An Optimization Technique for Linear Manifold Learning-Based Dimensionality Reduction: Evaluations on Hyperspectral Images

Manifold learning tries to find low-dimensional manifolds on high-dimensional data. It is useful to omit redundant data from input. Linear manifold learning algorithms have applicability for out-of-sample data, in which they are fast and practical especially for classification purposes. Locality preserving projection (LPP) and orthogonal locality preserving projection (OLPP) are two known linear manifold learning algorithms. In this study, scatter information of a distance matrix is used to construct a weight matrix with a supervised approach for the LPP and OLPP algorithms to improve classification accuracy rates. Low-dimensional data are classified with SVM and the results of the proposed method are compared with some other important existing linear manifold learning methods. Class-based enhancements and coefficients proposed for the formulization are reported visually. Furthermore, the change on weight matrices, band information, and correlation matrices with p-values are extracted and visualized to understand the effect of the proposed method. Experiments are conducted on hyperspectral imaging (HSI) with two different datasets. According to the experimental results, application of the proposed method with the LPP or OLPP algorithms outperformed traditional LPP, OLPP, neighborhood preserving embedding (NPE) and orthogonal neighborhood preserving embedding (ONPE) algorithms. Furthermore, the analytical findings on visualizations show consistency with obtained classification accuracy enhancements.

Manifold reduction techniques for the comparison of crank angle-resolved particle image velocimetry (PIV) data and Reynolds-averaged Navier-Stokes (RANS) simulations in a spark ignition direct injection (SIDI) engine

In this article, different manifold reduction techniques are implemented for the post-processing of Particle Image Velocimetry (PIV) images from a Spark Ignition Direct Injection (SIDI) engine. The methods are proposed to help make a more objective comparison between Reynolds-averaged Navier-Stokes (RANS) simulations and PIV experiments when Cycle-to-Cycle Variations (CCV) are present in the flow field. The two different methods used here are based on Singular Value Decomposition (SVD) principles where Proper Orthogonal Decomposition (POD) and Kernel Principal Component Analysis (KPCA) are used for representing linear and non-linear manifold reduction techniques. To the authors’ best knowledge, this is the first time a non-linear manifold reduction technique, such as KPCA, has ever been used in the study of in-cylinder flow fields. Both qualitative and quantitative studies are given to show the capability of each method in validating the simulation and incorporating CCV for each engine cycle. Traditional Relevance Index (RI) and two other previously developed novel indexes: the Weighted Relevance Index (WRI) and the Weighted Magnitude Index (WMI), are used for the quantitative study. The results indicate that both POD and KPCA show improvements in capturing the main flow field features compared to ensemble-averaged PIV experimental data and single cycle experimental flow fields while capturing CCV. Both methods present similar quantitative accuracy when using the three indexes. However, challenges were highlighted in the POD method for the selection of the number of POD modes needed for a representative reconstruction. When the flow field region presents a Gaussian distribution, the KPCA method is seen to provide a more objective numerical process as the reconstructed flow field will see convergence with an increasing number of modes due to its usage of Gaussian properties. No additional criterion is needed to determine how to reconstruct the main flow field feature. Using KPCA can, therefore, reduce the amount of analysis needed in the process of extracting the main flow field while incorporating CCV.

Non-linear manifold learning in fMRI uncovers a low-dimensional space of brain dynamics

Large-scale brain dynamics are believed to lie in a latent, low-dimensional space. Typically, the embeddings of brain scans are derived independently from different cognitive tasks or resting-state data, ignoring a potentially large-and shared-portion of this space. Here, we establish that a shared, robust, and interpretable low-dimensional space of brain dynamics can be recovered from a rich repertoire of task-based fMRI data. This occurs when relying on non-linear approaches as opposed to traditional linear methods. The embedding maintains proper temporal progression of the tasks, revealing brain states and the dynamics of network integration. We demonstrate that resting-state data embeds fully onto the same task embedding, indicating similar brain states are present in both task and resting-state data. Our findings suggest analysis of fMRI data from multiple cognitive tasks in a low-dimensional space is possible and desirable, and our proposed framework can thus provide an interpretable framework to investigate brain dynamics in the low-dimensional space.

msiPL: Non-linear Manifold and Peak Learning of Mass Spectrometry Imaging Data Using Artificial Neural Networks

