Search

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., Comment: Extended Abstract Track at the Unifying Representations in Neural Models Workshop (NeurIPS 2024)

In vision-language models (VLMs), visual tokens usually consume a significant amount of computational overhead, despite their sparser information density compared to text tokens. To address this, most existing methods learn a network to prune redundant visual tokens and require additional training data. Differently, we propose an efficient training-free token optimization mechanism dubbed SparseVLM without extra parameters or fine-tuning costs. Concretely, given that visual tokens complement text tokens in VLMs for linguistic reasoning, we select visual-relevant text tokens to rate the significance of vision tokens within the self-attention matrix extracted from the VLMs. Then we progressively prune irrelevant tokens. To maximize sparsity while retaining essential information, we introduce a rank-based strategy to adaptively determine the sparsification ratio for each layer, alongside a token recycling method that compresses pruned tokens into more compact representations. Experimental results show that our SparseVLM improves the efficiency of various VLMs across a range of image and video understanding tasks. In particular, LLaVA equipped with SparseVLM reduces 61% to 67% FLOPs with a compression ratio of 78% while maintaining 93% of the accuracy. Our code is available at https://github.com/Gumpest/SparseVLMs., Comment: 17 pages

Transition metal dichalcogenides (TMD) like MoSe$_2$ exhibit remarkable optical properties such as intense photoluminescence (PL) in the monolayer form. To date, narrow-linewidth PL is only achieved in micrometer-sized exfoliated TMD flakes encapsulated in hexagonal boron nitride (hBN). In this work, we develop a growth strategy to prepare monolayer MoSe$_2$ on hBN flakes by molecular beam epitaxy in the van der Waals regime. It constitutes the first step towards the development of large area single crystalline TMDs encapsulated in hBN for potential integration in electronic or opto-electronic devices. For this purpose, we define a two-step growth strategy to achieve monolayer-thick MoSe$_2$ grains on hBN flakes. The high quality of MoSe$_2$ allows us to detect very narrow PL linewidth down to 5.5 meV at 13 K, comparable to the one of encapsulated exfoliated MoSe$_2$ flakes. Moreover, sizeable PL can be detected at room temperature as well as clear reflectivity signatures of A, B and charged excitons., Comment: 6 pages, 6 figures

This paper presents an integrated perspective on robustness in regression. Specifically, we examine the relationship between traditional outlier-resistant robust estimation and robust optimization, which focuses on parameter estimation resistant to imaginary dataset-perturbations. While both are commonly regarded as robust methods, these concepts demonstrate a bias-variance trade-off, indicating that they follow roughly converse strategies., Comment: 17pages, 16figures

The impact of quantization on the overall performance of deep learning models is a well-studied problem. However, understanding and mitigating its effects on a more fine-grained level is still lacking, especially for harder tasks such as object detection with both classification and regression objectives. This work defines the performance for a subset of task-critical categories, i.e. the critical-category performance, as a crucial yet largely overlooked fine-grained objective for detection tasks. We analyze the impact of quantization at the category-level granularity, and propose methods to improve performance for the critical categories. Specifically, we find that certain critical categories have a higher sensitivity to quantization, and are prone to overfitting after quantization-aware training (QAT). To explain this, we provide theoretical and empirical links between their performance gaps and the corresponding loss landscapes with the Fisher information framework. Using this evidence, we apply a Fisher-aware mixed-precision quantization scheme, and a Fisher-trace regularization for the QAT on the critical-category loss landscape. The proposed methods improve critical-category metrics of the quantized transformer-based DETR detectors. They are even more significant in case of larger models and higher number of classes where the overfitting becomes more severe. For example, our methods lead to 10.4% and 14.5% mAP gains for, correspondingly, 4-bit DETR-R50 and Deformable DETR on the most impacted critical classes in the COCO Panoptic dataset., Comment: Poster presentation at the 2nd Workshop on Advancing Neural Network Training: Computational Efficiency, Scalability, and Resource Optimization (WANT@ICML 2024)

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., Comment: Accepted to UAI 2024. Preprint

Machine learning is increasingly used to predict lifestyle-related disease onset using health and medical data. However, the prediction effectiveness is hindered by dataset shift, which involves discrepancies in data distribution between the training and testing datasets, misclassifying out-of-distribution (OOD) data. To diminish dataset shift effects, this paper proposes the out-of-distribution reject option for prediction (ODROP), which integrates OOD detection models to preclude OOD data from the prediction phase. We investigated the efficacy of five OOD detection methods (variational autoencoder, neural network ensemble std, neural network ensemble epistemic, neural network energy, and neural network gaussian mixture based energy measurement) across two datasets, the Hirosaki and Wakayama health checkup data, in the context of three disease onset prediction tasks: diabetes, dyslipidemia, and hypertension. To evaluate the ODROP method, we trained disease onset prediction models and OOD detection models on Hirosaki data and used AUROC-rejection curve plots from Wakayama data. The variational autoencoder method showed superior stability and magnitude of improvement in Area Under the Receiver Operating Curve (AUROC) in five cases: AUROC in the Wakayama data was improved from 0.80 to 0.90 at a 31.1% rejection rate for diabetes onset and from 0.70 to 0.76 at a 34% rejection rate for dyslipidemia. We categorized dataset shifts into two types using SHAP clustering - those that considerably affect predictions and those that do not. We expect that this classification will help standardize measuring instruments. This study is the first to apply OOD detection to actual health and medical data, demonstrating its potential to substantially improve the accuracy and reliability of disease prediction models amidst dataset shift.

