Fine-tuning pre-trained Large Language Models (LLMs) is essential to align them with human values and intentions. This process often utilizes methods like pairwise comparisons and KL divergence against a reference LLM, focusing on the evaluation of full answers generated by the models. However, the generation of these responses occurs in a token level, following a sequential, auto-regressive fashion. In this paper, we introduce Token-level Direct Preference Optimization (TDPO), a novel approach to align LLMs with human preferences by optimizing policy at the token level. Unlike previous methods, which face challenges in divergence efficiency, TDPO incorporates forward KL divergence constraints for each token, improving alignment and diversity. Utilizing the Bradley-Terry model for a token-based reward system, TDPO enhances the regulation of KL divergence, while preserving simplicity without the need for explicit reward modeling. Experimental results across various text tasks demonstrate TDPO's superior performance in balancing alignment with generation diversity. Notably, fine-tuning with TDPO strikes a better balance than DPO in the controlled sentiment generation and single-turn dialogue datasets, and significantly improves the quality of generated responses compared to both DPO and PPO-based RLHF methods. Our code is open-sourced at https://github.com/Vance0124/Token-level-Direct-Preference-Optimization.
Temporal misalignment (time offset) between sensors is common in low cost visual-inertial odometry (VIO) systems. Such temporal misalignment introduces inconsistent constraints for state estimation, leading to a significant positioning drift especially in high dynamic motion scenarios. In this article, we focus on online temporal calibration to reduce the positioning drift caused by the time offset for high dynamic motion VIO. For the time offset observation model, most existing methods rely on accurate state estimation or stable visual tracking. For the prediction model, current methods oversimplify the time offset as a constant value with white Gaussian noise. However, these ideal conditions are seldom satisfied in real high dynamic scenarios, resulting in the poor performance. In this paper, we introduce online time offset modeling networks (TON) to enhance real-time temporal calibration. TON improves the accuracy of time offset observation and prediction modeling. Specifically, for observation modeling, we propose feature velocity observation networks to enhance velocity computation for features in unstable visual tracking conditions. For prediction modeling, we present time offset prediction networks to learn its evolution pattern. To highlight the effectiveness of our method, we integrate the proposed TON into both optimization-based and filter-based VIO systems. Simulation and real-world experiments are conducted to demonstrate the enhanced performance of our approach. Additionally, to contribute to the VIO community, we will open-source the code of our method on: https://github.com/Franky-X/FVON-TPN.
In recent years, Neural Radiance Fields (NeRFs) have demonstrated significant potential in encoding highly-detailed 3D geometry and environmental appearance, positioning themselves as a promising alternative to traditional explicit representation for 3D scene reconstruction. However, the predominant reliance on RGB imaging presupposes ideal lighting conditions: a premise frequently unmet in robotic applications plagued by poor lighting or visual obstructions. This limitation overlooks the capabilities of infrared (IR) cameras, which excel in low-light detection and present a robust alternative under such adverse scenarios. To tackle these issues, we introduce Thermal-NeRF, the first method that estimates a volumetric scene representation in the form of a NeRF solely from IR imaging. By leveraging a thermal mapping and structural thermal constraint derived from the thermal characteristics of IR imaging, our method showcasing unparalleled proficiency in recovering NeRFs in visually degraded scenes where RGB-based methods fall short. We conduct extensive experiments to demonstrate that Thermal-NeRF can achieve superior quality compared to existing methods. Furthermore, we contribute a dataset for IR-based NeRF applications, paving the way for future research in IR NeRF reconstruction.
In reinforcement Learning (RL), an instant reward signal is generated for each action of the agent, such that the agent learns to maximize the cumulative reward to obtain the optimal policy. However, in many real-world applications, the instant reward signals are not obtainable by the agent. Instead, the learner only obtains rewards at the ends of bags, where a bag is defined as a partial sequence of a complete trajectory. In this situation, the learner has to face the significant difficulty of exploring the unknown instant rewards in the bags, which could not be addressed by existing approaches, including those trajectory-based approaches that consider only complete trajectories and ignore the inner reward distributions. To formally study this situation, we introduce a novel RL setting termed Reinforcement Learning from Bagged Rewards (RLBR), where only the bagged rewards of sequences can be obtained. We provide the theoretical study to establish the connection between RLBR and standard RL in Markov Decision Processes (MDPs). To effectively explore the reward distributions within the bagged rewards, we propose a Transformer-based reward model, the Reward Bag Transformer (RBT), which uses the self-attention mechanism for interpreting the contextual nuances and temporal dependencies within each bag. Extensive experimental analyses demonstrate the superiority of our method, particularly in its ability to mimic the original MDP's reward distribution, highlighting its proficiency in contextual understanding and adaptability to environmental dynamics.
