Papers by Nina S. T. Hirata
Information Visualization, 2020
Dimensionality reduction methods, also known as projections, are often used to explore multidimen... more Dimensionality reduction methods, also known as projections, are often used to explore multidimensional data in machine learning, data science, and information visualization. However, several such methods, such as the well-known t-distributed stochastic neighbor embedding and its variants, are computationally expensive for large datasets, suffer from stability problems, and cannot directly handle out-of-sample data. We propose a learning approach to construct any such projections. We train a deep neural network based on sample set drawn from a given data universe, and their corresponding two-dimensional projections, compute with any user-chosen technique. Next, we use the network to infer projections of any dataset from the same universe. Our approach generates projections with similar characteristics as the learned ones, is computationally two to four orders of magnitude faster than existing projection methods, has no complex-to-set user parameters, handles out-of-sample data in a ...
arXiv (Cornell University), Feb 21, 2019
Dimensionality reduction methods, also known as projections, are frequently used for exploring mu... more Dimensionality reduction methods, also known as projections, are frequently used for exploring multidimensional data in machine learning, data science, and information visualization. Among these, t-SNE and its variants have become very popular for their ability to visually separate distinct data clusters. However, such methods are computationally expensive for large datasets, suffer from stability problems, and cannot directly handle out-of-sample data. We propose a learning approach to construct such projections. We train a deep neural network based on a collection of samples from a given data universe, and their corresponding projections, and next use the network to infer projections of data from the same, or similar, universes. Our approach generates projections with similar characteristics as the learned ones, is computationally two to three orders of magnitude faster than SNE-class methods, has no complex-to-set user parameters, handles out-of-sample data in a stable manner, and can be used to learn any projection technique. We demonstrate our proposal on several real-world high dimensional datasets from machine learning.
Automatic Programming of MMach’s for OCR
Springer eBooks, 1996
Binary Image Analysis problems can be solved by set operators implemented as programs for a Morph... more Binary Image Analysis problems can be solved by set operators implemented as programs for a Morphological Machine (MMach). These programs can be generated automatically by the description of the goals of the user as a collection of input-output image pairs and the estimation of the target operator from these data. In this paper, we present a software, installed as a Toolbox for the KHOROS system, that implements this technique and some impressive results of applying this tool in shape recognition for OCR.
<title>Automatic programming of binary morphological machines by PAC learning</title>
Proceedings of SPIE, Aug 11, 1995
Binary image analysis problems can be solved by set operators implemented as programs for a binar... more Binary image analysis problems can be solved by set operators implemented as programs for a binary morphological machine (BMM). This is a very general and powerful approach to solve this type of problem. However, the design of these programs is not a task manageable by nonexperts on mathematical morphology. In order to overcome this difficulty we have worked on tools that help users describe their goals at higher levels of abstraction and to translate them into BMM programs. Some of these tools are based on the representation of the goals of the user as a collection of input-output pairs of images and the estimation of the target operator from these data. PAC learning is a well suited methodology for this task, since in this theory 'concepts' are represented as Boolean functions that are equivalent to set operators. In order to apply this technique in practice we must have efficient learning algorithms. In this paper we introduce two PAC learning algorithms, both are based on the minimal representation of Boolean functions, which has a straightforward translation to the canonical decomposition of set operators. The first algorithm is based on the classical Quine-McCluskey algorithm for the simplification of Boolean functions, and the second one is based on a new idea for the construction of Boolean functions: the incremental splitting of intervals. We also present a comparative complexity analysis of the two algorithms. Finally, we give some application examples.
NILC: A two level learning algorithm with operator selection
Machine learning is a very promising way of solving some image processing tasks. However, existin... more Machine learning is a very promising way of solving some image processing tasks. However, existing approaches fails at integrating feature selection within the learning task. This paper introduces a new two stage learning algorithm called near infinitely linear combination (NILC) that performs at the same time variable selection and error minimization. Empirical evidence reported on different document processing tasks shows that our approach significantly outperforms existing approaches.
Chapter 3 Morphological Operator Design from Training Data A State of the Art Overview
Mathematical morphology offers a set of powerful tools for im- age processing and analysis. From ... more Mathematical morphology offers a set of powerful tools for im- age processing and analysis. From a practical perspective, the expected re- sults of many morphological operators can be intuitively explained in terms of geometrical and topological characteristics of the images. From a formal perspective, mathematical morphology is based on complete lattices, which provides a solid theoretical framework for the study of algebraic properties of the operators. Despite of these nice characteristics, designing morpholog- ical operators is not a trivial task; it requires knowledge and experience. In this chapter, a self-contained exposition on the design of translation-invariant morphological operators from training data is presented. The described train- ing procedure relies on the canonical sup-decomposition theorem of mor- phological operators, which in the context of binary images states that any translation-invariant operator can be expressed uniquely in terms of two ele- mentary operators, erosions and dilations, plus set operations. An important issue considered in this exposition is how the bias-variance tradeoff manifests within the training context and how its understanding can lead to approaches that generate better results. Several application examples that illustrate the usefulness of the described design procedure are also presented.