Two-dimensional (2D) materials like transition metal dichalcogenides (TMD) have proved to be serious candidates to replace silicon in several technologies with enhanced performances. In this respect, the two remaining challenges are the wafer scale growth of TMDs and their integration into operational devices using clean room compatible processes. In this work, two different CMOS-compatible protocols are developed for the fabrication of MoS$_2$-based memristors, and the resulting performances are compared. The quality of MoS$_2$ at each stage of the process is characterized by Raman spectroscopy and x-ray photoemission spectroscopy. In the first protocol, the structure of MoS$_2$ is preserved during transfer and patterning processes. However, a polymer layer with a minimum thickness of 3 nm remains at the surface of MoS$_2$ limiting the electrical switching performances. In the second protocol, the contamination layer is completely removed resulting in improved electrical switching performances and reproducibility. Based on physico-chemical and electrical results, the switching mechanism is discussed in terms of conduction through grain boundaries., Comment: 11 pages, 6 figures

Among various quantum machine learning (QML) algorithms, the quantum kernel method has especially attracted attention due to its compatibility with noisy intermediate-scale quantum devices and its potential to achieve quantum advantage. This method performs classification and regression by nonlinearly mapping data into quantum states in a higher dimensional Hilbert space. Thus far, the quantum kernel method has been implemented only on qubit-based systems, but continuous-variable (CV) systems can potentially offer superior computational power by utilizing its infinite-dimensional Hilbert space. Here, we demonstrate the implementation of the classification task with the CV quantum kernel method on a programmable photonic quantum processor. We experimentally prove that the CV quantum kernel method successfully classifies several datasets robustly even under the experimental imperfections, with high accuracies comparable to the classical kernel. This demonstration sheds light on the utility of CV quantum systems for QML and should stimulate further study in other CV QML algorithms., Comment: 8 pages and 4 figures

In optical quantum information processing with continuous variables, optical non-Gaussian quantum states are essential for universal and fault-tolerant quantum computation. Experimentally, their most typical generation method is photon subtraction (PS) where single-photon detection by an on/off detector probabilistically heralds the generation of squeezed single-photon states. In PS, however, trying to avoid unwanted multi-photon detection inevitably limits the generation rate, hindering the application of squeezed single-photon states. Here, we theoretically show that generalized photon subtraction (GPS), a simple extension of PS, can improve the generation rate while maintaining the quality of the generated states. Furthermore, we experimentally demonstrate the generation rate improvement for 2 dB- and 4 dB-squeezed single-photon states compared to PS, by more than one order of magnitude particularly for the case of 2 dB. Our results will accelerate the application of squeezed single-photon states to more advanced quantum information protocols., Comment: 13 pages, 7 figures

Quantum computing has been pursued with various hardware platforms, and an optical system is one of the most reasonable choices for large-scale computation. In the optical continuous-variable computation scheme, the incorporation of Gaussian gates and a highly non-classical non-Gaussian state enables universal quantum computation. Although basic technologies for Gaussian gates and non-Gaussian state generations have long been developed, these building blocks have not yet been integrated in a scalable fashion. Here, we integrate them to develop a scalable and programmable optical quantum computing platform that can sequentially perform an essential Gaussian gate, the squeezing gate, on a non-Gaussian input state. The key enablers are a loop-based optical circuit with dynamical and programmable controllability and its time-synchronization with the probabilistic non-Gaussian state generation. We verify the deterministic, programmable, and repeatable quantum gates on a typical non-Gaussian state by implementing up to three-step gates. The gates implemented are so high-quality that strong evidence of the states' non-classicalities, negativities of the Wigner functions, are preserved even after multistep gates. This platform is compatible with other non-Gaussian states and realizes large-scale universal quantum computing by incorporating other existing processing technologies., Comment: 9 pages, 3 figures for the main text, and 3 pages and 2 figures for the supplementary information

We present a loop-based optical processor enabling time-domain programmable beam-splitter (BS) operations for a phase-sensitive non-Gaussian state. The loop itself is of high quality, allowing for storage of a non-Gaussian state for up to seven round trips while preserving its Wigner negativity and phase coherence. We perform various BS operations on a non-Gaussian state and evaluate them as transformations of the state's waveforms. Our work integrates non-Gaussian states with time multiplexing, laying the foundation for large-scale universal quantum information processing., Comment: 5 pages, 4 figures

