Denis Gudovskiy

Jinze Yu, Jiaming Liu, Xiaobao Wei et al.

Recently, DEtection TRansformer (DETR), an end-to-end object detection pipeline, has achieved promising performance. However, it requires large-scale labeled data and suffers from domain shift, especially when no labeled data is available in the target domain. To solve this problem, we propose an end-to-end cross-domain detection Transformer based on the mean teacher framework, MTTrans, which can fully exploit unlabeled target domain data in object detection training and transfer knowledge between domains via pseudo labels. We further propose the comprehensive multi-level feature alignment to improve the pseudo labels generated by the mean teacher framework taking advantage of the cross-scale self-attention mechanism in Deformable DETR. Image and object features are aligned at the local, global, and instance levels with domain query-based feature alignment (DQFA), bi-level graph-based prototype alignment (BGPA), and token-wise image feature alignment (TIFA). On the other hand, the unlabeled target domain data pseudo-labeled and available for the object detection training by the mean teacher framework can lead to better feature extraction and alignment. Thus, the mean teacher framework and the comprehensive multi-level feature alignment can be optimized iteratively and mutually based on the architecture of Transformers. Extensive experiments demonstrate that our proposed method achieves state-of-the-art performance in three domain adaptation scenarios, especially the result of Sim10k to Cityscapes scenario is remarkably improved from 52.6 mAP to 57.9 mAP. Code will be released.

Dongwei Wang, Jinhee Kim, Seokho Han et al.

Dynamic runtime latency and memory constraints necessitate flexible large language model (LLM) deployment, where an LLM can be inferred with various quantization precisions based on available computational resources. Recent work on such any-precision quantization either relies on hardware-inefficient vector quantization or induces additional scaling factors when switching between bit-widths. Meanwhile, existing post-training quantization (PTQ) methods calibrated for a fixed low precision show poor generalizability under runtime precision change. In this work, we attribute the source of poor generalization across bit-widths to a precision-dependent \textit{outlier migration} phenomenon where the distribution of PTQ-sensitive tokens changes across precisions. Motivated by this observation, we propose \texttt{MoBiQuant}, a novel any-precision Mixture-of-Bits quantization framework that adjusts weight precision for flexible LLM inference based on token sensitivity. Specifically, we propose a many-in-one recursive residual quantization that can iteratively reconstruct higher-precision weights at runtime and mitigates \textit{outlier migration} with a token-aware router to dynamically select the optimal inference precision of each token.Extensive experiments show that \texttt{MoBiQuant} matches or surpasses frontier single-precision PTQ while exhibiting strong elasticity, achieving significant memory savings and throughput gains of up to $1.34\times$ over state-of-the-art any-precision methods.

Denis Gudovskiy, Wenzhao Zheng, Tomoyuki Okuno et al.

Continuous normalizing flows (CNFs) and diffusion models (DMs) generate high-quality data from a noise distribution. However, their sampling process demands multiple iterations to solve an ordinary differential equation (ODE) with high computational complexity. State-of-the-art methods focus on reducing the number of discrete time steps during sampling to improve efficiency. In this work, we explore a complementary direction in which the quality-complexity tradeoff can also be controlled in terms of the neural network length. We achieve this by rewiring the blocks in the transformer-based architecture to solve an inner discretized ODE w.r.t. its depth. Then, we apply a length consistency term during flow matching training, and as a result, the sampling can be performed with an arbitrary number of time steps and transformer blocks. Unlike others, our ODE$_t$(ODE$_l$) approach is solver-agnostic in time dimension and reduces both latency and, importantly, memory usage. CelebA-HQ and ImageNet generation experiments show a latency reduction of up to $2\times$ in the most efficient sampling mode, and FID improvement of up to $2.8$ points for high-quality sampling when applied to prior methods. We open-source our code and checkpoints at github.com/gudovskiy/odelt.

Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata

Normalizing flow-based generative models have been widely used in applications where the exact density estimation is of major importance. Recent research proposes numerous methods to improve their expressivity. However, conditioning on a context is largely overlooked area in the bijective flow research. Conventional conditioning with the vector concatenation is limited to only a few flow types. More importantly, this approach cannot support a practical setup where a set of context-conditioned (specialist) models are trained with the fixed pretrained general-knowledge (generalist) model. We propose ContextFlow++ approach to overcome these limitations using an additive conditioning with explicit generalist-specialist knowledge decoupling. Furthermore, we support discrete contexts by the proposed mixed-variable architecture with context encoders. Particularly, our context encoder for discrete variables is a surjective flow from which the context-conditioned continuous variables are sampled. Our experiments on rotated MNIST-R, corrupted CIFAR-10C, real-world ATM predictive maintenance and SMAP unsupervised anomaly detection benchmarks show that the proposed ContextFlow++ offers faster stable training and achieves higher performance metrics. Our code is publicly available at https://github.com/gudovskiy/contextflow.

Denis Gudovskiy, Shun Ishizaka, Kazuki Kozuka

Unsupervised anomaly detection with localization has many practical applications when labeling is infeasible and, moreover, when anomaly examples are completely missing in the train data. While recently proposed models for such data setup achieve high accuracy metrics, their complexity is a limiting factor for real-time processing. In this paper, we propose a real-time model and analytically derive its relationship to prior methods. Our CFLOW-AD model is based on a conditional normalizing flow framework adopted for anomaly detection with localization. In particular, CFLOW-AD consists of a discriminatively pretrained encoder followed by a multi-scale generative decoders where the latter explicitly estimate likelihood of the encoded features. Our approach results in a computationally and memory-efficient model: CFLOW-AD is faster and smaller by a factor of 10x than prior state-of-the-art with the same input setting. Our experiments on the MVTec dataset show that CFLOW-AD outperforms previous methods by 0.36% AUROC in detection task, by 1.12% AUROC and 2.5% AUPRO in localization task, respectively. We open-source our code with fully reproducible experiments.

Denis Gudovskiy, Gyuri Han, Takuya Yamaguchi et al.

Current home appliances are capable to execute a limited number of voice commands such as turning devices on or off, adjusting music volume or light conditions. Recent progress in machine reasoning gives an opportunity to develop new types of conversational user interfaces for home appliances. In this paper, we apply state-of-the-art visual reasoning model and demonstrate that it is feasible to ask a smart fridge about its contents and various properties of the food with close-to-natural conversation experience. Our visual reasoning model answers user questions about existence, count, category and freshness of each product by analyzing photos made by the image sensor inside the smart fridge. Users may chat with their fridge using off-the-shelf phone messenger while being away from home, for example, when shopping in the supermarket. We generate a visually realistic synthetic dataset to train machine learning reasoning model that achieves 95% answer accuracy on test data. We present the results of initial user tests and discuss how we modify distribution of generated questions for model training based on human-in-the-loop guidance. We open source code for the whole system including dataset generation, reasoning model and demonstration scripts.

Jerry Jiang, Haowen Sun, Denis Gudovskiy et al.

Spatial intelligence in vision-language models (VLMs) attracts research interest with the practical demand to reason in the 3D world.Despite promising results, most existing methods follow the conventional 2D pipeline in VLMs and use pixel-aligned representations for the vision modality. However, correspondence-based models with implicit 3D scene understanding often fail to achieve spatial consistency, and representation-based models with 3D geometric priors lack efficiency in vision sequence serialization. To address this, we propose a Proxy3D method with compact yet comprehensive 3D proxy representations for the vision modality. Given only video frames as input, we employ semantic and geometric encoders to extract scene features and then perform their semantic-aware clustering to obtain a set of proxies in the 3D space. For representation alignment, we further curate the SpaceSpan dataset and apply multi-stage training to adopt the proposed 3D proxy representations with the VLM. When using shorter sequences for vision information, our method achieves competitive or state-of-the-art performance in 3D visual question answering, visual grounding and general spatial intelligence benchmarks.

