Low-data Regime Research Articles

Score-Based Learning of Graphical Event Models with Background Knowledge Augmentation

Graphical event models (GEMs) are representations of temporal point process dynamics between different event types. Many real-world applications however involve limited event stream data, making it challenging to learn GEMs from data alone. In this paper, we introduce approaches that can work together in a score-based learning paradigm, to augment data with potentially different types of background knowledge. We propose novel scores for learning an important parametric class of GEMs; in particular, we propose a Bayesian score for leveraging prior information as well as a more practical simplification that involves fewer parameters, analogous to Bayesian networks. We also introduce a framework for incorporating easily assessed qualitative background knowledge from domain experts, in the form of statements such as `event X depends on event Y' or `event Y makes event X more likely'. The proposed framework has Bayesian interpretations and can be deployed by any score-based learner. Through an extensive empirical investigation, we demonstrate the practical benefits of background knowledge augmentation while learning GEMs for applications in the low-data regime.

Physics-informed neural ODE (PINODE): embedding physics into models using collocation points

Building reduced-order models (ROMs) is essential for efficient forecasting and control of complex dynamical systems. Recently, autoencoder-based methods for building such models have gained significant traction, but their demand for data limits their use when the data is scarce and expensive. We propose aiding a model’s training with the knowledge of physics using a collocation-based physics-informed loss term. Our innovation builds on ideas from classical collocation methods of numerical analysis to embed knowledge from a known equation into the latent-space dynamics of a ROM. We show that the addition of our physics-informed loss allows for exceptional data supply strategies that improves the performance of ROMs in data-scarce settings, where training high-quality data-driven models is impossible. Namely, for a problem of modeling a high-dimensional nonlinear PDE, our experiments show times 5 performance gains, measured by prediction error, in a low-data regime, times 10 performance gains in tasks of high-noise learning, times 100 gains in the efficiency of utilizing the latent-space dimension, and times 200 gains in tasks of far-out out-of-distribution forecasting relative to purely data-driven models. These improvements pave the way for broader adoption of network-based physics-informed ROMs in compressive sensing and control applications.

A foundational vision transformer improves diagnostic performance for electrocardiograms

The electrocardiogram (ECG) is a ubiquitous diagnostic modality. Convolutional neural networks (CNNs) applied towards ECG analysis require large sample sizes, and transfer learning approaches for biomedical problems may result in suboptimal performance when pre-training is done on natural images. We leveraged masked image modeling to create a vision-based transformer model, HeartBEiT, for electrocardiogram waveform analysis. We pre-trained this model on 8.5 million ECGs and then compared performance vs. standard CNN architectures for diagnosis of hypertrophic cardiomyopathy, low left ventricular ejection fraction and ST elevation myocardial infarction using differing training sample sizes and independent validation datasets. We find that HeartBEiT has significantly higher performance at lower sample sizes compared to other models. We also find that HeartBEiT improves explainability of diagnosis by highlighting biologically relevant regions of the EKG vs. standard CNNs. Domain specific pre-trained transformer models may exceed the classification performance of models trained on natural images especially in very low data regimes. The combination of the architecture and such pre-training allows for more accurate, granular explainability of model predictions.

PCA-Enhanced Autoencoders for Nonlinear Dimensionality Reduction in Low Data Regimes

Many scientific domains, such as nanophotonic design, gene expression, and materials design, are limited by high costs of acquiring data. This data is often intrinsically low-dimensional, nonlinear, and benefits from dimensionality reduction. Autoencoders (AE) provide nonlinear dimensionality reduction but are typically ineffective for low data regimes. Principal Component Analysis (PCA) is data-efficient but limited to linear dimensionality reduction. We propose a technique that harnesses the benefits of both methods by using PCA to initialize an AE. The proposed approach outperforms both PCA and standard AEs in low-data regimes and is comparable to the best of either of the two in other scenarios.

De Novo Design of Nurr1 Agonists via Fragment-Augmented Generative Deep Learning in Low-Data Regime.

Generative neural networks trained on SMILES can design innovative bioactive molecules de novo. These so-called chemical language models (CLMs) have typically been trained on tens of template molecules for fine-tuning. However, it is challenging to apply CLM to orphan targets with few known ligands. We have fine-tuned a CLM with a single potent Nurr1 agonist as template in a fragment-augmented fashion and obtained novel Nurr1 agonists using sampling frequency for design prioritization. Nanomolar potency and binding affinity of the top-ranking design and its structural novelty compared to available Nurr1 ligands highlight its value as an early chemical tool and as a lead for Nurr1 agonist development, as well as the applicability of CLM in very low-data scenarios.

