets, including image classification, self-supervised learning, object detection and segmentation, image generation, video prediction, attribute regression, and language modeling. Comprehensive results with popular optimizers and networks show that SEMA is a free lunch for DNN training by improving performances and boosting convergence speeds.

Multi-Sender Persuasion -- A Computational Perspective

We consider multiple senders with informational advantage signaling to convince a single self-interested actor towards certain actions. Generalizing the seminal Bayesian Persuasion framework, such settings are ubiquitous in computational economics, multi-agent learning, and machine learning with multiple objectives. The core solution concept here is the Nash equilibrium of senders' signaling policies. Theoretically, we prove that finding an equilibrium in general is PPAD-Hard; in fact, even computing a sender's best response is NP-Hard. Given these intrinsic difficulties, we turn to finding local Nash equilibria. We propose a novel differentiable neural network to approximate this game's non-linear and discontinuous utilities. Complementing this with the extra-gradient algorithm, we discover local equilibria that Pareto dominates full-revelation equilibria and those found by existing neural networks. Broadly, our theoretical and empirical contributions are of interest to a large class of economic problems.

PathMMU: A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology

The emergence of large multimodal models has unlocked remarkable potential in AI, particularly in pathology. However, the lack of specialized, high-quality benchmark impeded their development and precise evaluation. To address this, we introduce PathMMU, the largest and highest-quality expert-validated pathology benchmark for LMMs. It comprises 33,573 multimodal multi-choice questions and 21,599 images from various sources, and an explanation for the correct answer accompanies each question. The construction of PathMMU capitalizes on the robust capabilities of GPT-4V, utilizing approximately 30,000 gathered image-caption pairs to generate Q\&As. Significantly, to maximize PathMMU's authority, we invite six pathologists to scrutinize each question under strict standards in PathMMU's validation and test sets, while simultaneously setting an expert-level performance benchmark for PathMMU. We conduct extensive evaluations, including zero-shot assessments of 14 open-sourced and three closed-sourced LMMs and their robustness to image corruption. We also fine-tune representative LMMs to assess their adaptability to PathMMU. The empirical findings indicate that advanced LMMs struggle with the challenging PathMMU benchmark, with the top-performing LMM, GPT-4V, achieving only a 51.7\% zero-shot performance, significantly lower than the 71.4\% demonstrated by human pathologists. After fine-tuning, even open-sourced LMMs can surpass GPT-4V with a performance of over 60\%, but still fall short of the expertise shown by pathologists. We hope that the PathMMU will offer valuable insights and foster the development of more specialized, next-generation LLMs for pathology.

Federated Unlearning: a Perspective of Stability and Fairness

This paper explores the multifaceted consequences of federated unlearning (FU) with data heterogeneity. We introduce key metrics for FU assessment, concentrating on verification, global stability, and local fairness, and investigate the inherent trade-offs. Furthermore, we formulate the unlearning process with data heterogeneity through an optimization framework. Our key contribution lies in a comprehensive theoretical analysis of the trade-offs in FU and provides insights into data heterogeneity's impacts on FU. Leveraging these insights, we propose FU mechanisms to manage the trade-offs, guiding further development for FU mechanisms. We empirically validate that our FU mechanisms effectively balance trade-offs, confirming insights derived from our theoretical analysis.

Training-time Neuron Alignment through Permutation Subspace for Improving Linear Mode Connectivity and Model Fusion

In deep learning, stochastic gradient descent often yields functionally similar yet widely scattered solutions in the weight space even under the same initialization, causing barriers in the Linear Mode Connectivity (LMC) landscape. Overcoming these barriers is crucial for understanding deep learning dynamics and enhancing model-fusion algorithms. Previous studies highlight the role of permutation symmetry in reducing post-training barriers through network permutation. However, these post-hoc methods, demanding extra computations, are less effective for larger, complex models (e.g., ViT, LLM) due to numerous permutation matrices. Thus, in this paper, we study training-time neuron alignment. Our hypothesis suggests that training-time permutation subspace can reduce LMC barriers for free. We find that pruning at initialization supports this. Beyond pruning, we introduce TNA-PFN, a simple yet lossless algorithm using a partial gradient mask during training. TNA-PFN is theoretically and empirically validated for reducing LMC barriers. It excels in wide model fusion applications, especially in federated learning, two algorithms based on TNA-FPN that are proposed to show its prospects even under heterogeneous datasets. Moreover, TNA-PFN can enhance the generalization of model soup for vision transformers and ColD fusion for pretrained language models.