Learning Code Preference via Synthetic Evolution

Large Language Models (LLMs) have recently demonstrated remarkable coding capabilities. However, assessing code generation based on well-formed properties and aligning it with developer preferences remains challenging. In this paper, we explore two key questions under the new challenge of code preference learning: (i) How do we train models to predict meaningful preferences for code? and (ii) How do human and LLM preferences align with verifiable code properties and developer code tastes? To this end, we propose CodeFavor, a framework for training pairwise code preference models from synthetic evolution data, including code commits and code critiques. To evaluate code preferences, we introduce CodePrefBench, a benchmark comprising 1364 rigorously curated code preference tasks to cover three verifiable properties-correctness, efficiency, and security-along with human preference. Our evaluation shows that CodeFavor holistically improves the accuracy of model-based code preferences by up to 28.8%. Meanwhile, CodeFavor models can match the performance of models with 6-9x more parameters while being 34x more cost-effective. We also rigorously validate the design choices in CodeFavor via a comprehensive set of controlled experiments. Furthermore, we discover the prohibitive costs and limitations of human-based code preference: despite spending 23.4 person-minutes on each task, 15.1-40.3% of tasks remain unsolved. Compared to model-based preference, human preference tends to be more accurate under the objective of code correctness, while being sub-optimal for non-functional objectives.

Mitigating Adversarial Perturbations for Deep Reinforcement Learning via Vector Quantization

Recent studies reveal that well-performing reinforcement learning (RL) agents in training often lack resilience against adversarial perturbations during deployment. This highlights the importance of building a robust agent before deploying it in the real world. Most prior works focus on developing robust training-based procedures to tackle this problem, including enhancing the robustness of the deep neural network component itself or adversarially training the agent on strong attacks. In this work, we instead study an input transformation-based defense for RL. Specifically, we propose using a variant of vector quantization (VQ) as a transformation for input observations, which is then used to reduce the space of adversarial attacks during testing, resulting in the transformed observations being less affected by attacks. Our method is computationally efficient and seamlessly integrates with adversarial training, further enhancing the robustness of RL agents against adversarial attacks. Through extensive experiments in multiple environments, we demonstrate that using VQ as the input transformation effectively defends against adversarial attacks on the agent's observations.

On the Perturbed States for Transformed Input-robust Reinforcement Learning

Reinforcement Learning (RL) agents demonstrating proficiency in a training environment exhibit vulnerability to adversarial perturbations in input observations during deployment. This underscores the importance of building a robust agent before its real-world deployment. To alleviate the challenging point, prior works focus on developing robust training-based procedures, encompassing efforts to fortify the deep neural network component's robustness or subject the agent to adversarial training against potent attacks. In this work, we propose a novel method referred to as Transformed Input-robust RL (TIRL), which explores another avenue to mitigate the impact of adversaries by employing input transformation-based defenses. Specifically, we introduce two principles for applying transformation-based defenses in learning robust RL agents: (1) autoencoder-styled denoising to reconstruct the original state and (2) bounded transformations (bit-depth reduction and vector quantization (VQ)) to achieve close transformed inputs. The transformations are applied to the state before feeding it into the policy network. Extensive experiments on multiple MuJoCo environments demonstrate that input transformation-based defenses, i.e., VQ, defend against several adversaries in the state observations. The official code is available at https://github.com/tunglm2203/tirl

Active Learning for Fair and Stable Online Allocations

We explore an active learning approach for dynamic fair resource allocation problems. Unlike previous work that assumes full feedback from all agents on their allocations, we consider feedback from a select subset of agents at each epoch of the online resource allocation process. Despite this restriction, our proposed algorithms provide regret bounds that are sub-linear in number of time-periods for various measures that include fairness metrics commonly used in resource allocation problems and stability considerations in matching mechanisms. The key insight of our algorithms lies in adaptively identifying the most informative feedback using dueling upper and lower confidence bounds. With this strategy, we show that efficient decision-making does not require extensive feedback and produces efficient outcomes for a variety of problem classes.

Towards Robust Policy: Enhancing Offline Reinforcement Learning with Adversarial Attacks and Defenses

Offline reinforcement learning (RL) addresses the challenge of expensive and high-risk data exploration inherent in RL by pre-training policies on vast amounts of offline data, enabling direct deployment or fine-tuning in real-world environments. However, this training paradigm can compromise policy robustness, leading to degraded performance in practical conditions due to observation perturbations or intentional attacks. While adversarial attacks and defenses have been extensively studied in deep learning, their application in offline RL is limited. This paper proposes a framework to enhance the robustness of offline RL models by leveraging advanced adversarial attacks and defenses. The framework attacks the actor and critic components by perturbing observations during training and using adversarial defenses as regularization to enhance the learned policy. Four attacks and two defenses are introduced and evaluated on the D4RL benchmark. The results show the vulnerability of both the actor and critic to attacks and the effectiveness of the defenses in improving policy robustness. This framework holds promise for enhancing the reliability of offline RL models in practical scenarios.

