Jie Liu

Zhanhui Zhou, Jie Liu, Jing Shao et al.

A single language model, even when aligned with labelers through reinforcement learning from human feedback (RLHF), may not suit all human preferences. Recent approaches therefore prefer customization, gathering multi-dimensional feedback, and creating distinct reward models for each dimension. Different language models are then optimized for various preferences using multi-objective RLHF (MORLHF) with varying reward weights. However, RL fine-tuning is unstable and resource-heavy, especially with diverse and usually conflicting objectives. In this paper, we present Multi-Objective Direct Preference Optimization (MODPO), an RL-free extension of Direct Preference Optimization (DPO) for multiple alignment objectives. Essentially, MODPO folds language modeling directly into reward modeling, training language models as implicit collective reward models that combine all objectives with specific weights. MODPO theoretically yields the same optimal solutions as MORLHF but is practically more stable and efficient. Empirical results in safety alignment and long-form question answering show that MODPO matches or outperforms existing methods, producing a Pareto front of language models catering to diverse preferences with three times less computational resources compared to MORLHF. Code is available at https://github.com/ZHZisZZ/modpo.

Chuming Li, Jie Liu, Yinmin Zhang et al.

Multi-agent reinforcement learning (MARL) suffers from the non-stationarity problem, which is the ever-changing targets at every iteration when multiple agents update their policies at the same time. Starting from first principle, in this paper, we manage to solve the non-stationarity problem by proposing bidirectional action-dependent Q-learning (ACE). Central to the development of ACE is the sequential decision-making process wherein only one agent is allowed to take action at one time. Within this process, each agent maximizes its value function given the actions taken by the preceding agents at the inference stage. In the learning phase, each agent minimizes the TD error that is dependent on how the subsequent agents have reacted to their chosen action. Given the design of bidirectional dependency, ACE effectively turns a multiagent MDP into a single-agent MDP. We implement the ACE framework by identifying the proper network representation to formulate the action dependency, so that the sequential decision process is computed implicitly in one forward pass. To validate ACE, we compare it with strong baselines on two MARL benchmarks. Empirical experiments demonstrate that ACE outperforms the state-of-the-art algorithms on Google Research Football and StarCraft Multi-Agent Challenge by a large margin. In particular, on SMAC tasks, ACE achieves 100% success rate on almost all the hard and super-hard maps. We further study extensive research problems regarding ACE, including extension, generalization, and practicability. Code is made available to facilitate further research.

Letian Wang, Jie Liu, Hao Shao et al.

When autonomous vehicles are deployed on public roads, they will encounter countless and diverse driving situations. Many manually designed driving policies are difficult to scale to the real world. Fortunately, reinforcement learning has shown great success in many tasks by automatic trial and error. However, when it comes to autonomous driving in interactive dense traffic, RL agents either fail to learn reasonable performance or necessitate a large amount of data. Our insight is that when humans learn to drive, they will 1) make decisions over the high-level skill space instead of the low-level control space and 2) leverage expert prior knowledge rather than learning from scratch. Inspired by this, we propose ASAP-RL, an efficient reinforcement learning algorithm for autonomous driving that simultaneously leverages motion skills and expert priors. We first parameterized motion skills, which are diverse enough to cover various complex driving scenarios and situations. A skill parameter inverse recovery method is proposed to convert expert demonstrations from control space to skill space. A simple but effective double initialization technique is proposed to leverage expert priors while bypassing the issue of expert suboptimality and early performance degradation. We validate our proposed method on interactive dense-traffic driving tasks given simple and sparse rewards. Experimental results show that our method can lead to higher learning efficiency and better driving performance relative to previous methods that exploit skills and priors differently. Code is open-sourced to facilitate further research.

Dimitrios Stamoulis, Ruizhou Ding, Di Wang et al.

Can we reduce the search cost of Neural Architecture Search (NAS) from days down to only few hours? NAS methods automate the design of Convolutional Networks (ConvNets) under hardware constraints and they have emerged as key components of AutoML frameworks. However, the NAS problem remains challenging due to the combinatorially large design space and the significant search time (at least 200 GPU-hours). In this work, we alleviate the NAS search cost down to less than 3 hours, while achieving state-of-the-art image classification results under mobile latency constraints. We propose a novel differentiable NAS formulation, namely Single-Path NAS, that uses one single-path over-parameterized ConvNet to encode all architectural decisions based on shared convolutional kernel parameters, hence drastically decreasing the search overhead. Single-Path NAS achieves state-of-the-art top-1 ImageNet accuracy (75.62%), hence outperforming existing mobile NAS methods in similar latency settings (~80ms). In particular, we enhance the accuracy-runtime trade-off in differentiable NAS by treating the Squeeze-and-Excitation path as a fully searchable operation with our novel single-path encoding. Our method has an overall cost of only 8 epochs (24 TPU-hours), which is up to 5,000x faster compared to prior work. Moreover, we study how different NAS formulation choices affect the performance of the designed ConvNets. Furthermore, we exploit the efficiency of our method to answer an interesting question: instead of empirically tuning the hyperparameters of the NAS solver (as in prior work), can we automatically find the hyperparameter values that yield the desired accuracy-runtime trade-off? We open-source our entire codebase at: https://github.com/dstamoulis/single-path-nas.

