Large Language Models (LLMs) are difficult to fully fine-tune (e.g., with instructions or human feedback) due to their sheer number of parameters. A family of parameter-efficient sparse fine-tuning (SFT) methods have proven promising in terms of performance but their memory requirements increase proportionally to the size of the LLMs. In this work, we scale sparse fine-tuning to state-of-the-art LLMs like LLaMA 2 7B and 13B. At any given time, for a desired density level, we maintain an array of parameter indices and the deltas of these parameters relative to their pretrained values. We iterate among: (a) updating the active deltas, (b) pruning indices (based on the change of magnitude of their deltas) and (c) regrowth of indices. For regrowth, we explore two criteria based on either the accumulated gradients of a few candidate parameters or their approximate momenta estimated using the efficient SM3 optimizer. We experiment with instruction-tuning of LLMs on standard dataset mixtures, finding that SFT is often superior to popular parameter-efficient fine-tuning methods like LoRA (low-rank adaptation) in terms of performance and comparable in terms of run time. We additionally show that SFT is compatible with both quantization and efficient optimizers, to facilitate scaling to ever-larger model sizes. We release the code for SFT at //github.com/AlanAnsell/peft and for the instruction-tuning experiments at //github.com/ducdauge/sft-llm.
相關內容
In the Network Revenue Management (NRM) problem, products composed of up to L resources are sold to stochastically arriving customers. We take a randomized rounding approach to NRM, motivated by developments in Online Contention Resolution Schemes (OCRS). The goal is to take a fractional solution to NRM that satisfies the resource constraints in expectation, and implement it in an online policy that satisfies the resource constraints in any state, while (approximately) preserving all of the sales that were prescribed by the fractional solution. OCRS cannot be naively applied to NRM or revenue management problems in general, because customer substitution induces a negative correlation in products being demanded. We start by deriving an OCRS that achieves a guarantee of 1/(1+L) for NRM with customer substitution, matching a common benchmark in the literature. We then show how to beat this benchmark for all integers L>1 assuming no substitution, i.e., in the standard OCRS setting. By contrast, we show that this benchmark is unbeatable using OCRS or any fractional relaxation if there is customer substitution, for all integers L that are the power of a prime number. Finally, we show how to beat 1/(1+L) even with customer substitution, if the products comprise one item from each of up to L groups. Our results have corresponding implications for Online Combinatorial Auctions, in which buyers bid for bundles of up to L items, and buyers being single-minded is akin to no substitution. Our final result also beats 1/(1+L) for Prophet Inequality on the intersection of L partition matroids. All in all, our paper provides a unifying framework for applying OCRS to these problems, delineating the impact of substitution, and establishing a separation between the guarantees achievable with vs. without substitution under general resource constraints parametrized by L.
Although person or identity verification has been predominantly explored using individual modalities such as face and voice, audio-visual fusion has recently shown immense potential to outperform unimodal approaches. Audio and visual modalities are often expected to pose strong complementary relationships, which plays a crucial role in effective audio-visual fusion. However, they may not always strongly complement each other, they may also exhibit weak complementary relationships, resulting in poor audio-visual feature representations. In this paper, we propose a Dynamic Cross-Attention (DCA) model that can dynamically select the cross-attended or unattended features on the fly based on the strong or weak complementary relationships, respectively, across audio and visual modalities. In particular, a conditional gating layer is designed to evaluate the contribution of the cross-attention mechanism and choose cross-attended features only when they exhibit strong complementary relationships, otherwise unattended features. Extensive experiments are conducted on the Voxceleb1 dataset to demonstrate the robustness of the proposed model. Results indicate that the proposed model consistently improves the performance on multiple variants of cross-attention while outperforming the state-of-the-art methods.