AI for drug discovery has been a research hotspot in recent years, and SMILES-based language models has been increasingly applied in drug molecular design. However, no work has explored whether and how language models understand the chemical spatial structure from 1D sequences. In this work, we pre-train a transformer model on chemical language and fine-tune it toward drug design objectives, and investigate the correspondence between high-frequency SMILES substrings and molecular fragments. The results indicate that language models can understand chemical structures from the perspective of molecular fragments, and the structural knowledge learned through fine-tuning is reflected in the high-frequency SMILES substrings generated by the model.
De novo drug design is a pivotal issue in pharmacology and a new area of focus in AI for science research. A central challenge in this field is to generate molecules with specific properties while also producing a wide range of diverse candidates. Although advanced technologies such as transformer models and reinforcement learning have been applied in drug design, their potential has not been fully realized. Therefore, we propose MolRL-MGPT, a reinforcement learning algorithm with multiple GPT agents for drug molecular generation. To promote molecular diversity, we encourage the agents to collaborate in searching for desirable molecules in diverse directions. Our algorithm has shown promising results on the GuacaMol benchmark and exhibits efficacy in designing inhibitors against SARS-CoV-2 protein targets. The codes are available at: https://github.com/HXYfighter/MolRL-MGPT.
Occlusions between consecutive frames have long posed a significant challenge in optical flow estimation. The inherent ambiguity introduced by occlusions directly violates the brightness constancy constraint and considerably hinders pixel-to-pixel matching. To address this issue, multi-frame optical flow methods leverage adjacent frames to mitigate the local ambiguity. Nevertheless, prior multi-frame methods predominantly adopt recursive flow estimation, resulting in a considerable computational overlap. In contrast, we propose a streamlined in-batch framework that eliminates the need for extensive redundant recursive computations while concurrently developing effective spatio-temporal modeling approaches under in-batch estimation constraints. Specifically, we present a Streamlined In-batch Multi-frame (SIM) pipeline tailored to video input, attaining a similar level of time efficiency to two-frame networks. Furthermore, we introduce an efficient Integrative Spatio-temporal Coherence (ISC) modeling method for effective spatio-temporal modeling during the encoding phase, which introduces no additional parameter overhead. Additionally, we devise a Global Temporal Regressor (GTR) that effectively explores temporal relations during decoding. Benefiting from the efficient SIM pipeline and effective modules, StreamFlow not only excels in terms of performance on the challenging KITTI and Sintel datasets, with particular improvement in occluded areas but also attains a remarkable $63.82\%$ enhancement in speed compared with previous multi-frame methods. The code will be available soon at https://github.com/littlespray/StreamFlow.
Skin lesion segmentation is of great significance for skin lesion analysis and subsequent treatment. It is still a challenging task due to the irregular and fuzzy lesion borders, and diversity of skin lesions. In this paper, we propose Triple-UNet to automatically segment skin lesions. It is an organic combination of three UNet architectures with suitable modules. In order to concatenate the first and second sub-networks more effectively, we design a region of interest enhancement module (ROIE). The ROIE enhances the target object region of the image by using the predicted score map of the first UNet. The features learned by the first UNet and the enhanced image help the second UNet obtain a better score map. Finally, the results are fine-tuned by the third UNet. We evaluate our algorithm on a publicly available dataset of skin lesion segmentation. Experiments show that Triple-UNet outperforms the state-of-the-art on skin lesion segmentation.
In this paper, we propose an Invariant Extended Kalman Filter (IEKF) based Visual-Inertial Odometry (VIO) using multiple features in man-made environments. Conventional EKF-based VIO usually suffers from system inconsistency and angular drift that naturally occurs in feature-based methods. However, in man-made environments, notable structural regularities, such as lines and vanishing points, offer valuable cues for localization. To exploit these structural features effectively and maintain system consistency, we design a right invariant filter-based VIO scheme incorporating point, line, and vanishing point features. We demonstrate that the conventional additive error definition for point features can also preserve system consistency like the invariant error definition by proving a mathematically equivalent measurement model. And a similar conclusion is established for line features. Additionally, we conduct an invariant filter-based observability analysis proving that vanishing point measurement maintains unobservable directions naturally. Both simulation and real-world tests are conducted to validate our methods' pose accuracy and consistency. The experimental results validate the competitive performance of our method, highlighting its ability to deliver accurate and consistent pose estimation in man-made environments.
The planning of how to synthesize molecules, also known as retrosynthesis, has been a growing focus of the machine learning and chemistry communities in recent years. Despite the appearance of steady progress, we argue that imperfect benchmarks and inconsistent comparisons mask systematic shortcomings of existing techniques. To remedy this, we present a benchmarking library called syntheseus which promotes best practice by default, enabling consistent meaningful evaluation of single-step and multi-step retrosynthesis algorithms. We use syntheseus to re-evaluate a number of previous retrosynthesis algorithms, and find that the ranking of state-of-the-art models changes when evaluated carefully. We end with guidance for future works in this area.