arXiv subjects

Fei Wen

Publications and source records attributed to Fei Wen.

At least 19 recordsLinked to original sources

arXiv2026

Curvature-Aware Zeroth-Order Optimization for Memory-Efficient Test-Time Adaptation

Test-time adaptation (TTA) aims to enhance the cross-domain performance of pre-trained models by adapting to unlabeled test data. While most existing TTA methods rely on backpropagation (BP) for finetuning, BP-free methods such as zeroth-order (ZO) methods are more desired in practical on-device scenarios. ZO methods rely only on forward computation, which can largely reduce the complexity and memory overhead of on-device deployment. However, ZO methods suffer from much higher variance compared with first-order methods in estimating the gradient. To address this, we propose an improved ZO method to substantially boost the performance of ZO optimization based TTA. First, we provide an observation to reveal the persistent low-rank Hessian structure of the loss during the adaptation process. Based on this insight, we then propose a loss-landscape curvature-aware zeroth-order (CAZO) method, which leverages a sliding-average estimation of the diagonal Hessian to construct a covariance matrix for anisotropic perturbation sampling. CAZO operates by freezing pretrained weights and optimizing minimal adapter parameters via forward-only passes based gradient estimation, which can substantially reduce the memory overhead compared to BP-based methods. Extensive experiments demonstrate that CAZO significantly outperforms existing TTA methods, achieving state-of-the-art performance while maintaining an excellent balance between accuracy and memory efficiency. Code is available at https://github.com/Hollyming/CAZO.

cs.CV↗

arXiv2026

Boosting Vision-Language-Action Finetuning with Feasible Action Neighborhood Prior

In real-world robotic manipulation, states typically admit a neighborhood of near-equivalent actions. That is, for each state, there exist a feasible action neighborhood (FAN) rather than a single correct action, within which motions yield indistinguishable progress. However, prevalent VLA training methodologies are directly inherited from linguistic settings and do not exploit the FAN property, thus leading to poor generalization and low sample efficiency. To address this limitation, we introduce a FAN-guided regularizer that shapes the model's output distribution to align with the geometry of FAN. Concretely, we introduce a Gaussian prior that promotes locally smooth and unimodal predictions around the preferred direction and magnitude. In extensive experiments across both reinforced finetuning (RFT) and supervised finetuning (SFT), our method achieves significant improvement in sample efficiency, and success rate in both in-distribution and out-of-distribution (OOD) scenarios. By aligning with the intrinsic action tolerance of physical manipulation, FAN-guided regularization provides a principled and practical method for sample-efficient, and generalizable VLA adaptation.

cs.RO↗

arXiv2026

TRINE: A Token-Aware, Runtime-Adaptive FPGA Inference Engine for Multimodal AI

Multimodal stacks that mix ViTs, CNNs, GNNs, and transformer NLP strain embedded platforms because their compute/memory patterns diverge and hard real-time targets leave little slack. TRINE is a single-bitstream FPGA accelerator and compiler that executes end-to-end multimodal inference without reconfiguration. Layers are unified as DDMM/SDDMM/SpMM and mapped to a mode-switchable engine that toggles at runtime among weight/output-stationary systolic, 1xCS SIMD, and a routable adder tree (RADT) on a shared PE array. A width-matched, two-stage top-k unit enables in-stream token pruning, while dependency-aware layer offloading (DALO) overlaps independent kernels across reconfigurable processing units to sustain utilization. Evaluated on Alveo U50 and ZCU104, TRINE reduces latency by up to 22.57x vs. RTX 4090 and 6.86x vs. Jetson Orin Nano at 20-21 W; token pruning alone yields up to 7.8x on ViT-heavy pipelines, and DALO contributes up to 79% throughput improvement. With int8 quantization, accuracy drops remain <2.5% across representative tasks, delivering state-of-the-art latency and energy efficiency for unified vision, language, and graph workloads-in one bitstream.

cs.AR↗

arXiv2026

SSMG-Nav: Enhancing Lifelong Object Navigation with Semantic Skeleton Memory Graph

