Guillermo L. Grinblat

Deep learning for plant identification using vein morphological patterns

Display Omitted Deep convolutional neural network (CNN) for plant identification focusing on leaf vein patterns.No task-specific feature extractors needed.Improved the state of the art accuracy on a legume species recognition task.Visualization of relevant vein patterns. We propose using a deep convolutional neural network (CNN) for the problem of plant identification from leaf vein patterns. In particular, we consider classifying three different legume species: white bean, red bean and soybean. The introduction of a CNN avoids the use of handcrafted feature extractors as it is standard in state of the art pipeline. Furthermore, this deep learning approach significantly improves the accuracy of the referred pipeline. We also show that the reported accuracy is reached by increasing the model depth. Finally, by analyzing the resulting models with a simple visualization technique, we are able to unveil relevant vein patterns.

Toward Large-Scale Vulnerability Discovery using Machine Learning

With sustained growth of software complexity, finding security vulnerabilities in operating systems has become an important necessity. Nowadays, OS are shipped with thousands of binary executables. Unfortunately, methodologies and tools for an OS scale program testing within a limited time budget are still missing. In this paper we present an approach that uses lightweight static and dynamic features to predict if a test case is likely to contain a software vulnerability using machine learning techniques. To show the effectiveness of our approach, we set up a large experiment to detect easily exploitable memory corruptions using 1039 Debian programs obtained from its bug tracker, collected 138,308 unique execution traces and statically explored 76,083 different subsequences of function calls. We managed to predict with reasonable accuracy which programs contained dangerous memory corruptions. We also developed and implemented VDiscover, a tool that uses state-of-the-art Machine Learning techniques to predict vulnerabilities in test cases. Such tool will be released as open-source to encourage the research of vulnerability discovery at a large scale, together with VDiscovery, a public dataset that collects raw analyzed data.

Abrupt change detection with One-Class Time-Adaptive Support Vector Machines

Abstract We recently introduced an algorithm for training a sequence of coupled Support Vector Machines which shows promising results in the field of non-stationary classification problems Grinblat, Uzal, Ceccatto, and Granitto (2011) . In this paper we analyze its application to the abrupt change detection problem. With this goal, we first introduce and analyze an extension of it to deal with the One-Class Support Vector Machine (OC-SVM) problem, and then discuss its use as an improved abrupt change detection method. Finally, we apply the proposed procedure to artificial and real-world examples, and demonstrate that it is competitive by comparison against other abrupt change detection methods.

REPMAC: A New Hybrid Approach to Highly Imbalanced Classification Problems

The class imbalance problem (when one of the classes has much less samples than the others) is of great importance in machine learning, because it corresponds to many critical applications. In this work we introduce the recursive partitioning of the majority class (REPMAC) algorithm, a new hybrid method to solve imbalanced problems. Using a clustering method, REPMAC recursively splits the majority class in several subsets, creating a decision tree, until the resulting sub-problems are balanced or easy to solve. At that point, a classifier is fitted to each sub-problem. We evaluate the new method on 7 datasets from the UCI repository, finding that REPMAC is more efficient than other methods usually applied to imbalanced datasets.

SVM based feature selection: why are we using the dual?

Most Support Vector Machines (SVM) implementations are based on solving the dual optimization problem. Of course, feature selection algorithms based on SVM are not different and, in particular, the most used method in the area, Guyon et al.s Recursive Feature Elimination (SVM-RFE) is also based on the dual problem. However, this is just one of the options available to find a solution to the original SVM optimization problem. In this work we discuss some potential problems that arise when ranking features with the dual-based version of SVM-RFE and propose a primal-based version of this well-known method, PSVM-RFE. We show that our new method is able to produce a better detection of relevant features, in particular in situations involving non-linear decision boundaries. Using several artificial and real-world datasets we compare both versions of SVM-RFE, finding that PSVM-RFE is preferable in most situations.

Nonstationary regression with support vector machines

Abstract In this work, we introduce a method for data analysis in nonstationary environments: time-adaptive support vector regression (TA-SVR). The proposed approach extends a previous development that was limited to classification problems. Focusing our study on time series applications, we show that TA-SVR can improve the accuracy of several aspects of nonstationary data analysis, namely the tasks of modelling and prediction, input relevance estimation, and reconstruction of a hidden forcing profile.

Seed-per-pod estimation for plant breeding using deep learning

Abstract Commercial and scientific plant breeding programs require the phenotyping of large populations. Phenotyping is typically a manual task (costly, time-consuming and sometimes arbitrary). The use of computer vision techniques is a potential solution to some of these specific tasks. In the last years, Deep Learning, and in particular Convolutional Neural Networks (CNNs), have shown a number of advantages over traditional methods in the area. In this work we introduce a computer vision method that estimates the number of seeds into soybean pods, a difficult task that usually requires the intervention of human experts. To this end we developed a classic approach, based on tailored features extraction (FE) followed by a Support Vector Machines (SVM) classification model, and also the referred CNNs. We show how standard CNNs can be easily configured and how a simple method can be used to visualize the key features learned by the model in order to infer the correct class. We processed different seasons batches with both methods obtaining 50.4% (FE + SVM) and 86.2% (CNN) of accuracy in test, highlighting the particularly high increase in generalization capabilities of a deep learning approach over a classic machine vision approach in this task. Dataset and code are publicly available.

Unsupervized data-driven partitioning of multiclass problems

Many classification problems of high technological value are multiclass. In the last years, several improved solutions based on the combination of simple classifiers were introduced. An interesting kind of methods creates a hierarchy of sub-problems by clustering prototypes of each one of the classes, but the solution produced by the clustering stage is heavily influenced by the labels information. In this work we introduce a new strategy to solve multiclass problems that makes more use of spatial information than other methods. Based on our previous work on imbalanced problems, we construct a hierarchy of subproblems, but opposite to previous developments, based only on spatial information and not using class labels at any time. We consider different clustering methods (either agglomerative or divisive) for this task. We use an SVM for each sub-problem (if needed, because in several cases the clustering method directly gives a subset with samples of a single class). Using publicly available datasets we compare the new method with several previous approaches, finding promising results.

Evaluation of a new hybrid algorithm for highly imbalanced classification problems

Many times in classification problems, particularly in critical real world applications, one of the classes has much less samples than the others usually known as the class imbalance problem. In this work we discuss and evaluate the use of the REPMAC algorithm to solve imbalanced problems. Using a clustering method, REPMAC recursively splits the majority class in several subsets, creating a decision tree, until the resulting sub-problems are balanced or easy to solve. We use two diverse clustering methods and three different classifiers coupled with REPMAC to evaluate the new method on several benchmark datasets spanning a wide range of number of features, samples and imbalance degree. We also apply our method to a real world problem, the identification of weed seeds. We find that the good performance of REPMAC is almost independent of the classifier or the clustering method coupled to it, which suggests that its success is mostly related to the use of an appropriate strategy to cope with imbalanced problems.