Yinmin Zhang, Jie Liu, Chuming Li et al.

Offline-to-online Reinforcement Learning (O2O RL) aims to improve the performance of offline pretrained policy using only a few online samples. Built on offline RL algorithms, most O2O methods focus on the balance between RL objective and pessimism, or the utilization of offline and online samples. In this paper, from a novel perspective, we systematically study the challenges that remain in O2O RL and identify that the reason behind the slow improvement of the performance and the instability of online finetuning lies in the inaccurate Q-value estimation inherited from offline pretraining. Specifically, we demonstrate that the estimation bias and the inaccurate rank of Q-value cause a misleading signal for the policy update, making the standard offline RL algorithms, such as CQL and TD3-BC, ineffective in the online finetuning. Based on this observation, we address the problem of Q-value estimation by two techniques: (1) perturbed value update and (2) increased frequency of Q-value updates. The first technique smooths out biased Q-value estimation with sharp peaks, preventing early-stage policy exploitation of sub-optimal actions. The second one alleviates the estimation bias inherited from offline pretraining by accelerating learning. Extensive experiments on the MuJoco and Adroit environments demonstrate that the proposed method, named SO2, significantly alleviates Q-value estimation issues, and consistently improves the performance against the state-of-the-art methods by up to 83.1%.

Kecheng Chen, Xinyu Luo, Tiexin Qin et al.

Foundation medical segmentation models, with MedSAM being the most popular, have achieved promising performance across organs and lesions. However, MedSAM still suffers from compromised performance on specific lesions with intricate structures and appearance, as well as bounding box prompt-induced perturbations. Although current test-time adaptation (TTA) methods for medical image segmentation may tackle this issue, partial (e.g., batch normalization) or whole parametric updates restrict their effectiveness due to limited update signals or catastrophic forgetting in large models. Meanwhile, these approaches ignore the computational complexity during adaptation, which is particularly significant for modern foundation models. To this end, our theoretical analyses reveal that directly refining image embeddings is feasible to approach the same goal as parametric updates under the MedSAM architecture, which enables us to realize high computational efficiency and segmentation performance without the risk of catastrophic forgetting. Under this framework, we propose to encourage maximizing factorized conditional probabilities of the posterior prediction probability using a proposed distribution-approximated latent conditional random field loss combined with an entropy minimization loss. Experiments show that we achieve about 3\% Dice score improvements across three datasets while reducing computational complexity by over 7 times.

Jie Liu, Pan Zhou, Zehao Xiao et al.

Interactive 3D segmentation has emerged as a promising solution for generating accurate object masks in complex 3D scenes by incorporating user-provided clicks. However, two critical challenges remain underexplored: (1) effectively generalizing from sparse user clicks to produce accurate segmentation, and (2) quantifying predictive uncertainty to help users identify unreliable regions. In this work, we propose NPISeg3D, a novel probabilistic framework that builds upon Neural Processes (NPs) to address these challenges. Specifically, NPISeg3D introduces a hierarchical latent variable structure with scene-specific and object-specific latent variables to enhance few-shot generalization by capturing both global context and object-specific characteristics. Additionally, we design a probabilistic prototype modulator that adaptively modulates click prototypes with object-specific latent variables, improving the model's ability to capture object-aware context and quantify predictive uncertainty. Experiments on four 3D point cloud datasets demonstrate that NPISeg3D achieves superior segmentation performance with fewer clicks while providing reliable uncertainty estimations.

Zhiqiang Cao, Zhijun Li, Pan Heng et al.

As one of the basic tasks of computer vision, object detection has been widely used in many intelligent applications. However, object detection algorithms are usually heavyweight in computation, hindering their implementations on resource-constrained edge devices. Current edge-cloud collaboration methods, such as CNN partition over Edge-cloud devices, are not suitable for object detection since the huge data size of the intermediate results will introduce extravagant communication costs. To address this challenge, we propose a small-big model framework that deploys a big model in the cloud and a small model on the edge devices. Upon receiving data, the edge device operates a difficult-case discriminator to classify the images into easy cases and difficult cases according to the specific semantics of the images. The easy cases will be processed locally at the edge, and the difficult cases will be uploaded to the cloud. Experimental results on the VOC, COCO, HELMET datasets using two different object detection algorithms demonstrate that the small-big model system can detect 94.01%-97.84% of objects with only about 50% images uploaded to the cloud when using SSD. In addition, the small-big model averagely reaches 91.22%- 92.52% end-to-end mAP of the scheme that uploading all images to the cloud.