Navigating to out-of-sight targets from human instructions in unfamiliar environments is a core capability for service robots. Despite substantial progress, most approaches underutilize reusable, persistent memory, constraining performance in lifelong settings. Many are additionally limited to single-modality inputs and employ myopic greedy policies, which often induce inefficient back-and-forth maneuvers (BFMs). To address such limitations, we introduce SSMG-Nav, a framework for object navigation built on a \textit{Semantic Skeleton Memory Graph} (SSMG) that consolidates past observations into a spatially aligned, persistent memory anchored by topological keypoints (e.g., junctions, room centers). SSMG clusters nearby entities into subgraphs, unifying entity- and space-level semantics to yield a compact set of candidate destinations. To support multimodal targets (images, objects, and text), we integrate a vision-language model (VLM). For each subgraph, a multimodal prompt synthesized from memory guides the VLM to infer a target belief over destinations. A long-horizon planner then trades off this belief against traversability costs to produce a visit sequence that minimizes expected path length, thereby reducing backtracking. Extensive experiments on challenging lifelong benchmarks and standard ObjectNav benchmarks demonstrate that, compared to strong baselines, our method achieves higher success rates and greater path efficiency, validating the effectiveness of SSMG-Nav.

cs.RO↗

arXiv2026

Switchcodec: Adaptive residual-expert sparse quantization for high-fidelity neural audio coding

Recent neural audio compression models often rely on residual vector quantization for high-fidelity coding, but using a fixed number of per-frame codebooks is suboptimal for the wide variability of audio content-especially for signals that are either very simple or highly complex. To address this limitation, we propose SwitchCodec, a neural audio codec based on Residual Experts Vector Quantization (REVQ). REVQ combines a shared quantizer with dynamically routed expert quantizers that are activated according to the input audio, decoupling bitrate from codebook capacity and improving compression efficiency. This design ensures full training and utilization of each quantizer. In addition, a variable-bitrate mechanism adjusts the number of active expert quantizers at inference, enabling multi-bitrate operation without retraining. Experiments demonstrate that SwitchCodec surpasses existing baselines on both objective metrics and subjective listening tests.

cs.SD↗

arXiv2025

ZO-ASR: Zeroth-Order Fine-Tuning of Speech Foundation Models without Back-Propagation

Fine-tuning pre-trained speech foundation models for Automatic Speech Recognition (ASR) is prevalent, yet constrained by substantial GPU memory requirements. We introduce ZO-ASR, a memory-efficient Zeroth-Order (ZO) method that avoids Back-Propagation (BP) and activation memory by estimating gradients via forward passes. When combined with SGD optimizer, ZO-ASR-SGD fine-tunes ASR models using only inference memory. Our evaluation spans supervised and unsupervised tasks. For Supervised Domain Adaptation on Whisper-Large-V3, ZO-ASR's multiple query mechanism enhances robustness and achieves up to an 18.9\% relative Word Error Rate reduction over zero-shot baselines, outperforming existing ZO methods. For unsupervised Test-Time Adaptation on Wav2Vec2-Base, ZO-ASR exhibits moderately lower performance compared to first-order optimizer Adam. Our BP-free approach provides a viable solution for fine-tuning ASR models in computationally resource-constrained or gradient-inaccessible scenarios.

cs.MM↗

arXiv2025

A High-Throughput Spiking Neural Network Processor Enabling Synaptic Delay Emulation

Synaptic delay has attracted significant attention in neural network dynamics for integrating and processing complex spatiotemporal information. This paper introduces a high-throughput Spiking Neural Network (SNN) processor that supports synaptic delay-based emulation for edge applications. The processor leverages a multicore pipelined architecture with parallel compute engines, capable of real-time processing of the computational load associated with synaptic delays. We develop a SoC prototype of the proposed processor on PYNQ Z2 FPGA platform and evaluate its performance using the Spiking Heidelberg Digits (SHD) benchmark for low-power keyword spotting tasks. The processor achieves 93.4% accuracy in deployment and an average throughput of 104 samples/sec at a typical operating frequency of 125 MHz and 282 mW power consumption.

cs.NE↗

arXiv2025

An Online Adaptation Method for Robust Depth Estimation and Visual Odometry in the Open World