Automatic programming of binary morphological machines by design of statistically optimal operators in the context of computational learning theory
Journal of Electronic Imaging, 1997
Representation of set operators by artificial neural networks and design of such operators by inf... more Representation of set operators by artificial neural networks and design of such operators by inference of network parameters is a popular technique in binary image analysis. We propose an alternative.
arXiv (Cornell University), Apr 23, 2020
In Astronomy, a huge amount of image data is generated daily by photometric surveys, which scan t... more In Astronomy, a huge amount of image data is generated daily by photometric surveys, which scan the sky to collect data from stars, galaxies and other celestial objects. In this paper, we propose a technique to leverage unlabeled astronomical images to pre-train deep convolutional neural networks, in order to learn a domain-specific feature extractor which improves the results of machine learning techniques in setups with small amounts of labeled data available. We show that our technique produces results which are in many cases better than using ImageNet pre-training.
Automated in situ plankton image classification is a challenging task. To take advantage of recen... more Automated in situ plankton image classification is a challenging task. To take advantage of recent progress in machine learning techniques, a large amount of labeled data is necessary. However, beyond being time consuming, labeling is a task that may require frequent redoing due to variations in plankton population as well as image characteristics. Transfer learning, which is a machine learning technique concerned with transferring knowledge obtained in some data domain to a second distinct data domain, appears as a potential approach to be employed in this scenario. We use convolutional neural networks, trained on publicly available distinct datasets, to extract features from our plankton image data and then train SVM classifiers to perform the classification. Results show evidences that indicate the effectiveness of transfer learning in real plankton image classification situations.
Monthly Notices of the Royal Astronomical Society, Aug 5, 2019
The Southern Photometric Local Universe Survey (S-PLUS) is imaging ∼9300 deg 2 of the celestial s... more The Southern Photometric Local Universe Survey (S-PLUS) is imaging ∼9300 deg 2 of the celestial sphere in 12 optical bands using a dedicated 0.8 m robotic telescope, the T80-South, at the Cerro Tololo Inter-american Observatory, Chile. The telescope is equipped with a 9.2k × 9.2k e2v detector with 10 μm pixels, resulting in a field of view of 2 deg 2 with a plate scale of 0.55 arcsec pixel −1. The survey consists of four main subfields, which include two non-contiguous fields at high Galactic latitudes (|b| > 30 • , 8000 deg 2) and two areas of the Galactic Disc and Bulge (for an additional 1300 deg 2). S-PLUS uses the Javalambre 12-band magnitude system, which includes the 5 ugriz broad-band filters and 7 narrow-band filters centred on prominent stellar spectral features: the Balmer jump/[OII], Ca H + K, H δ, G band, Mg b triplet, H α, and the Ca triplet. S-PLUS delivers accurate photometric redshifts (δ z /(1 + z) = 0.02 or better) for galaxies with r < 19.7 AB mag and z < 0.4, thus producing a 3D map of the local Universe over a volume of more than 1 (Gpc/h) 3. The final S-PLUS catalogue will also enable the study of star formation and stellar populations in and around the Milky Way and nearby galaxies, as well as searches for quasars, variable sources, and low-metallicity stars. In this paper we introduce the main characteristics of the survey, illustrated with science verification data highlighting the unique capabilities of S-PLUS. We also present the first public data release of ∼336 deg 2 of the Stripe 82 area, in 12 bands, to a limiting magnitude of r = 21, available at datalab.noao.edu/splus.
Visualizing High-Dimensional Functions with Dense Maps
SN Computer Science
2021 34th SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI)
We address the problem of training Object Detection models using significantly less bounding box ... more We address the problem of training Object Detection models using significantly less bounding box annotated images. For that, we take advantage of cheaper and more abundant image classification data. Our proposal consists in automatically generating artificial detection samples, with no need of expensive detection level supervision, using images with classification labels only. We also detail a pretraining initialization strategy for detection architectures using these artificially synthesized samples, before finetuning on real detection data, and experimentally show how this consistently leads to more data efficient models. With the proposed approach, we were able to effectively use only classification data to improve results on the harder and more supervision hungry object detection problem. We achieve results equivalent to those of the full data scenario using only a small fraction of the original detection data for Face, Bird, and Car detection.
Proceedings of the International Astronomical Union, 2020
We present a machine learning methodology to separate quasars from galaxies and stars using data ... more We present a machine learning methodology to separate quasars from galaxies and stars using data from S-PLUS in the Stripe-82 region. In terms of quasar classification, we achieved 95.49% for precision and 95.26% for recall using a Random Forest algorithm. For photometric redshift estimation, we obtained a precision of 6% using k-Nearest Neighbour.
Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, 2018
Automated in situ plankton image classification is a challenging task. To take advantage of recen... more Automated in situ plankton image classification is a challenging task. To take advantage of recent progress in machine learning techniques, a large amount of labeled data is necessary. However, beyond being time consuming, labeling is a task that may require frequent redoing due to variations in plankton population as well as image characteristics. Transfer learning, which is a machine learning technique concerned with transferring knowledge obtained in some data domain to a second distinct data domain, appears as a potential approach to be employed in this scenario. We use convolutional neural networks, trained on publicly available distinct datasets, to extract features from our plankton image data and then train SVM classifiers to perform the classification. Results show evidences that indicate the effectiveness of transfer learning in real plankton image classification situations.
Communications in Computer and Information Science, 2022
Visualization of multidimensional data is a difficult task, for which there are many tools. Among... more Visualization of multidimensional data is a difficult task, for which there are many tools. Among these tools, dimensionality reduction methods were shown to be particularly helpful to explore data visually. Techniques with good visual separation are very popular, such as those from the SNE-class, but those often are computationally expensive and non-parametric. An approach based on neural networks was recently proposed to address those shortcomings, but it introduces some fuzziness in the generated projection, which is not desired. In this paper we thoroughly explain the parameter space of this neural network approach and propose a new neighborhood-based learning paradigm, which further improves the quality of the projections learned by the neural networks, and we illustrate our approach on large real-world datasets.
An Evaluation of Deep Learning Techniques for Qr Code Detection
2019 IEEE International Conference on Image Processing (ICIP), 2019
In this work, we employ deep learning models for detecting QR Codes in natural scenes. A series o... more In this work, we employ deep learning models for detecting QR Codes in natural scenes. A series of different model configurations are evaluated in terms of Average Precision, and an architecture modification that allows detection aided by object subparts annotations is proposed. This modification is implemented in our best scoring model, which is compared to a traditional technique, achieving a substantial improvement in the considered metrics. The dataset used in our evaluation, with bounding box annotations for both QR Codes and their Finder Patterns (FIPs), will be made publicly available. This dataset is significantly bigger than known available options at the moment, so we expect it to provide a common benchmark tool for QR Code detection in natural scenes.
Information Visualization, 2020
Dimensionality reduction methods, also known as projections, are often used to explore multidimen... more Dimensionality reduction methods, also known as projections, are often used to explore multidimensional data in machine learning, data science, and information visualization. However, several such methods, such as the well-known t-distributed stochastic neighbor embedding and its variants, are computationally expensive for large datasets, suffer from stability problems, and cannot directly handle out-of-sample data. We propose a learning approach to construct any such projections. We train a deep neural network based on sample set drawn from a given data universe, and their corresponding two-dimensional projections, compute with any user-chosen technique. Next, we use the network to infer projections of any dataset from the same universe. Our approach generates projections with similar characteristics as the learned ones, is computationally two to four orders of magnitude faster than existing projection methods, has no complex-to-set user parameters, handles out-of-sample data in a ...
NILC: A two level learning algorithm with operator selection
2016 IEEE International Conference on Image Processing (ICIP), 2016
Machine learning is a very promising way of solving some image processing tasks. However, existin... more Machine learning is a very promising way of solving some image processing tasks. However, existing approaches fails at integrating feature selection within the learning task. This paper introduces a new two stage learning algorithm called near infinitely linear combination (NILC) that performs at the same time variable selection and error minimization. Empirical evidence reported on different document processing tasks shows that our approach significantly outperforms existing approaches.
Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 2022
An algorithm for digital signal analysis using convolutional neural networks (CNN) was developed ... more An algorithm for digital signal analysis using convolutional neural networks (CNN) was developed in this work. The main objective of this algorithm is to make the analysis of experiments with active target time projection chambers more efficient. The code is divided in three steps: baseline correction, signal deconvolution and peak detection and integration. The CNNs were able to learn the signal processing models with relative errors of less than 6%. The analysis based on CNNs provides the same results as the traditional deconvolution algorithms, but considerably more efficient in terms of computing time (about 65 times faster). This opens up new possibilities to improve existing codes and to simplify the analysis of the large amount of data produced in active target experiments.
IEEE Transactions on Visualization and Computer Graphics, 2019
Dimensionality reduction methods, also known as projections, are frequently used in multidimensio... more Dimensionality reduction methods, also known as projections, are frequently used in multidimensional data exploration in machine learning, data science, and information visualization. Tens of such techniques have been proposed, aiming to address a wide set of requirements, such as ability to show the high-dimensional data structure, distance or neighborhood preservation, computational scalability, stability to data noise and/or outliers, and practical ease of use. However, it is far from clear for practitioners how to choose the best technique for a given use context. We present a survey of a wide body of projection techniques that helps answering this question. For this, we characterize the input data space, projection techniques, and the quality of projections, by several quantitative metrics. We sample these three spaces according to these metrics, aiming at good coverage with bounded effort. We describe our measurements and outline observed dependencies of the measured variables. Based on these results, we draw several conclusions that help comparing projection techniques, explain their results for different types of data, and ultimately help practitioners when choosing a projection for a given context. Our methodology, datasets, projection implementations, metrics, visualizations, and results are publicly open, so interested stakeholders can examine and/or extend this benchmark.
Uploads
Papers by Nina S. T. Hirata