ChampKit: A framework for rapid evaluation of deep neural networks for patch-based histopathology classification

Background and ObjectiveHistopathology is the gold standard for diagnosis of many cancers. Recent advances in computer vision, specifically deep learning, have facilitated the analysis of histopathology images for many tasks, including the detection of immune cells and microsatellite instability. However, it remains difficult to identify optimal models and training configurations for different histopathology classification tasks due to the abundance of available architectures and the lack of systematic evaluations. Our objective in this work is to present a software tool that addresses this need and enables robust, systematic evaluation of neural network models for patch classification in histology in a light-weight, easy-to-use package for both algorithm developers and biomedical researchers. MethodsHere we present ChampKit (Comprehensive Histopathology Assessment of Model Predictions toolKit): an extensible, fully reproducible evaluation toolkit that is a one-stop-shop to train and evaluate deep neural networks for patch classification. ChampKit curates a broad range of public datasets. It enables training and evaluation of models supported by timm directly from the command line, without the need for users to write any code. External models are enabled through a straightforward API and minimal coding. As a result, Champkit facilitates the evaluation of existing and new models and deep learning architectures on pathology datasets, making it more accessible to the broader scientific community. To demonstrate the utility of ChampKit, we establish baseline performance for a subset of possible models that could be employed with ChampKit, focusing on several popular deep learning models, namely ResNet18, ResNet50, and R26-ViT, a hybrid vision transformer. In addition, we compare each model trained either from random weight initialization or with transfer learning from ImageNet pretrained models. For ResNet18, we also consider transfer learning from a self-supervised pretrained model. ResultsThe main result of this paper is the ChampKit software. Using ChampKit, we were able to systemically evaluate multiple neural networks across six datasets. We observed mixed results when evaluating the benefits of pretraining versus random intialization, with no clear benefit except in the low data regime, where transfer learning was found to be beneficial. Surprisingly, we found that transfer learning from self-supervised weights rarely improved performance, which is counter to other areas of computer vision. ConclusionsChoosing the right model for a given digital pathology dataset is nontrivial. ChampKit provides a valuable tool to fill this gap by enabling the evaluation of hundreds of existing (or user-defined) deep learning models across a variety of pathology tasks. Source code and data for the tool are freely accessible at https://github.com/SBU-BMI/champkit.

Motor decoding from the posterior parietal cortex using deep neural networks

Objective. Motor decoding is crucial to translate the neural activity for brain-computer interfaces (BCIs) and provides information on how motor states are encoded in the brain. Deep neural networks (DNNs) are emerging as promising neural decoders. Nevertheless, it is still unclear how different DNNs perform in different motor decoding problems and scenarios, and which network could be a good candidate for invasive BCIs. Approach. Fully-connected, convolutional, and recurrent neural networks (FCNNs, CNNs, RNNs) were designed and applied to decode motor states from neurons recorded from V6A area in the posterior parietal cortex (PPC) of macaques. Three motor tasks were considered, involving reaching and reach-to-grasping (the latter under two illumination conditions). DNNs decoded nine reaching endpoints in 3D space or five grip types using a sliding window approach within the trial course. To evaluate decoders simulating a broad variety of scenarios, the performance was also analyzed while artificially reducing the number of recorded neurons and trials, and while performing transfer learning from one task to another. Finally, the accuracy time course was used to analyze V6A motor encoding. Main results. DNNs outperformed a classic Naïve Bayes classifier, and CNNs additionally outperformed XGBoost and Support Vector Machine classifiers across the motor decoding problems. CNNs resulted the top-performing DNNs when using less neurons and trials, and task-to-task transfer learning improved performance especially in the low data regime. Lastly, V6A neurons encoded reaching and reach-to-grasping properties even from action planning, with the encoding of grip properties occurring later, closer to movement execution, and appearing weaker in darkness. Significance. Results suggest that CNNs are effective candidates to realize neural decoders for invasive BCIs in humans from PPC recordings also reducing BCI calibration times (transfer learning), and that a CNN-based data-driven analysis may provide insights about the encoding properties and the functional roles of brain regions.

Reconstructing Kernel-Based Machine Learning Force Fields with Superlinear Convergence.