Recently, learning-based robotic navigation systems have gained extensive research attention and made significant progress. However, the diversity of open-world scenarios poses a major challenge for the generalization of such systems to practical scenarios. Specifically, learned systems for scene measurement and state estimation tend to degrade when the application scenarios deviate from the training data, resulting to unreliable depth and pose estimation. Toward addressing this problem, this work aims to develop a visual odometry system that can fast adapt to diverse novel environments in an online manner. To this end, we construct a self-supervised online adaptation framework for monocular visual odometry aided by an online-updated depth estimation module. Firstly, we design a monocular depth estimation network with lightweight refiner modules, which enables efficient online adaptation. Then, we construct an objective for self-supervised learning of the depth estimation module based on the output of the visual odometry system and the contextual semantic information of the scene. Specifically, a sparse depth densification module and a dynamic consistency enhancement module are proposed to leverage camera poses and contextual semantics to generate pseudo-depths and valid masks for the online adaptation. Finally, we demonstrate the robustness and generalization capability of the proposed method in comparison with state-of-the-art learning-based approaches on urban, in-house datasets and a robot platform. Code is publicly available at: https://github.com/jixingwu/SOL-SLAM.

cs.RO↗

arXiv2024

A Skeleton-Based Topological Planner for Exploration in Complex Unknown Environments

The capability of autonomous exploration in complex, unknown environments is important in many robotic applications. While recent research on autonomous exploration have achieved much progress, there are still limitations, e.g., existing methods relying on greedy heuristics or optimal path planning are often hindered by repetitive paths and high computational demands. To address such limitations, we propose a novel exploration framework that utilizes the global topology information of observed environment to improve exploration efficiency while reducing computational overhead. Specifically, global information is utilized based on a skeletal topological graph representation of the environment geometry. We first propose an incremental skeleton extraction method based on wavefront propagation, based on which we then design an approach to generate a lightweight topological graph that can effectively capture the environment's structural characteristics. Building upon this, we introduce a finite state machine that leverages the topological structure to efficiently plan coverage paths, which can substantially mitigate the back-and-forth maneuvers (BFMs) problem. Experimental results demonstrate the superiority of our method in comparison with state-of-the-art methods. The source code will be made publicly available at: https://github.com/Haochen-Niu/STGPlanner.

cs.RO↗

arXiv2024

VLTP: Vision-Language Guided Token Pruning for Task-Oriented Segmentation

Vision Transformers (ViTs) have emerged as the backbone of many segmentation models, consistently achieving state-of-the-art (SOTA) performance. However, their success comes at a significant computational cost. Image token pruning is one of the most effective strategies to address this complexity. However, previous approaches fall short when applied to more complex task-oriented segmentation (TOS), where the class of each image patch is not predefined but dependent on the specific input task. This work introduces the Vision Language Guided Token Pruning (VLTP), a novel token pruning mechanism that can accelerate ViT-based segmentation models, particularly for TOS guided by multi-modal large language model (MLLM). We argue that ViT does not need to process every image token through all of its layers -- only the tokens related to reasoning tasks are necessary. We design a new pruning decoder to take both image tokens and vision-language guidance as input to predict the relevance of each image token to the task. Only image tokens with high relevance are passed to deeper layers of the ViT. Experiments show that the VLTP framework reduces the computational costs of ViT by approximately 25% without performance degradation and by around 40% with only a 1% performance drop. The code associated with this study can be found at this URL.

cs.CV↗

arXiv2024

Explicit Mutual Information Maximization for Self-Supervised Learning

Recently, self-supervised learning (SSL) has been extensively studied. Theoretically, mutual information maximization (MIM) is an optimal criterion for SSL, with a strong theoretical foundation in information theory. However, it is difficult to directly apply MIM in SSL since the data distribution is not analytically available in applications. In practice, many existing methods can be viewed as approximate implementations of the MIM criterion. This work shows that, based on the invariance property of MI, explicit MI maximization can be applied to SSL under a generic distribution assumption, i.e., a relaxed condition of the data distribution. We further illustrate this by analyzing the generalized Gaussian distribution. Based on this result, we derive a loss function based on the MIM criterion using only second-order statistics. We implement the new loss for SSL and demonstrate its effectiveness via extensive experiments.

