LIONlab seminars
By members and visitors
Learning from examples in optical imaging
May 12, 2016 Speaker: Prof. Demetri Psaltis, EPFL Abstract: Optical tomography has been widely investigated for biomedical imaging applications. In recent years optical tomography has been combined with digital holography and has been employed to produce high-quality images of phase objects such as cells. We describe a method for imaging 3D phase objects in a tomographic configuration implemented by training an artificial neural network to reproduce the complex amplitude of the experimentally measured scattered light. The network is designed such that the voxel values of the refractive index of the 3D object are the variables that are adapted during the training process.
Independent Component Analysis
April 27, 2016 Speaker: Marco Zugliani, uniTN intern at LIONlab Abstract: Independent component analysis (ICA) is a method for separating an observed set of signal mixtures into a set of statistically independent component signals, or source signals. The goal of ICA is to find a linear representation of non-Gaussian data so that the components are statistically independent, or as independent as possible. This technique has different application, e.g. image noise reduction, face recognition, medical signal processing.
Graph Partitioning Problem
April 20, 2016 Speaker: Tahir E. Kalayci, postdoctoral research fellow at LION lab - University of Trento, Italy Abstract: Graphs are frequently used by computer scientists as abstractions when modelling an application problem and cutting a graph into smaller pieces is one of the fundamental algorithmic operations. Even if the final application concerns a different problem (such as traversal, finding paths, trees, and flows), partitioning large graphs is often an important sub-problem for complexity reduction or parallelism. There are different approaches like global optimization, iterative improvement heuristics, multilevel graph partitioning, evolutionary methods and further meta-heuristics for solving graph partitioning problem. In this talk, we are going to focus on bi-partitioning of the graphs and introduce heuristic based algorithms for solving this problem.
scikit-learn, machine learning for the 99.9%
April 13, 2016 Speaker: Fabian Pedregosa, postdoctoral fellow at Chaire Havas-Dauphine / INRIA Paris Abstract: scikit-learn is a machine learning library for Python. Although nowadays largely a community effort, the library was born from the effort of a small number of developers. In this talk I will describe the library vision, goals, and identify key aspects that made it possible to transition from a small number of developers to the current ~100 active contributors. The talk will finish with a brief hands-on introductory tutorial.
Parallel stochastic optimization for machine learning
April 6, 2016 Speaker: Fabian Pedregosa, postdoctoral fellow at Chaire Havas-Dauphine / INRIA Paris Abstract: With its cheap per-iteration cost, stochastic gradient descent has become in recent years the workhorse of huge-scale optimization and is behind the success of recent AI achievements such as deep learning. At the same time, practical parallel variants of stochastic gradient descent have been developed. These algorithms achieve a theoretical linear speedup in the number of cores under strong assumptions, such as sparsity of gradients. After a brief review of the current literature, I will present our main contribution, the ASAGA algorithm, a fully parallel version of the incremental gradient algorithm SAGA that (unlike plain stochastic gradient descent) enjoys fast linear convergence rates. Furthermore, we prove that ASAGA can obtain a theoretical linear speed-up on multicore systems under weaker assumptions that previous literature. We present results of an implementation on a 40-core architecture illustrating the practical speed-up as well as the hardware overhead. This is joint work with my colleagues at INRIA-Paris Rémi Leblond and Simon Lacoste-Julien.
Approximated Nearest Neighbors
March 16, 2016 Speaker: Andrea Mariello, PhD student at LIONlab - University of Trento, Italy Abstract: The problem of finding the nearest neighbor or the first K neighbors of a point occurs very often in various machine learning and pattern recognition applications. For instance, in a classification task we can load all points in a dataset, with their class labels, at first, and then we can assign a new point (the query point) to the class of its nearest neighbor or the class voted by the majority of its first K neighbors. Since this model is lazy, lacking a preprocessing phase, we have to start a new search for each query point. When we increase the number of points in the dataset as well as the number of queries and data dimensionality, finding exact neighbors of all query points by a linear search on the dataset translates to an overall complexity of O(d N^2), where d is the number of dimensions and N the number of points. If our problem is not bound to exact solutions, we can significantly improve the search speed by using Approximated Nearest Neighbors (ANN) techniques. In this seminar, we are going to introduce one of these techniques, which is based on hashing and k-means clustering, that achieves a complexity of O(d N^(3/2)).