We propose a variant of Hamiltonian Monte Carlo (HMC), called the Repelling-Attracting Hamiltonian Monte Carlo (RAHMC), for sampling from multimodal distributions. The key idea that underpins RAHMC is a departure from the conservative dynamics of Hamiltonian systems, which form the basis of traditional HMC, and turning instead to the dissipative dynamics of conformal Hamiltonian systems. In particular, RAHMC involves two stages: a mode-repelling stage to encourage the sampler to move away from regions of high probability density; and, a mode-attracting stage, which facilitates the sampler to find and settle near alternative modes. We achieve this by introducing just one additional tuning parameter -- the coefficient of friction. The proposed method adapts to the geometry of the target distribution, e.g., modes and density ridges, and can generate proposals that cross low-probability barriers with little to no computational overhead in comparison to traditional HMC. Notably, RAHMC requires no additional information about the target distribution or memory of previously visited modes. We establish the theoretical basis for RAHMC, and we discuss repelling-attracting extensions to several variants of HMC in literature. Finally, we provide a tuning-free implementation via dual-averaging, and we demonstrate its effectiveness in sampling from, both, multimodal and unimodal distributions in high dimensions.
Large Language Models (LLMs) have demonstrated exceptional performance in biochemical tasks, especially the molecule caption translation task, which aims to bridge the gap between molecules and natural language texts. However, previous methods in adapting LLMs to the molecule-caption translation task required extra domain-specific pre-training stages, suffered weak alignment between molecular and textual spaces, or imposed stringent demands on the scale of LLMs. To resolve the challenges, we propose In-Context Molecule Adaptation (ICMA), as a new paradigm allowing LLMs to learn the molecule-text alignment from context examples via In-Context Molecule Tuning. Specifically, ICMA incorporates the following three stages: Cross-modal Retrieval, Post-retrieval Re-ranking, and In-context Molecule Tuning. Initially, Cross-modal Retrieval utilizes BM25 Caption Retrieval and Molecule Graph Retrieval to retrieve informative context examples. Additionally, we also propose Post-retrieval Re-ranking with Sequence Reversal and Random Walk to further improve the quality of retrieval results. Finally, In-Context Molecule Tuning unlocks the in-context molecule learning capability of LLMs with retrieved examples and adapts the parameters of LLMs for the molecule-caption translation task. Experimental results demonstrate that ICMT can empower LLMs to achieve state-of-the-art or comparable performance without extra training corpora and intricate structures, showing that LLMs are inherently in-context molecule learners.
This work considers a repeated principal-agent bandit game, where the principal can only interact with her environment through the agent. The principal and the agent have misaligned objectives and the choice of action is only left to the agent. However, the principal can influence the agent's decisions by offering incentives which add up to his rewards. The principal aims to iteratively learn an incentive policy to maximize her own total utility. This framework extends usual bandit problems and is motivated by several practical applications, such as healthcare or ecological taxation, where traditionally used mechanism design theories often overlook the learning aspect of the problem. We present nearly optimal (with respect to a horizon $T$) learning algorithms for the principal's regret in both multi-armed and linear contextual settings. Finally, we support our theoretical guarantees through numerical experiments.
Ensuring safety in Reinforcement Learning (RL), typically framed as a Constrained Markov Decision Process (CMDP), is crucial for real-world exploration applications. Current approaches in handling CMDP struggle to balance optimality and feasibility, as direct optimization methods cannot ensure state-wise in-training safety, and projection-based methods correct actions inefficiently through lengthy iterations. To address these challenges, we propose Adaptive Chance-constrained Safeguards (ACS), an adaptive, model-free safe RL algorithm using the safety recovery rate as a surrogate chance constraint to iteratively ensure safety during exploration and after achieving convergence. Theoretical analysis indicates that the relaxed probabilistic constraint sufficiently guarantees forward invariance to the safe set. And extensive experiments conducted on both simulated and real-world safety-critical tasks demonstrate its effectiveness in enforcing safety (nearly zero-violation) while preserving optimality (+23.8%), robustness, and fast response in stochastic real-world settings.