Kernel machines have sustained continuous progress in the field of quantum chemistry. In particular, they have proven to be successful in the low-data regime of force field reconstruction. This is because many equivariances and invariances due to physical symmetries can be incorporated into the kernel function to compensate for much larger data sets. So far, the scalability of kernel machines has however been hindered by its quadratic memory and cubical runtime complexity in the number of training points. While it is known that iterative Krylov subspace solvers can overcome these burdens, their convergence crucially relies on effective preconditioners, which are elusive in practice. Effective preconditioners need to partially presolve the learning problem in a computationally cheap and numerically robust manner. Here, we consider the broad class of Nyström-type methods to construct preconditioners based on successively more sophisticated low-rank approximations of the original kernel matrix, each of which provides a different set of computational trade-offs. All considered methods aim to identify a representative subset of inducing (kernel) columns to approximate the dominant kernel spectrum.

A Data-Efficient Deep Learning Strategy for Tissue Characterization via Quantitative Ultrasound: Zone Training.

Deep learning (DL) powered biomedical ultrasound imaging is an emerging research field where researchers adapt the image analysis capabilities of DL algorithms to biomedical ultrasound imaging settings. A major roadblock to wider adoption of DL powered biomedical ultrasound imaging is that acquisition of large and diverse datasets is expensive in clinical settings, which is a requirement for successful DL implementation. Hence, there is a constant need for developing data-efficient DL techniques to turn DL powered biomedical ultrasound imaging into reality. In this work, we develop a data-efficient DL training strategy for classifying tissues based on the ultrasonic backscattered RF data, i.e., quantitative ultrasound (QUS), which we named zone training. In zone training, we propose to divide the complete field of view of an ultrasound image into multiple zones associated with different regions of a diffraction pattern and then, train separate DL networks for each zone. The main advantage of zone training is that it requires less training data to achieve high accuracy. In this work, three different tissue-mimicking phantoms were classified by a DL network. The results demonstrated that zone training can require a factor of 2-3 less training data in low data regime to achieve similar classification accuracies compared to a conventional training strategy.

A Robust Continuous Authentication System Using Smartphone Sensors and Wasserstein Generative Adversarial Networks

Since the continuous authentication (CA) system based on smartphone sensors has been facing the challenge of the low-data regime under some practical scenarios, which leads to low accuracy of CA, it needs to be solved urgently. To this end, currently, the generative adversarial networks (GAN) provide a powerful method to train the result generative model that could generate very convincing verisimilar data. The framework of the GAN and its variants shed much light on improving the performance of CA. Therefore, in this article, we propose a continuous authentication system on smartphones based on a Wasserstein generative adversarial network (WGAN) for sensor data augmentation, which utilizes accelerometers, gyroscopes, and magnetometers of smartphone sensors to sense phone movements caused by user operation behavior. Specifically, based on sensor data under different user activities, the WGAN is used to create additional data in training data for data augmentation. With the augmented data, we design a convolutional neural network to learn and extract deep features from sensor data, and then use four classifiers of RF, OCSVM, DT, and KNN to train these features. Finally, we train and test on the HMOG dataset, and the results show that the EER of the authentication system is between 3.68% and 6.39% on the sensor data with a time window of 2 s.

Sensor Fusion Approach for Multiple Human Motion Detection for Indoor Surveillance Use-Case.

Multi-human detection and tracking in indoor surveillance is a challenging task due to various factors such as occlusions, illumination changes, and complex human-human and human-object interactions. In this study, we address these challenges by exploring the benefits of a low-level sensor fusion approach that combines grayscale and neuromorphic vision sensor (NVS) data. We first generate a custom dataset using an NVS camera in an indoor environment. We then conduct a comprehensive study by experimenting with different image features and deep learning networks, followed by a multi-input fusion strategy to optimize our experiments with respect to overfitting. Our primary goal is to determine the best input feature types for multi-human motion detection using statistical analysis. We find that there is a significant difference between the input features of optimized backbones, with the best strategy depending on the amount of available data. Specifically, under a low-data regime, event-based frames seem to be the preferred input feature type, while higher data availability benefits the combined use of grayscale and optical flow features. Our results demonstrate the potential of sensor fusion and deep learning techniques for multi-human tracking in indoor surveillance, although it is acknowledged that further studies are needed to confirm our findings.

Physics-Informed Neural Network with Fourier Features for Radiation Transport in Heterogeneous Media

Typical machine learning (ML) methods are difficult to apply to radiation transport due to the large computational cost associated with simulating problems to create training data. Physics-informed Neural Networks (PiNNs) are a ML method that train a neural network with the residual of a governing equation as the loss function. This allows PiNNs to be trained in a low-data regime in the absence of (experimental or synthetic) data. PiNNs also are trained on points sampled within the phase-space volume of the problem, which means they are not required to be evaluated on a mesh, providing a distinct advantage in solving the linear Boltzmann transport equation, which is difficult to discretize. We have applied PiNNs to solve the streaming and interaction terms of the linear Boltzmann transport equation to create an accurate ML model that is wrapped inside a traditional source iteration process. We present an application of Fourier Features to PiNNs that yields good performance on heterogeneous problems. We also introduce a sampling method based on heuristics that improves the performance of PiNN simulations. The results are presented in a suite of one-dimensional radiation transport problems where PiNNs show very good agreement when compared to fine-mesh answers from traditional discretization techniques.