Mass spectrometry imaging (MSI) is an emerging technology that holds potential for improving clinical diagnosis, biomarker discovery, metabolomics research and pharmaceutical applications. The large data size and high dimensional nature of MSI pose computational and memory complexities that hinder accurate identification of biologically-relevant molecular patterns. We propose msiPL, a robust and generic probabilistic generative model based on a fully-connected variational autoencoder for unsupervised analysis and peak learning of MSI data. The method can efficiently learn and visualize the underlying non-linear spectral manifold, reveal biologically-relevant clusters of tumor heterogeneity and identify underlying informative m/z peaks. The method provides a probabilistic parametric mapping to allow a trained model to rapidly analyze a new unseen MSI dataset in a few seconds. The computational model features a memory-efficient implementation using a minibatch processing strategy to enable the analyses of big MSI data (encompassing more than 1 million high-dimensional datapoints) with significantly less memory. We demonstrate the robustness and generic applicability of the application on MSI data of large size from different biological systems and acquired using different mass spectrometers at different centers, namely: 2D Matrix-Assisted Laser Desorption Ionization (MALDI) Fourier Transform Ion Cyclotron Resonance (FT ICR) MSI data of human prostate cancer, 3D MALDI Time-of-Flight (TOF) MSI data of human oral squamous cell carcinoma, 3D Desorption Electrospray Ionization (DESI) Orbitrap MSI data of human colorectal adenocarcinoma, 3D MALDI TOF MSI data of mouse kidney, and 3D MALDI FT ICR MSI data of a patient-derived xenograft (PDX) mouse brain model of glioblastoma.SignificanceMass spectrometry imaging (MSI) provides detailed molecular characterization of a tissue specimen while preserving spatial distributions. However, the complex nature of MSI data slows down the processing time and poses computational and memory challenges that hinder the analysis of multiple specimens required to extract biologically relevant patterns. Moreover, the subjectivity in the selection of parameters for conventional pre-processing approaches can lead to bias. Here, we present a generative probabilistic deep-learning model that can analyze and non-linearly visualize MSI data independent of the nature of the specimen and of the MSI platform. We demonstrate robustness of the method with application to different tissue types, and envision it as a new generation of rapid and robust analysis for mass spectrometry data.

Functional regression on the manifold with contamination

Summary We propose a new method for functional nonparametric regression with a predictor that resides on a finite-dimensional manifold, but is observable only in an infinite-dimensional space. Contamination of the predictor due to discrete or noisy measurements is also accounted for. By using functional local linear manifold smoothing, the proposed estimator enjoys a polynomial rate of convergence that adapts to the intrinsic manifold dimension and the contamination level. This is in contrast to the logarithmic convergence rate in the literature of functional nonparametric regression. We also observe a phase transition phenomenon related to the interplay between the manifold dimension and the contamination level. We demonstrate via simulated and real data examples that the proposed method has favourable numerical performance relative to existing commonly used methods.

Predicting clinical drug response from model systems by non-linear subspace-based transfer learning

Pre-clinical models have been the workhorse of cancer research for decades. While powerful, these models do not fully recapitulate the complexity of human tumors. Consequently, translating biomarkers of drug response from pre-clinical models to human tumors has been particularly challenging. To explicitly take these differences into account and enable an efficient exploitation of the vast pre-clinical drug response resources, we developed TRANSACT, a novel computational framework for clinical drug response prediction. First, TRANSACT employs non-linear manifold learning to capture biological processes active in pre-clinical models and human tumors. Then, TRANSACT builds predictors on cell line response only and transfers these to Patient-Derived Xenografts (PDXs) and human tumors. TRANSACT outperforms four competing approaches, including Deep Learning approaches, for a set of 15 drugs on PDXs, TCGA cohorts and 226 metastatic tumors from the Hartwig Medical Foundation data. For only four drugs Deep Learning outperforms TRANSACT. We further derived an algorithmic approach to interpret TRANSACT and used it to validate the approach by identifying known biomarkers to targeted therapies and we propose novel putative biomarkers of resistance to Paclitaxel and Gemcitabine.

Pressure distribution classification and segmentation of human hands in contact with the robot body

This article deals with the problem of the recognition of human hand touch by a robot equipped with large area tactile sensors covering its body. This problem is relevant in the domain of physical human–robot interaction for discriminating between human and non-human contacts and to trigger and to drive cooperative tasks or robot motions, or to ensure a safe interaction. The underlying assumption used in this article is that voluntary physical interaction tasks involve hand touch over the robot body, and therefore the capability to recognize hand contacts is a key element to discriminate a purposive human touch from other types of interaction. The proposed approach is based on a geometric transformation of the tactile data, formed by pressure measurements associated to a non-uniform cloud of 3D points ( taxels) spread over a non-linear manifold corresponding to the robot body, into tactile images representing the contact pressure distribution in two dimensions. Tactile images can be processed using deep learning algorithms to recognize human hands and to compute the pressure distribution applied by the various hand segments: palm and single fingers. Experimental results, performed on a real robot covered with robot skin, show the effectiveness of the proposed methodology. Moreover, to evaluate its robustness, various types of failures have been simulated. A further analysis concerning the transferability of the system has been performed, considering contacts occurring on a different sensorized robot part.