Cyber threats continue to evolve in complexity, thereby traditional Cyber Threat Intelligence (CTI) methods struggle to keep pace. AI offers a potential solution, automating and enhancing various tasks, from data ingestion to resilience verification. This paper explores the potential of integrating Artificial Intelligence (AI) into CTI. We provide a blueprint of an AI-enhanced CTI processing pipeline, and detail its components and functionalities. The pipeline highlights the collaboration of AI and human expertise, which is necessary to produce timely and high-fidelity cyber threat intelligence. We also explore the automated generation of mitigation recommendations, harnessing AI's capabilities to provide real-time, contextual, and predictive insights. However, the integration of AI into CTI is not without challenges. Thereby, we discuss ethical dilemmas, potential biases, and the imperative for transparency in AI-driven decisions. We address the need for data privacy, consent mechanisms, and the potential misuse of technology. Moreover, we highlights the importance of addressing biases both during CTI analysis and AI models warranting their transparency and interpretability. Lastly, our work points out future research directions such as the exploration of advanced AI models to augment cyber defences, and the human-AI collaboration optimization. Ultimately, the fusion of AI with CTI appears to hold significant potential in cybersecurity domain.
Large Language Models (LLMs) have emerged as powerful tools in the field of Natural Language Processing (NLP) and have recently gained significant attention in the domain of Recommendation Systems (RS). These models, trained on massive amounts of data using self-supervised learning, have demonstrated remarkable success in learning universal representations and have the potential to enhance various aspects of recommendation systems by some effective transfer techniques such as fine-tuning and prompt tuning, and so on. The crucial aspect of harnessing the power of language models in enhancing recommendation quality is the utilization of their high-quality representations of textual features and their extensive coverage of external knowledge to establish correlations between items and users. To provide a comprehensive understanding of the existing LLM-based recommendation systems, this survey presents a taxonomy that categorizes these models into two major paradigms, respectively Discriminative LLM for Recommendation (DLLM4Rec) and Generative LLM for Recommendation (GLLM4Rec), with the latter being systematically sorted out for the first time. Furthermore, we systematically review and analyze existing LLM-based recommendation systems within each paradigm, providing insights into their methodologies, techniques, and performance. Additionally, we identify key challenges and several valuable findings to provide researchers and practitioners with inspiration.
Connecting Vision and Language plays an essential role in Generative Intelligence. For this reason, in the last few years, a large research effort has been devoted to image captioning, i.e. the task of describing images with syntactically and semantically meaningful sentences. Starting from 2015 the task has generally been addressed with pipelines composed of a visual encoding step and a language model for text generation. During these years, both components have evolved considerably through the exploitation of object regions, attributes, and relationships and the introduction of multi-modal connections, fully-attentive approaches, and BERT-like early-fusion strategies. However, regardless of the impressive results obtained, research in image captioning has not reached a conclusive answer yet. This work aims at providing a comprehensive overview and categorization of image captioning approaches, from visual encoding and text generation to training strategies, used datasets, and evaluation metrics. In this respect, we quantitatively compare many relevant state-of-the-art approaches to identify the most impactful technical innovations in image captioning architectures and training strategies. Moreover, many variants of the problem and its open challenges are analyzed and discussed. The final goal of this work is to serve as a tool for understanding the existing state-of-the-art and highlighting the future directions for an area of research where Computer Vision and Natural Language Processing can find an optimal synergy.
Graph Neural Networks (GNNs) have recently become increasingly popular due to their ability to learn complex systems of relations or interactions arising in a broad spectrum of problems ranging from biology and particle physics to social networks and recommendation systems. Despite the plethora of different models for deep learning on graphs, few approaches have been proposed thus far for dealing with graphs that present some sort of dynamic nature (e.g. evolving features or connectivity over time). In this paper, we present Temporal Graph Networks (TGNs), a generic, efficient framework for deep learning on dynamic graphs represented as sequences of timed events. Thanks to a novel combination of memory modules and graph-based operators, TGNs are able to significantly outperform previous approaches being at the same time more computationally efficient. We furthermore show that several previous models for learning on dynamic graphs can be cast as specific instances of our framework. We perform a detailed ablation study of different components of our framework and devise the best configuration that achieves state-of-the-art performance on several transductive and inductive prediction tasks for dynamic graphs.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191