Application of quantum computing to a linear non-Gaussian acyclic model for novel medical knowledge discovery.

Recently, the utilization of real-world medical data collected from clinical sites has been attracting attention. Especially as the number of variables in real-world medical data increases, causal discovery becomes more and more effective. On the other hand, it is necessary to develop new causal discovery algorithms suitable for small data sets for situations where sample sizes are insufficient to detect reasonable causal relationships, such as rare diseases and emerging infectious diseases. This study aims to develop a new causal discovery algorithm suitable for a small number of real-world medical data using quantum computing, one of the emerging information technologies attracting attention for its application in machine learning. In this study, a new algorithm that applies the quantum kernel to a linear non-Gaussian acyclic model, one of the causal discovery algorithms, is developed. Experiments on several artificial data sets showed that the new algorithm proposed in this study was more accurate than existing methods with the Gaussian kernel under various conditions in the low-data regime. When the new algorithm was applied to real-world medical data, a case was confirmed in which the causal structure could be correctly estimated even when the amount of data was small, which was not possible with existing methods. Furthermore, the possibility of implementing the new algorithm on real quantum hardware was discussed. This study suggests that the new proposed algorithm using quantum computing might be a good choice among the causal discovery algorithms in the low-data regime for novel medical knowledge discovery.

Meta-learning biologically plausible plasticity rules with random feedback pathways

Backpropagation is widely used to train artificial neural networks, but its relationship to synaptic plasticity in the brain is unknown. Some biological models of backpropagation rely on feedback projections that are symmetric with feedforward connections, but experiments do not corroborate the existence of such symmetric backward connectivity. Random feedback alignment offers an alternative model in which errors are propagated backward through fixed, random backward connections. This approach successfully trains shallow models, but learns slowly and does not perform well with deeper models or online learning. In this study, we develop a meta-learning approach to discover interpretable, biologically plausible plasticity rules that improve online learning performance with fixed random feedback connections. The resulting plasticity rules show improved online training of deep models in the low data regime. Our results highlight the potential of meta-learning to discover effective, interpretable learning rules satisfying biological constraints.

Deep generative fuel design in low data regimes via multi-objective imitation

Commercial fuel discovery faces a constantly decreasing return of investment due to due to increasingly tight environmental criteria and reducing potential uses for each new fuel. In this paper, a deep generative model, termed Latent Interspace Generative Adversarial Network with a Domain of Stacking (LIGANDS), has been established to screen desired fuel molecules in the large chemical space without setting design rules manually. A variational autoencoder, a generative adversarial network and a stacking model are well integrated in LIGANDS through model convergence. Given only the structures of 255 typical high-energy–density fuels in low data regimes, LIGANDS generated 3461 new fuel molecules with similar property distribution and improved energy performance as the qualified candidates of next-generation fuels. To expand and enrich the fuel-relevant chemical space with innovative molecular entities on demand, in-depth multi-objective imitation on the key properties of target fuel is realized by LIGANDS through optimizing generative molecular structures and their distribution.

Federated End-to-End Unrolled Models for Magnetic Resonance Image Reconstruction.

Image reconstruction is the process of recovering an image from raw, under-sampled signal measurements, and is a critical step in diagnostic medical imaging, such as magnetic resonance imaging (MRI). Recently, data-driven methods have led to improved image quality in MRI reconstruction using a limited number of measurements, but these methods typically rely on the existence of a large, centralized database of fully sampled scans for training. In this work, we investigate federated learning for MRI reconstruction using end-to-end unrolled deep learning models as a means of training global models across multiple clients (data sites), while keeping individual scans local. We empirically identify a low-data regime across a large number of heterogeneous scans, where a small number of training samples per client are available and non-collaborative models lead to performance drops. In this regime, we investigate the performance of adaptive federated optimization algorithms as a function of client data distribution and communication budget. Experimental results show that adaptive optimization algorithms are well suited for the federated learning of unrolled models, even in a limited-data regime (50 slices per data site), and that client-sided personalization can improve reconstruction quality for clients that did not participate in training.

