Joint Demonstration and Preference Learning Improves Policy Alignment with Human Feedback

Aligning human preference and value is an important requirement for building contemporary foundation models and embodied AI. However, popular approaches such as reinforcement learning with human feedback (RLHF) break down the task into successive stages, such as supervised fine-tuning (SFT), reward modeling (RM), and reinforcement learning (RL), each performing one specific learning task. Such a sequential approach results in serious issues such as significant under-utilization of data and distribution mismatch between the learned reward model and generated policy, which eventually lead to poor alignment performance. We develop a single stage approach named Alignment with Integrated Human Feedback (AIHF), capable of integrating both human preference and demonstration to train reward models and the policy. The proposed approach admits a suite of efficient algorithms, which can easily reduce to, and leverage, popular alignment algorithms such as RLHF and Directly Policy Optimization (DPO), and only requires minor changes to the existing alignment pipelines. We demonstrate the efficiency of the proposed solutions with extensive experiments involving alignment problems in LLMs and robotic control problems in MuJoCo. We observe that the proposed solutions outperform the existing alignment algorithms such as RLHF and DPO by large margins, especially when the amount of high-quality preference data is relatively limited.

SLTrain: a sparse plus low-rank approach for parameter and memory efficient pretraining

Large language models (LLMs) have shown impressive capabilities across various tasks. However, training LLMs from scratch requires significant computational power and extensive memory capacity. Recent studies have explored low-rank structures on weights for efficient fine-tuning in terms of parameters and memory, either through low-rank adaptation or factorization. While effective for fine-tuning, low-rank structures are generally less suitable for pretraining because they restrict parameters to a low-dimensional subspace. In this work, we propose to parameterize the weights as a sum of low-rank and sparse matrices for pretraining, which we call SLTrain. The low-rank component is learned via matrix factorization, while for the sparse component, we employ a simple strategy of uniformly selecting the sparsity support at random and learning only the non-zero entries with the fixed support. While being simple, the random fixed-support sparse learning strategy significantly enhances pretraining when combined with low-rank learning. Our results show that SLTrain adds minimal extra parameters and memory costs compared to pretraining with low-rank parameterization, yet achieves substantially better performance, which is comparable to full-rank training. Remarkably, when combined with quantization and per-layer updates, SLTrain can reduce memory requirements by up to 73% when pretraining the LLaMA 7B model.

Getting More Juice Out of the SFT Data: Reward Learning from Human Demonstration Improves SFT for LLM Alignment

Aligning human preference and value is an important requirement for contemporary foundation models. State-of-the-art techniques such as Reinforcement Learning from Human Feedback (RLHF) often consist of two stages: 1) supervised fine-tuning (SFT), where the model is fine-tuned by learning from human demonstration data; 2) Preference learning, where preference data is used to learn a reward model, which is in turn used by a reinforcement learning (RL) step to fine-tune the model. Such reward model serves as a proxy to human preference, and it is critical to guide the RL step towards improving the model quality. In this work, we argue that the SFT stage significantly benefits from learning a reward model as well. Instead of using the human demonstration data directly via supervised learning, we propose to leverage an Inverse Reinforcement Learning (IRL) technique to (explicitly or implicitly) build an reward model, while learning the policy model. This approach leads to new SFT algorithms that are not only efficient to implement, but also promote the ability to distinguish between the preferred and non-preferred continuations. Moreover, we identify a connection between the proposed IRL based approach, and certain self-play approach proposed recently, and showed that self-play is a special case of modeling a reward-learning agent. Theoretically, we show that the proposed algorithms converge to the stationary solutions of the IRL problem. Empirically, we align 1B and 7B models using proposed methods and evaluate them on a reward benchmark model and the HuggingFace Open LLM Leaderboard. The proposed methods show significant performance improvement over existing SFT approaches. Our results indicate that it is beneficial to explicitly or implicitly leverage reward learning throughout the entire alignment process.

SA-Attack: Speed-adaptive stealthy adversarial attack on trajectory prediction

Trajectory prediction is critical for the safe planning and navigation of automated vehicles. The trajectory prediction models based on the neural networks are vulnerable to adversarial attacks. Previous attack methods have achieved high attack success rates but overlook the adaptability to realistic scenarios and the concealment of the deceits. To address this problem, we propose a speed-adaptive stealthy adversarial attack method named SA-Attack. This method searches the sensitive region of trajectory prediction models and generates the adversarial trajectories by using the vehicle-following method and incorporating information about forthcoming trajectories. Our method has the ability to adapt to different speed scenarios by reconstructing the trajectory from scratch. Fusing future trajectory trends and curvature constraints can guarantee the smoothness of adversarial trajectories, further ensuring the stealthiness of attacks. The empirical study on the datasets of nuScenes and Apolloscape demonstrates the attack performance of our proposed method. Finally, we also demonstrate the adaptability and stealthiness of SA-Attack for different speed scenarios. Our code is available at the repository: https://github.com/eclipse-bot/SA-Attack.

Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark

In the evolving landscape of natural language processing (NLP), fine-tuning pre-trained Large Language Models (LLMs) with first-order (FO) optimizers like SGD and Adam has become standard. Yet, as LLMs grow {in size}, the substantial memory overhead from back-propagation (BP) for FO gradient computation presents a significant challenge. Addressing this issue is crucial, especially for applications like on-device training where memory efficiency is paramount. This paper proposes a shift towards BP-free, zeroth-order (ZO) optimization as a solution for reducing memory costs during LLM fine-tuning, building on the initial concept introduced by MeZO. Unlike traditional ZO-SGD methods, our work expands the exploration to a wider array of ZO optimization techniques, through a comprehensive, first-of-its-kind benchmarking study across five LLM families (Roberta, OPT, LLaMA, Vicuna, Mistral), three task complexities, and five fine-tuning schemes. Our study unveils previously overlooked optimization principles, highlighting the importance of task alignment, the role of the forward gradient method, and the balance between algorithm complexity and fine-tuning performance. We further introduce novel enhancements to ZO optimization, including block-wise descent, hybrid training, and gradient sparsity. Our study offers a promising direction for achieving further memory-efficient LLM fine-tuning. Codes to reproduce all our experiments are at https://github.com/ZO-Bench/ZO-LLM .

Zeroth-order Riemannian Averaging Stochastic Approximation Algorithms

We present Zeroth-order Riemannian Averaging Stochastic Approximation (\texttt{Zo-RASA}) algorithms for stochastic optimization on Riemannian manifolds. We show that \texttt{Zo-RASA} achieves optimal sample complexities for generating $\epsilon$-approximation first-order stationary solutions using only one-sample or constant-order batches in each iteration. Our approach employs Riemannian moving-average stochastic gradient estimators, and a novel Riemannian-Lyapunov analysis technique for convergence analysis. We improve the algorithm's practicality by using retractions and vector transport, instead of exponential mappings and parallel transports, thereby reducing per-iteration complexity. Additionally, we introduce a novel geometric condition, satisfied by manifolds with bounded second fundamental form, which enables new error bounds for approximating parallel transport with vector transport.

A Riemannian ADMM

We consider a class of Riemannian optimization problems where the objective is the sum of a smooth function and a nonsmooth function, considered in the ambient space. This class of problems finds important applications in machine learning and statistics such as the sparse principal component analysis, sparse spectral clustering, and orthogonal dictionary learning. We propose a Riemannian alternating direction method of multipliers (ADMM) to solve this class of problems. Our algorithm adopts easily computable steps in each iteration. The iteration complexity of the proposed algorithm for obtaining an $\epsilon$-stationary point is analyzed under mild assumptions. To the best of our knowledge, this is the first Riemannian ADMM with provable convergence guarantee for solving Riemannian optimization problem with nonsmooth objective. Numerical experiments are conducted to demonstrate the advantage of the proposed method.

Solving combinational optimization problems with evolutionary single-pixel imaging

Single-pixel imaging (SPI) is a novel optical imaging technique by replacing the pixelated sensor array in a conventional camera with a single-pixel detector. In previous works, SPI is usually used for capturing object images or performing image processing tasks. In this work, we propose a SPI scheme for processing other types of data in addition to images. An Ising machine model is implemented optically with SPI for solving combinational optimization problems including number partition and graph maximum cut. Simulated and experimental results show that our proposed scheme can optimize the Hamiltonian function with evolutionary illumination patterns.

Federated Learning on Riemannian Manifolds

Federated learning (FL) has found many important applications in smart-phone-APP based machine learning applications. Although many algorithms have been studied for FL, to the best of our knowledge, algorithms for FL with nonconvex constraints have not been studied. This paper studies FL over Riemannian manifolds, which finds important applications such as federated PCA and federated kPCA. We propose a Riemannian federated SVRG (RFedSVRG) method to solve federated optimization over Riemannian manifolds. We analyze its convergence rate under different scenarios. Numerical experiments are conducted to compare RFedSVRG with the Riemannian counterparts of FedAvg and FedProx. We observed from the numerical experiments that the advantages of RFedSVRG are significant.

Playing Tic-Tac-Toe Games with Intelligent Single-pixel Imaging

Single-pixel imaging (SPI) is a novel optical imaging technique by replacing a two-dimensional pixelated sensor with a single-pixel detector and pattern illuminations. SPI have been extensively used for various tasks related to image acquisition and processing. In this work, a novel non-image-based task of playing Tic-Tac-Toe games interactively is merged into the framework of SPI. An optoelectronic artificial intelligent (AI) player with minimal digital computation can detect the game states, generate optimal moves and display output results mainly by pattern illumination and single-pixel detection. Simulated and experimental results demonstrate the feasibility of proposed scheme and its unbeatable performance against human players.