Jie Liu, Chuming Li, Feng Liang et al.

As a variant of standard convolution, a dilated convolution can control effective receptive fields and handle large scale variance of objects without introducing additional computational costs. To fully explore the potential of dilated convolution, we proposed a new type of dilated convolution (referred to as inception convolution), where the convolution operations have independent dilation patterns among different axes, channels and layers. To develop a practical method for learning complex inception convolution based on the data, a simple but effective search algorithm, referred to as efficient dilation optimization (EDO), is developed. Based on statistical optimization, the EDO method operates in a low-cost manner and is extremely fast when it is applied on large scale datasets. Empirical results validate that our method achieves consistent performance gains for image recognition, object detection, instance segmentation, human detection, and human pose estimation. For instance, by simply replacing the 3x3 standard convolution in the ResNet-50 backbone with inception convolution, we significantly improve the AP of Faster R-CNN from 36.4% to 39.2% on MS COCO.

Jie Liu, Chen Lin, Chuming Li et al.

Several variants of stochastic gradient descent (SGD) have been proposed to improve the learning effectiveness and efficiency when training deep neural networks, among which some recent influential attempts would like to adaptively control the parameter-wise learning rate (e.g., Adam and RMSProp). Although they show a large improvement in convergence speed, most adaptive learning rate methods suffer from compromised generalization compared with SGD. In this paper, we proposed an Adaptive Gradient Method with Resilience and Momentum (AdaRem), motivated by the observation that the oscillations of network parameters slow the training, and give a theoretical proof of convergence. For each parameter, AdaRem adjusts the parameter-wise learning rate according to whether the direction of one parameter changes in the past is aligned with the direction of the current gradient, and thus encourages long-term consistent parameter updating with much fewer oscillations. Comprehensive experiments have been conducted to verify the effectiveness of AdaRem when training various models on a large-scale image recognition dataset, e.g., ImageNet, which also demonstrate that our method outperforms previous adaptive learning rate-based algorithms in terms of the training speed and the test error, respectively.

Dimitrios Stamoulis, Ruizhou Ding, Di Wang et al.

Can we automatically design a Convolutional Network (ConvNet) with the highest image classification accuracy under the latency constraint of a mobile device? Neural Architecture Search (NAS) for ConvNet design is a challenging problem due to the combinatorially large design space and search time (at least 200 GPU-hours). To alleviate this complexity, we propose Single-Path NAS, a novel differentiable NAS method for designing device-efficient ConvNets in less than 4 hours. 1. Novel NAS formulation: our method introduces a single-path, over-parameterized ConvNet to encode all architectural decisions with shared convolutional kernel parameters. 2. NAS efficiency: Our method decreases the NAS search cost down to 8 epochs (30 TPU-hours), i.e., up to 5,000x faster compared to prior work. 3. On-device image classification: Single-Path NAS achieves 74.96% top-1 accuracy on ImageNet with 79ms inference latency on a Pixel 1 phone, which is state-of-the-art accuracy compared to NAS methods with similar latency (<80ms).

Xiao Yan, Xinyan Dai, Jie Liu et al.

Recently, locality sensitive hashing (LSH) was shown to be effective for MIPS and several algorithms including $L_2$-ALSH, Sign-ALSH and Simple-LSH have been proposed. In this paper, we introduce the norm-range partition technique, which partitions the original dataset into sub-datasets containing items with similar 2-norms and builds hash index independently for each sub-dataset. We prove that norm-range partition reduces the query processing complexity for all existing LSH based MIPS algorithms under mild conditions. The key to performance improvement is that norm-range partition allows to use smaller normalization factor most sub-datasets. For efficient query processing, we also formulate a unified framework to rank the buckets from the hash indexes of different sub-datasets. Experiments on real datasets show that norm-range partition significantly reduces the number of probed for LSH based MIPS algorithms when achieving the same recall.

Jie Liu, Martin Takac

We propose a projected semi-stochastic gradient descent method with mini-batch for improving both the theoretical complexity and practical performance of the general stochastic gradient descent method (SGD). We are able to prove linear convergence under weak strong convexity assumption. This requires no strong convexity assumption for minimizing the sum of smooth convex functions subject to a compact polyhedral set, which remains popular across machine learning community. Our PS2GD preserves the low-cost per iteration and high optimization accuracy via stochastic gradient variance-reduced technique, and admits a simple parallel implementation with mini-batches. Moreover, PS2GD is also applicable to dual problem of SVM with hinge loss.