An Efficient Data Augmentation Method for Automatic Modulation Recognition from Low-Data Imbalanced-Class Regime

The application of deep neural networks to address automatic modulation recognition (AMR) challenges has gained increasing popularity. Despite the outstanding capability of deep learning in automatic feature extraction, predictions based on low-data regimes with imbalanced classes of modulation signals generally result in low accuracy due to an insufficient number of training examples, which hinders the wide adoption of deep learning methods in practical applications of AMR. The identification of the minority class of samples can be crucial, as they tend to be of higher value. However, in AMR tasks, there is a lack of attention and effective solutions to the problem of Imbalanced-class in a low-data regime. In this work, we present a practical automatic data augmentation method for radio signals, called SigAugment, which incorporates eight individual transformations and effectively improves the performance of AMR tasks without additional searches. It surpasses existing data augmentation methods and mainstream methods for solving low-data and imbalanced-class problems on multiple datasets. By simply embedding SigAugment into the training pipeline of an existing model, it can achieve state-of-the-art performance on benchmark datasets and dramatically improve the classification accuracy of minority classes in the low-data imbalanced-class regime. SigAugment can be trained for uniform use on different types of models and datasets and works right out of the box.

A thermodynamics-informed active learning approach to perception and reasoning about fluids

Learning and reasoning about physical phenomena is still a challenge in robotics development, and computational sciences play a capital role in the search for accurate methods able to provide explanations for past events and rigorous forecasts of future situations. We propose a thermodynamics-informed active learning strategy for fluid perception and reasoning from observations. As a model problem, we take the sloshing phenomena of different fluids contained in a glass. Starting from full-field and high-resolution synthetic data for a particular fluid, we develop a method for the tracking (perception) and simulation (reasoning) of any previously unseen liquid whose free surface is observed with a commodity camera. This approach demonstrates the importance of physics and knowledge not only in data-driven (gray-box) modeling but also in real-physics adaptation in low-data regimes and partial observations of the dynamics. The presented method is extensible to other domains such as the development of cognitive digital twins able to learn from observation of phenomena for which they have not been trained explicitly.

Putting Chemical Knowledge to Work in Machine Learning for Reactivity.

Machine learning has been used to study chemical reactivity for a long time in fields such as physical organic chemistry, chemometrics and cheminformatics. Recent advances in computer science have resulted in deep neural networks that can learn directly from the molecular structure. Neural networks are a good choice when large amounts of data are available. However, many datasets in chemistry are small, and models utilizing chemical knowledge are required for good performance. Adding chemical knowledge can be achieved either by adding more information about the molecules or by adjusting the model architecture itself. The current method of choice for adding more information is descriptors based on computed quantum-chemical properties. Exciting new research directions show that it is possible to augment deep learning with such descriptors for better performance in the low-data regime. To modify the models, differentiable programming enables seamless merging of neural networks with mathematical models from chemistry and physics. The resulting methods are also more data-efficient and make better predictions for molecules that are different from the initial dataset on which they were trained. Application of these chemistry-informed machine learning methods promise to accelerate research in fields such as drug design, materials design, catalysis and reactivity.

Modular machine learning-based elastoplasticity: Generalization in the context of limited data

The development of highly accurate constitutive models for materials that undergo path-dependent processes continues to be a complex challenge in computational solid mechanics. Challenges arise both in considering the appropriate model assumptions and from the viewpoint of data availability, verification, and validation. Recently, data-driven modeling approaches have been proposed that aim to establish stress-evolution laws that avoid user-chosen functional forms by relying on machine learning representations and algorithms. However, these approaches not only require a significant amount of data but also need data that probes the full stress space with a variety of complex loading paths. Furthermore, they rarely enforce all necessary thermodynamic principles as hard constraints. Hence, they are in particular not suitable for low-data or limited-data regimes, where the first arises from the cost of obtaining the data and the latter from the experimental limitations of obtaining labeled data, which is commonly the case in engineering applications. In this work, we discuss a hybrid framework that can work on a variable amount of data by relying on the modularity of the elastoplasticity formulation where each component of the model can be chosen to be either a classical phenomenological or a data-driven model depending on the amount of available information and the complexity of the response. The method is tested on synthetic uniaxial data coming from simulations as well as cyclic experimental data for structural materials. The discovered material models are found to not only interpolate well but also allow for accurate extrapolation in a thermodynamically consistent manner far outside the domain of the training data. This ability to extrapolate from limited data was the main reason for the early and continued success of phenomenological models and the main shortcoming in machine learning-enabled constitutive modeling approaches. Training aspects and details of the implementation of these models into Finite Element simulations are discussed and analyzed.