cs.CV↗

arXiv2024

Brain-Inspired Online Adaptation for Remote Sensing with Spiking Neural Network

On-device computing, or edge computing, is becoming increasingly important for remote sensing, particularly in applications like deep network-based perception on on-orbit satellites and unmanned aerial vehicles (UAVs). In these scenarios, two brain-like capabilities are crucial for remote sensing models: (1) high energy efficiency, allowing the model to operate on edge devices with limited computing resources, and (2) online adaptation, enabling the model to quickly adapt to environmental variations, weather changes, and sensor drift. This work addresses these needs by proposing an online adaptation framework based on spiking neural networks (SNNs) for remote sensing. Starting with a pretrained SNN model, we design an efficient, unsupervised online adaptation algorithm, which adopts an approximation of the BPTT algorithm and only involves forward-in-time computation that significantly reduces the computational complexity of SNN adaptation learning. Besides, we propose an adaptive activation scaling scheme to boost online SNN adaptation performance, particularly in low time-steps. Furthermore, for the more challenging remote sensing detection task, we propose a confidence-based instance weighting scheme, which substantially improves adaptation performance in the detection task. To our knowledge, this work is the first to address the online adaptation of SNNs. Extensive experiments on seven benchmark datasets across classification, segmentation, and detection tasks demonstrate that our proposed method significantly outperforms existing domain adaptation and domain generalization approaches under varying weather conditions. The proposed method enables energy-efficient and fast online adaptation on edge devices, and has much potential in applications such as remote perception on on-orbit satellites and UAV.

cs.LG↗

arXiv2024

Recoverable Anonymization for Pose Estimation: A Privacy-Enhancing Approach

Human pose estimation (HPE) is crucial for various applications. However, deploying HPE algorithms in surveillance contexts raises significant privacy concerns due to the potential leakage of sensitive personal information (SPI) such as facial features, and ethnicity. Existing privacy-enhancing methods often compromise either privacy or performance, or they require costly additional modalities. We propose a novel privacy-enhancing system that generates privacy-enhanced portraits while maintaining high HPE performance. Our key innovations include the reversible recovery of SPI for authorized personnel and the preservation of contextual information. By jointly optimizing a privacy-enhancing module, a privacy recovery module, and a pose estimator, our system ensures robust privacy protection, efficient SPI recovery, and high-performance HPE. Experimental results demonstrate the system's robust performance in privacy enhancement, SPI recovery, and HPE.

cs.CV↗

arXiv2024

Analyzing and Bridging the Gap between Maximizing Total Reward and Discounted Reward in Deep Reinforcement Learning

The optimal objective is a fundamental aspect of reinforcement learning (RL), as it determines how policies are evaluated and optimized. While total return maximization is the ideal objective in RL, discounted return maximization is the practical objective due to its stability. This can lead to a misalignment of objectives. To better understand the problem, we theoretically analyze the performance gap between the policy maximizes the total return and the policy maximizes the discounted return. Our analysis reveals that increasing the discount factor can be ineffective at eliminating this gap when environment contains cyclic states,a frequent scenario. To address this issue, we propose two alternative approaches to align the objectives. The first approach achieves alignment by modifying the terminal state value, treating it as a tunable hyper-parameter with its suitable range defined through theoretical analysis. The second approach focuses on calibrating the reward data in trajectories, enabling alignment in practical Deep RL applications using off-policy algorithms. This method enhances robustness to the discount factor and improve performance when the trajectory length is large. Our proposed methods demonstrate that adjusting reward data can achieve alignment, providing an insight that can be leveraged to design new optimization objectives to fundamentally enhance the performance of RL algorithms.

cs.LG↗

arXiv2024

Probing Implicit Bias in Semi-gradient Q-learning: Visualizing the Effective Loss Landscapes via the Fokker--Planck Equation