Rongyu Zhang, Yulin Luo, Jiaming Liu et al. · berkeley

The Mixture-of-Experts (MoE) approach has demonstrated outstanding scalability in multi-task learning including low-level upstream tasks such as concurrent removal of multiple adverse weather effects. However, the conventional MoE architecture with parallel Feed Forward Network (FFN) experts leads to significant parameter and computational overheads that hinder its efficient deployment. In addition, the naive MoE linear router is suboptimal in assigning task-specific features to multiple experts which limits its further scalability. In this work, we propose an efficient MoE architecture with weight sharing across the experts. Inspired by the idea of linear feature modulation (FM), our architecture implicitly instantiates multiple experts via learnable activation modulations on a single shared expert block. The proposed Feature Modulated Expert (FME) serves as a building block for the novel Mixture-of-Feature-Modulation-Experts (MoFME) architecture, which can scale up the number of experts with low overhead. We further propose an Uncertainty-aware Router (UaR) to assign task-specific features to different FM modules with well-calibrated weights. This enables MoFME to effectively learn diverse expert functions for multiple tasks. The conducted experiments on the multi-deweather task show that our MoFME outperforms the baselines in the image restoration quality by 0.1-0.2 dB and achieves SOTA-compatible performance while saving more than 72% of parameters and 39% inference time over the conventional MoE counterpart. Experiments on the downstream segmentation and classification tasks further demonstrate the generalizability of MoFME to real open-world applications.

Rongyu Zhang, Zefan Cai, Huanrui Yang et al.

Finetuning a pretrained vision model (PVM) is a common technique for learning downstream vision tasks. However, the conventional finetuning process with randomly sampled data points results in diminished training efficiency. To address this drawback, we propose a novel approach, Vision-language Collaborative Active Finetuning (VeCAF). With the emerging availability of labels and natural language annotations of images through web-scale crawling or controlled generation, VeCAF makes use of these information to perform parametric data selection for PVM finetuning. VeCAF incorporates the finetuning objective to select significant data points that effectively guide the PVM towards faster convergence to meet the performance goal. This process is assisted by the inherent semantic richness of the text embedding space which we use to augment image features. Furthermore, the flexibility of text-domain augmentation allows VeCAF to handle out-of-distribution scenarios without external data. Extensive experiments show the leading performance and high computational efficiency of VeCAF that is superior to baselines in both in-distribution and out-of-distribution image classification tasks. On ImageNet, VeCAF uses up to 3.3x less training batches to reach the target performance compared to full finetuning, and achieves an accuracy improvement of 2.7% over the state-of-the-art active finetuning method with the same number of batches.

Stefano Rando, Luca Romani, Matteo Migliarini et al.

State Space Models (SSMs) have recently enjoyed a rise to prominence in the field of deep learning for sequence modeling, especially as an alternative to Transformers. Their success stems from avoiding two well-known drawbacks of attention-based models: quadratic complexity with respect to the sequence length and inability to model long-range dependencies. The SSM variant Mamba has demonstrated performance comparable to Transformers without any form of attention, thanks to the use of a selective mechanism for the state parameters. Selectivity, however, is only evaluated empirically and the reasons of its effectiveness remain unclear. In this work, we show how selectivity is related to the sequence processing. Our analysis shows that selective time intervals in Mamba act as linear approximators of information. Then, we propose our SeRpEnt architecture, a SSM that further exploits selectivity to compress sequences in an information-aware fashion. It employs a resampling mechanism that aggregates elements based on their information content. Our empirical results in the Long Range Arena benchmark and other language modeling tasks show benefits of the SeRpEnt's resampling mechanism.

Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata

Continuous normalizing flows (CNFs) can model data distributions with expressive infinite-length architectures. But this modeling involves computationally expensive process of solving an ordinary differential equation (ODE) during maximum likelihood training. Recently proposed flow matching (FM) framework allows to substantially simplify the training phase using a regression objective with the interpolated forward vector field. In this paper, we propose an interpolant-free dual flow matching (DFM) approach without explicit assumptions about the modeled vector field. DFM optimizes the forward and, additionally, a reverse vector field model using a novel objective that facilitates bijectivity of the forward and reverse transformations. Our experiments with the SMAP unsupervised anomaly detection show advantages of DFM when compared to the CNF trained with either maximum likelihood or FM objectives with the state-of-the-art performance metrics.

Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata

Recent semantic segmentation models accurately classify test-time examples that are similar to a training dataset distribution. However, their discriminative closed-set approach is not robust in practical data setups with distributional shifts and out-of-distribution (OOD) classes. As a result, the predicted probabilities can be very imprecise when used as confidence scores at test time. To address this, we propose a generative model for concurrent in-distribution misclassification (IDM) and OOD detection that relies on a normalizing flow framework. The proposed flow-based detector with an energy-based inputs (FlowEneDet) can extend previously deployed segmentation models without their time-consuming retraining. Our FlowEneDet results in a low-complexity architecture with marginal increase in the memory footprint. FlowEneDet achieves promising results on Cityscapes, Cityscapes-C, FishyScapes and SegmentMeIfYouCan benchmarks in IDM/OOD detection when applied to pretrained DeepLabV3+ and SegFormer semantic segmentation models.