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., Comment: 13 pages

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., Comment: aaai2024

Uncertainty estimation is crucial for machine learning models to detect out-of-distribution (OOD) inputs. However, the conventional discriminative deep learning classifiers produce uncalibrated closed-set predictions for OOD data. A more robust classifiers with the uncertainty estimation typically require a potentially unavailable OOD dataset for outlier exposure training, or a considerable amount of additional memory and compute to build ensemble models. In this work, we improve on uncertainty estimation without extra OOD data or additional inference costs using an alternative Split-Ensemble method. Specifically, we propose a novel subtask-splitting ensemble training objective, where a common multiclass classification task is split into several complementary subtasks. Then, each subtask's training data can be considered as OOD to the other subtasks. Diverse submodels can therefore be trained on each subtask with OOD-aware objectives. The subtask-splitting objective enables us to share low-level features across submodels to avoid parameter and computational overheads. In particular, we build a tree-like Split-Ensemble architecture by performing iterative splitting and pruning from a shared backbone model, where each branch serves as a submodel corresponding to a subtask. This leads to improved accuracy and uncertainty estimation across submodels under a fixed ensemble computation budget. Empirical study with ResNet-18 backbone shows Split-Ensemble, without additional computation cost, improves accuracy over a single model by 0.8%, 1.8%, and 25.5% on CIFAR-10, CIFAR-100, and Tiny-ImageNet, respectively. OOD detection for the same backbone and in-distribution datasets surpasses a single model baseline by, correspondingly, 2.2%, 8.1%, and 29.6% mean AUROC., Comment: ICML2024. Project website is available at https://antonioo-c.github.io/projects/split-ensemble

\'Cwik and Mielniczuk (1989) introduced a univariate kernel density ratio estimator, which directly estimates the ratio without estimating the two densities of interest. This study presents its straightforward multivariate adaptation., Comment: 8 pages, 9 figures

Detailed measurements of the spectral structure of cosmic-ray electrons and positrons from 10.6 GeV to 7.5 TeV are presented from over 7 years of observations with the CALorimetric Electron Telescope (CALET) on the International Space Station. Because of the excellent energy resolution (a few percent above 10 GeV) and the outstanding e/p separation (10$^5$), CALET provides optimal performance for a detailed search of structures in the energy spectrum. The analysis uses data up to the end of 2022, and the statistics of observed electron candidates has increased more than 3 times since the last publication in 2018. By adopting an updated boosted decision tree analysis, a sufficient proton rejection power up to 7.5 TeV is achieved, with a residual proton contamination less than 10%. The observed energy spectrum becomes gradually harder in the lower energy region from around 30 GeV, consistently with AMS-02, but from 300 to 600 GeV it is considerably softer than the spectra measured by DAMPE and Fermi-LAT. At high energies, the spectrum presents a sharp break around 1 TeV, with a spectral index change from -3.15 to -3.91, and a broken power law fitting the data in the energy range from 30 GeV to 4.8 TeV better than a single power law with 6.9 sigma significance, which is compatible with the DAMPE results. The break is consistent with the expected effects of radiation loss during the propagation from distant sources (except the highest energy bin). We have fitted the spectrum with a model consistent with the positron flux measured by AMS-02 below 1 TeV and interpreted the electron + positron spectrum with possible contributions from pulsars and nearby sources. Above 4.8 TeV, a possible contribution from known nearby supernova remnants, including Vela, is addressed by an event-by-event analysis providing a higher proton-rejection power than a purely statistical analysis., Comment: main text: 7 pages, 4 figures; supplemental material: 10 pages, 8 figures, 1 table

There has been extensive activity exploring the doping of semiconducting two-dimensional (2D) transition metal dichalcogenides in order to tune their electronic and magnetic properties. The outcome of doping depends on various factors, including the intrinsic properties of the host material, the nature of the dopants used, their spatial distribution as well as their interactions with other types of defects. A thorough atomic-level analysis is essential to fully understand these mechanisms. In this work, vanadium doped WSe$_2$ monolayer grown by molecular beam epitaxy is investigated using four-dimensional scanning transmission electron microscopy (4D-STEM). Through center of mass-based reconstruction, atomic scale maps are produced, allowing the visualization of both the electric field and the electrostatic potential around individual V atoms. To provide quantitative insights, these results are successfully compared with multislice image simulations based on ab initio calculations, accounting for lens aberrations. Finally, a negative charge around the V dopants is detected as a drop in the electrostatic potential, unambiguously demonstrating that 4D-STEM can be used to detect and to accurately analyze single dopant charge states in semiconducting 2D materials., Comment: 17 pages, 4 figures and Supporting Information