Gabriel Fricout

A fast learning algorithm for image segmentation with max-pooling convolutional networks

We present a fast algorithm for training MaxPooling Convolutional Networks to segment images. This type of network yields record-breaking performance in a variety of tasks, but is normally trained on a computationally expensive patch-by-patch basis. Our new method processes each training image in a single pass, which is vastly more efficient. We validate the approach in different scenarios and report a 1500-fold speed-up. In an application to automated steel defect detection and segmentation, we obtain excellent performance with short training times.

Steel defect classification with Max-Pooling Convolutional Neural Networks

We present a Max-Pooling Convolutional Neural Network approach for supervised steel defect classification. On a classification task with 7 defects, collected from a real production line, an error rate of 7% is obtained. Compared to SVM classifiers trained on commonly used feature descriptors our best net performs at least two times better. Not only we do obtain much better results, but the proposed method also works directly on raw pixel intensities of detected and segmented steel defects, avoiding further time consuming and hard to optimize ad-hoc preprocessing.

Kalman Temporal Differences: The deterministic case

This paper deals with value function and Q-function approximation in deterministic Markovian decision processes. A general statistical framework based on the Kalman filtering paradigm is introduced. Its principle is to adopt a parametric representation of the value function, to model the associated parameter vector as a random variable and to minimize the mean-squared error of the parameters conditioned on past observed transitions. From this general framework, which will be called Kalman Temporal Differences (KTD), and using an approximation scheme called the unscented transform, a family of algorithms is derived, namely KTD-V, KTD-SARSA and KTD-Q, which aim respectively at estimating the value function of a given policy, the Q-function of a given policy and the optimal Q-function. The proposed approach holds for linear and nonlinear parameterization. This framework is discussed and potential advantages and shortcomings are highlighted.

Tracking in Reinforcement Learning

Reinforcement learning induces non-stationarity at several levels. Adaptation to non-stationary environments is of course a desired feature of a fair RL algorithm. Yet, even if the environment of the learning agent can be considered as stationary, generalized policy iteration frameworks, because of the interleaving of learning and control, will produce non-stationarity of the evaluated policy and so of its value function. Tracking the optimal solution instead of trying to converge to it is therefore preferable. In this paper, we propose to handle this tracking issue with a Kalman-based temporal difference framework. Complexity and convergence analysis are studied. Empirical investigations of its ability to handle non-stationarity is finally provided.

Bayesian Reward Filtering

A wide variety of function approximation schemes have been applied to reinforcement learning. However, Bayesian filtering approaches, which have been shown efficient in other fields such as neural network training, have been little studied. We propose a general Bayesian filtering framework for reinforcement learning, as well as a specific implementation based on sigma point Kalman filtering and kernel machines. This allows us to derive an efficient off-policy model-free approximate temporal differences algorithm which will be demonstrated on two simple benchmarks.

A Sparse Nonlinear Bayesian Online Kernel Regression

In a large number of applications, engineers have to estimate values of an unknown function given some observed samples. This task is referred to as function approximation or as generalization. One way to solve the problem is to regress a family of parameterized functions so as to make it fit at best the observed samples. Yet, usually batch methods are used and parameterization is habitually linear. Moreover, very few approaches try to quantify uncertainty reduction occurring when acquiring more samples (thus more information), which can be quite useful depending on the application. In this paper we propose a sparse nonlinear Bayesian online kernel regression. Sparsity is achieved in a preprocessing step by using a dictionary method. The nonlinear Bayesian kernel regression can therefore be considered as achieved online by a Sigma Point Kalman filter. First experiments on a cardinal sine regression show that our approach is promising.

Multi-scale pyramidal pooling network for generic steel defect classification

We introduce a Multi-Scale Pyramidal Pooling Network tailored to generic steel defect classification, featuring a novel pyramidal pooling layer at multiple scales and a novel encoding layer. Thanks to the former, the network does not require all images of a given classification task to be of equal size. The latter narrows the gap to bag-of-features approaches. On various benchmark datasets, we evaluate and compare our system to convolutional neural networks and state-of-the-art computer vision methods. We also present results on a real industrial steel defect classification problem, where existing architectures are not applicable as they require equally sized input images. Our method substantially outperforms previous methods based on engineered features. It can be seen as a fully supervised hierarchical bag-of-features extension that is trained online and can be fine-tuned for any given task.

Numerical simulation of thin paint film flow

PurposeBeing able to predict the visual appearance of a painted steel sheet, given its topography before paint application, is of crucial importance for car makers. Accurate modeling of the industrial painting process is required.ResultsThe equations describing the leveling of the paint film are complex and their numerical simulation requires advanced mathematical tools, which are described in detail in this paper. Simulations are validated using a large experimental data base obtained with a wavefront sensor developed by Phasics™.ConclusionsThe conducted simulations are complex and require the development of advanced numerical tools, like those presented in this paper.

Online Bayesian kernel regression from nonlinear mapping of observations

In a large number of applications, engineers have to estimate a function linked to the state of a dynamic system. To do so, a sequence of samples drawn from this unknown function is observed while the system is transiting from state to state and the problem is to generalize these observations to unvisited states. Several solutions can be envisioned among which regressing a family of parameterized functions so as to make it fit at best to the observed samples. However classical methods cannot handle the case where actual samples are not directly observable but only a nonlinear mapping of them is available, which happen when a special sensor has to be used or when solving the Bellman equation in order to control the system. This paper introduces a method based on Bayesian filtering and kernel machines designed to solve the tricky problem at sight. First experimental results are promising.

A Reference Architecture for Quality Improvement in Steel Production

Es gibt weltweit einen erhohten Bedarf an Stahl, aber die Stahlherstellung ist ein enorm anspruchsvoller und kostenintensiver Prozess, bei dem gute Qualitat schwer zu erreichen ist. Die Verbesserung der Qualitat ist noch immer die groste Herausforderung, der sich die Stahlbranche gegenuber sieht. Das EUProjekt PRESED (Predictive Sensor Data Mining for Product Quality Improvement) [Vorrausschauende Sensordatengewinnung zur Verbesserung der Produktqualitat] stellt sich dieser Herausforderung durch die Fokussierung auf weitverbreitete, wiederkehrende Probleme. Die Vielfalt und Richtigkeit der Daten sowie die Veranderung der Eigenschaften des untersuchten Materials erschwert die Interpretation der Daten. In dieser Abhandlung stellen wir die Referenzarchitektur von PRESED vor, die speziell angefertigt wurde, um die zentralen Anliegen der Verwaltung und Operationalisierung von Daten zu thematisieren. Die Architektur kombiniert grose und intelligente Datenkonzepte mit Datengewinnungsalgorithmen. Datenvorverarbeitung und vorausschauende Analyseaufgaben werden durch ein plastisches Datenmodell unterstutzt. Der Ansatz erlaubt es den Nutzern, Prozesse zu gestalten und mehrere Algorithmen zu bewerten, die sich gezielt mit dem vorliegenden Problem befassen. Das Konzept umfasst die Sicherung und Nutzung vollstandiger Produktionsdaten, anstatt sich auf aggregierte Werte zu verlassen. Erste Ergebnisse der Datenmodellierung zeigen, dass die detailgenaue Vorverarbeitung von Zeitreihendaten durch Merkmalserkennung und Prognosen im Vergleich zu traditionell verwendeter Aggregationsstatistik uberlegene Erkenntnisse bietet.