DimCL: Dimensional Contrastive Learning For Improving Self-Supervised Learning

Self-supervised learning (SSL) has gained remarkable success, for which contrastive learning (CL) plays a key role. However, the recent development of new non-CL frameworks has achieved comparable or better performance with high improvement potential, prompting researchers to enhance these frameworks further. Assimilating CL into non-CL frameworks has been thought to be beneficial, but empirical evidence indicates no visible improvements. In view of that, this paper proposes a strategy of performing CL along the dimensional direction instead of along the batch direction as done in conventional contrastive learning, named Dimensional Contrastive Learning (DimCL). DimCL aims to enhance the feature diversity, and it can serve as a regularizer to prior SSL frameworks. DimCL has been found to be effective, and the hardness-aware property is identified as a critical reason for its success. Extensive experimental results reveal that assimilating DimCL into SSL frameworks leads to performance improvement by a non-trivial margin on various datasets and backbone architectures.

LAD: A Hybrid Deep Learning System for Benign Paroxysmal Positional Vertigo Disorders Diagnostic

Herein, we introduce "Look and Diagnose" (LAD), a hybrid deep learning-based system that aims to support doctors in the medical field in diagnosing effectively the Benign Paroxysmal Positional Vertigo (BPPV) disorder. Given the body postures of the patient in the Dix-Hallpike and lateral head turns test, the visual information of both eyes is captured and fed into LAD for analyzing and classifying into one of six possible disorders the patient might be suffering from. The proposed system consists of two streams: (1) an RNN-based stream that takes raw RGB images of both eyes to extract visual features and optical flow of each eye followed by ternary classification to determine left/right posterior canal (PC) or other; and (2) pupil detector stream that detects the pupil when it is classified as Non-PC and classifies the direction and strength of the beating to categorize the Non-PC types into the remaining four classes: Geotropic BPPV (left and right) and Apogeotropic BPPV (left and right). Experimental results show that with the patient's body postures, the system can accurately classify given BPPV disorder into the six types of disorders with an accuracy of 91% on the validation set. The proposed method can successfully classify disorders with an accuracy of 93% for the Posterior Canal disorder and 95% for the Geotropic and Apogeotropic disorder, paving a potential direction for research with the medical data.

SoftGroup++: Scalable 3D Instance Segmentation with Octree Pyramid Grouping

Existing state-of-the-art 3D point cloud instance segmentation methods rely on a grouping-based approach that groups points to obtain object instances. Despite improvement in producing accurate segmentation results, these methods lack scalability and commonly require dividing large input into multiple parts. To process a scene with millions of points, the existing fastest method SoftGroup \cite{vu2022softgroup} requires tens of seconds, which is under satisfaction. Our finding is that $k$-Nearest Neighbor ($k$-NN), which serves as the prerequisite of grouping, is a computational bottleneck. This bottleneck severely worsens the inference time in the scene with a large number of points. This paper proposes SoftGroup++ to address this computational bottleneck and further optimize the inference speed of the whole network. SoftGroup++ is built upon SoftGroup, which differs in three important aspects: (1) performs octree $k$-NN instead of vanilla $k$-NN to reduce time complexity from $\mathcal{O}(n^2)$ to $\mathcal{O}(n \log n)$, (2) performs pyramid scaling that adaptively downsamples backbone outputs to reduce search space for $k$-NN and grouping, and (3) performs late devoxelization that delays the conversion from voxels to points towards the end of the model such that intermediate components operate at a low computational cost. Extensive experiments on various indoor and outdoor datasets demonstrate the efficacy of the proposed SoftGroup++. Notably, SoftGroup++ processes large scenes of millions of points by a single forward without dividing the input into multiple parts, thus enriching contextual information. Especially, SoftGroup++ achieves 2.4 points AP$_{50}$ improvement while nearly $6\times$ faster than the existing fastest method on S3DIS dataset. The code and trained models will be made publicly available.

Few-shot Object Counting and Detection

We tackle a new task of few-shot object counting and detection. Given a few exemplar bounding boxes of a target object class, we seek to count and detect all objects of the target class. This task shares the same supervision as the few-shot object counting but additionally outputs the object bounding boxes along with the total object count. To address this challenging problem, we introduce a novel two-stage training strategy and a novel uncertainty-aware few-shot object detector: Counting-DETR. The former is aimed at generating pseudo ground-truth bounding boxes to train the latter. The latter leverages the pseudo ground-truth provided by the former but takes the necessary steps to account for the imperfection of pseudo ground-truth. To validate the performance of our method on the new task, we introduce two new datasets named FSCD-147 and FSCD-LVIS. Both datasets contain images with complex scenes, multiple object classes per image, and a huge variation in object shapes, sizes, and appearance. Our proposed approach outperforms very strong baselines adapted from few-shot object counting and few-shot object detection with a large margin in both counting and detection metrics. The code and models are available at https://github.com/VinAIResearch/Counting-DETR.

Beyond NaN: Resiliency of Optimization Layers in The Face of Infeasibility

Prior work has successfully incorporated optimization layers as the last layer in neural networks for various problems, thereby allowing joint learning and planning in one neural network forward pass. In this work, we identify a weakness in such a set-up where inputs to the optimization layer lead to undefined output of the neural network. Such undefined decision outputs can lead to possible catastrophic outcomes in critical real time applications. We show that an adversary can cause such failures by forcing rank deficiency on the matrix fed to the optimization layer which results in the optimization failing to produce a solution. We provide a defense for the failure cases by controlling the condition number of the input matrix. We study the problem in the settings of synthetic data, Jigsaw Sudoku, and in speed planning for autonomous driving, building on top of prior frameworks in end-to-end learning and optimization. We show that our proposed defense effectively prevents the framework from failing with undefined output. Finally, we surface a number of edge cases which lead to serious bugs in popular equation and optimization solvers which can be abused as well.