Konstantinos Kallidromitis, Denis Gudovskiy, Kazuki Kozuka et al.

Recent contrastive methods show significant improvement in self-supervised learning in several domains. In particular, contrastive methods are most effective where data augmentation can be easily constructed e.g. in computer vision. However, they are less successful in domains without established data transformations such as time series data. In this paper, we propose a novel self-supervised learning framework that combines contrastive learning with neural processes. It relies on recent advances in neural processes to perform time series forecasting. This allows to generate augmented versions of data by employing a set of various sampling functions and, hence, avoid manually designed augmentations. We extend conventional neural processes and propose a new contrastive loss to learn times series representations in a self-supervised setup. Therefore, unlike previous self-supervised methods, our augmentation pipeline is task-agnostic, enabling our method to perform well across various applications. In particular, a ResNet with a linear classifier trained using our approach is able to outperform state-of-the-art techniques across industrial, medical and audio datasets improving accuracy over 10% in ECG periodic data. We further demonstrate that our self-supervised representations are more efficient in the latent space, improving multiple clustering indexes and that fine-tuning our method on 10% of labels achieves results competitive to fully-supervised learning.

Denis Gudovskiy, Luca Rigazio, Shun Ishizaka et al.

AutoAugment has sparked an interest in automated augmentation methods for deep learning models. These methods estimate image transformation policies for train data that improve generalization to test data. While recent papers evolved in the direction of decreasing policy search complexity, we show that those methods are not robust when applied to biased and noisy data. To overcome these limitations, we reformulate AutoAugment as a generalized automated dataset optimization (AutoDO) task that minimizes the distribution shift between test data and distorted train dataset. In our AutoDO model, we explicitly estimate a set of per-point hyperparameters to flexibly change distribution of train data. In particular, we include hyperparameters for augmentation, loss weights, and soft-labels that are jointly estimated using implicit differentiation. We develop a theoretical probabilistic interpretation of this framework using Fisher information and show that its complexity scales linearly with the dataset size. Our experiments on SVHN, CIFAR-10/100, and ImageNet classification show up to 9.3% improvement for biased datasets with label noise compared to prior methods and, importantly, up to 36.6% gain for underrepresented SVHN classes.

Denis Gudovskiy, Alec Hodgkinson, Takuya Yamaguchi et al.

Active learning (AL) aims to minimize labeling efforts for data-demanding deep neural networks (DNNs) by selecting the most representative data points for annotation. However, currently used methods are ill-equipped to deal with biased data. The main motivation of this paper is to consider a realistic setting for pool-based semi-supervised AL, where the unlabeled collection of train data is biased. We theoretically derive an optimal acquisition function for AL in this setting. It can be formulated as distribution shift minimization between unlabeled train data and weakly-labeled validation dataset. To implement such acquisition function, we propose a low-complexity method for feature density matching using self-supervised Fisher kernel (FK) as well as several novel pseudo-label estimators. Our FK-based method outperforms state-of-the-art methods on MNIST, SVHN, and ImageNet classification while requiring only 1/10th of processing. The conducted experiments show at least 40% drop in labeling efforts for the biased class-imbalanced data compared to existing methods.

Denis Gudovskiy, Alec Hodgkinson, Takuya Yamaguchi et al.

Explaining predictions of deep neural networks (DNNs) is an important and nontrivial task. In this paper, we propose a practical approach to interpret decisions made by a DNN object detector that has fidelity comparable to state-of-the-art methods and sufficient computational efficiency to process large datasets. Our method relies on recent theory and approximates Shapley feature importance values. We qualitatively and quantitatively show that the proposed explanation method can be used to find image features which cause failures in DNN object detection. The developed software tool combined into the "Explain to Fix" (E2X) framework has a factor of 10 higher computational efficiency than prior methods and can be used for cluster processing using graphics processing units (GPUs). Lastly, we propose a potential extension of the E2X framework where the discovered missing features can be added into training dataset to overcome failures after model retraining.