Semi-gradient Q-learning is applied in many fields, but due to the absence of an explicit loss function, studying its dynamics and implicit bias in the parameter space is challenging. This paper introduces the Fokker--Planck equation and employs partial data obtained through sampling to construct and visualize the effective loss landscape within a two-dimensional parameter space. This visualization reveals how the global minima in the loss landscape can transform into saddle points in the effective loss landscape, as well as the implicit bias of the semi-gradient method. Additionally, we demonstrate that saddle points, originating from the global minima in loss landscape, still exist in the effective loss landscape under high-dimensional parameter spaces and neural network settings. This paper develop a novel approach for probing implicit bias in semi-gradient Q-learning.

cs.LG↗

arXiv2024

TaskCLIP: Extend Large Vision-Language Model for Task Oriented Object Detection

Task-oriented object detection aims to find objects suitable for accomplishing specific tasks. As a challenging task, it requires simultaneous visual data processing and reasoning under ambiguous semantics. Recent solutions are mainly all-in-one models. However, the object detection backbones are pre-trained without text supervision. Thus, to incorporate task requirements, their intricate models undergo extensive learning on a highly imbalanced and scarce dataset, resulting in capped performance, laborious training, and poor generalizability. In contrast, we propose TaskCLIP, a more natural two-stage design composed of general object detection and task-guided object selection. Particularly for the latter, we resort to the recently successful large Vision-Language Models (VLMs) as our backbone, which provides rich semantic knowledge and a uniform embedding space for images and texts. Nevertheless, the naive application of VLMs leads to sub-optimal quality, due to the misalignment between embeddings of object images and their visual attributes, which are mainly adjective phrases. To this end, we design a transformer-based aligner after the pre-trained VLMs to re-calibrate both embeddings. Finally, we employ a trainable score function to post-process the VLM matching results for object selection. Experimental results demonstrate that our TaskCLIP outperforms the state-of-the-art DETR-based model TOIST by 3.5% and only requires a single NVIDIA RTX 4090 for both training and inference.

cs.CV↗

arXiv2024

HDReason: Algorithm-Hardware Codesign for Hyperdimensional Knowledge Graph Reasoning

In recent times, a plethora of hardware accelerators have been put forth for graph learning applications such as vertex classification and graph classification. However, previous works have paid little attention to Knowledge Graph Completion (KGC), a task that is well-known for its significantly higher algorithm complexity. The state-of-the-art KGC solutions based on graph convolution neural network (GCN) involve extensive vertex/relation embedding updates and complicated score functions, which are inherently cumbersome for acceleration. As a result, existing accelerator designs are no longer optimal, and a novel algorithm-hardware co-design for KG reasoning is needed. Recently, brain-inspired HyperDimensional Computing (HDC) has been introduced as a promising solution for lightweight machine learning, particularly for graph learning applications. In this paper, we leverage HDC for an intrinsically more efficient and acceleration-friendly KGC algorithm. We also co-design an acceleration framework named HDReason targeting FPGA platforms. On the algorithm level, HDReason achieves a balance between high reasoning accuracy, strong model interpretability, and less computation complexity. In terms of architecture, HDReason offers reconfigurability, high training throughput, and low energy consumption. When compared with NVIDIA RTX 4090 GPU, the proposed accelerator achieves an average 10.6x speedup and 65x energy efficiency improvement. When conducting cross-models and cross-platforms comparison, HDReason yields an average 4.2x higher performance and 3.4x better energy efficiency with similar accuracy versus the state-of-the-art FPGA-based GCN training platform.

cs.AR↗

arXiv2024

A priori Estimates for Deep Residual Network in Continuous-time Reinforcement Learning

Deep reinforcement learning excels in numerous large-scale practical applications. However, existing performance analyses ignores the unique characteristics of continuous-time control problems, is unable to directly estimate the generalization error of the Bellman optimal loss and require a boundedness assumption. Our work focuses on continuous-time control problems and proposes a method that is applicable to all such problems where the transition function satisfies semi-group and Lipschitz properties. Under this method, we can directly analyze the \emph{a priori} generalization error of the Bellman optimal loss. The core of this method lies in two transformations of the loss function. To complete the transformation, we propose a decomposition method for the maximum operator. Additionally, this analysis method does not require a boundedness assumption. Finally, we obtain an \emph{a priori} generalization error without the curse of dimensionality.

cs.LG↗