We are currently in the "large-scale era" of machine learning (ML), where the exciting capabilities of

A well-known numerical bottleneck in the differentially-private stochastic gradient descent (DP-SGD) algorithm is the computation of the gradient norm for each example in a large input batch. When the loss function in DP-SGD consists of an intermediate linear operation, existing methods in the literature have proposed decompositions of gradients that are amenable to fast norm computations. In this paper, we present a framework that generalizes the above approach to arbitrary (possibly nonlinear) intermediate operations. Moreover, we show that for certain operations, such as fully-connected and embedding layer computations, further improvements to the runtime and storage costs of existing decompositions can be deduced using certain components of our framework. Finally, preliminary numerical experiments are given to demonstrate the substantial effects of the aforementioned improvements.

Abstract--Future networks are envisioned to connect massive artificial intelligence (AI) agents, enabling their extensive collaboration on diverse tasks. Compared to traditional entities, these agents naturally suit the semantic communication (SC), which can significantly enhance the bandwidth efficiency. Nevertheless, SC requires the knowledge among agents to be aligned, while agents have distinct expert knowledge for their individual tasks in practice. In this paper, we propose a distillation-enabled knowledge alignment protocol (DeKAP), which distills the expert knowledge of each agent into parameter-efficient low-rank matrices, allocates them across the network, and allows agents to simultaneously maintain aligned knowledge for multiple tasks. We formulate the joint minimization of alignment loss, communication overhead, and storage cost as a large-scale integer linear programming problem and develop a highly efficient greedy algorithm. From computer simulation, the DeKAP establishes knowledge alignment with the lowest communication and computation resources compared to conventional approaches. Future communication networks will usher in a new era of the "Internet of Intelligence," where human beings, devices, and a wide range of artificial intelligence (AI) agents are seamlessly interconnected [1].

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by retrieving relevant documents from external sources and incorporating them into the context. While it improves reliability by providing factual texts, it significantly increases inference costs as context length grows and introduces challenging issue of RAG hallucination, primarily caused by the lack of corresponding parametric knowledge in LLMs. An efficient solution is to enhance the knowledge of LLMs at test-time. Parametric RAG (PRAG) addresses this by embedding document into LLMs parameters to perform test-time knowledge enhancement, effectively reducing inference costs through offline training. However, its high training and storage costs, along with limited generalization ability, significantly restrict its practical adoption. To address these challenges, we propose Dynamic Parametric RAG (DyPRAG), a novel framework that leverages a lightweight parameter translator model to efficiently convert documents into parametric knowledge. DyPRAG not only reduces inference, training, and storage costs but also dynamically generates parametric knowledge, seamlessly enhancing the knowledge of LLMs and resolving knowledge conflicts in a plug-and-play manner at test-time. Extensive experiments on multiple datasets demonstrate the effectiveness and generalization capabilities of DyPRAG, offering a powerful and practical RAG paradigm which enables superior knowledge fusion and mitigates RAG hallucination in real-world applications. Our code is available at https://github.com/Trae1ounG/DyPRAG.

Deploying LLMs on edge devices presents serious technical challenges. Memory elasticity is crucial for edge devices with unified memory, where memory is shared and fluctuates dynamically. Existing solutions suffer from either poor transition granularity or high storage costs. We propose FlexQuant, a novel elasticity framework that generates an ensemble of quantized models, providing an elastic hosting solution with 15x granularity improvement and 10x storage reduction compared to SoTA methods. FlexQuant works with most quantization methods and creates a family of trade-off options under various storage limits through our pruning method. It brings great performance and flexibility to the edge deployment of LLMs.

With the advent of billion-parameter foundation models, efficient fine-tuning has become increasingly important for the adaptation of models to downstream tasks. However, especially in computer vision, it can be hard to achieve good performance when access to quality labeled data is lacking. In this work, we propose a method adapting pretrained generalist models in a self-supervised manner by learning binary masks. These self-supervised masking networks (SMNs) are up to 79x more efficient to store and significantly improve performance on label-efficient downstream tasks. We validate the usefulness of learning binary masks as a fine-tuning method on 8 datasets and 3 model architectures, and we demonstrate the effectiveness of SMNs in 3 label-efficient settings.

Large language models (LLMs) have recently demonstrated promising performance in many tasks. However, the high storage and computational cost of LLMs has become a challenge for deploying LLMs. Weight quantization has been widely used for model compression, which can reduce both storage and computational cost. Most existing weight quantization methods for LLMs use a rank-one codebook for quantization, which results in substantial accuracy loss when the compression ratio is high. In this paper, we propose a novel weight quantization method, called low-rank codebook based quantization (LCQ), for LLMs. LCQ adopts a low-rank codebook, the rank of which can be larger than one, for quantization. Experiments show that LCQ can achieve better accuracy than existing methods with a negligibly extra storage cost.

We study a new class of MDPs that employs multinomial logit (MNL) function approximation to ensure valid probability distributions over the state space. Despite its benefits, introducing non-linear function approximation raises significant challenges in both computational and statistical efficiency. The best-known method of Hwang and Oh [2023] has achieved an $\widetilde{\mathcal{O}}(\kappa^{-1}dH^2\sqrt{K})$ regret, where $\kappa$ is a problem-dependent quantity, $d$ is the feature space dimension, $H$ is the episode length, and $K$ is the number of episodes. While this result attains the same rate in $K$ as the linear cases, the method requires storing all historical data and suffers from an $\mathcal{O}(K)$ computation cost per episode. Moreover, the quantity $\kappa$ can be exponentially small, leading to a significant gap for the regret compared to the linear cases. In this work, we first address the computational concerns by proposing an online algorithm that achieves the same regret with only $\mathcal{O}(1)$ computation cost. Then, we design two algorithms that leverage local information to enhance statistical efficiency. They not only maintain an $\mathcal{O}(1)$ computation cost per episode but achieve improved regrets of $\widetilde{\mathcal{O}}(\kappa^{-1/2}dH^2\sqrt{K})$ and $\widetilde{\mathcal{O}}(dH^2\sqrt{K} + \kappa^{-1}d^2H^2)$ respectively. Finally, we establish a lower bound, justifying the optimality of our results in $d$ and $K$. To the best of our knowledge, this is the first work that achieves almost the same computational and statistical efficiency as linear function approximation while employing non-linear function approximation for reinforcement learning.