Polytechnic University of Turin

About: Polytechnic University of Turin is a education organization based out in Turin, Piemonte, Italy. It is known for research contribution in the topics: Finite element method & Nonlinear system. The organization has 11553 authors who have published 41395 publications receiving 789320 citations. The organization is also known as: POLITO & Politecnico di Torino.

Entropy-SGD: Biasing Gradient Descent Into Wide Valleys

Abstract: This paper proposes a new optimization algorithm called Entropy-SGD for training deep neural networks that is motivated by the local geometry of the energy landscape. Local extrema with low generalization error have a large proportion of almost-zero eigenvalues in the Hessian with very few positive or negative eigenvalues. We leverage upon this observation to construct a local-entropy-based objective function that favors well-generalizable solutions lying in large flat regions of the energy landscape, while avoiding poorly-generalizable solutions located in the sharp valleys. Conceptually, our algorithm resembles two nested loops of SGD where we use Langevin dynamics in the inner loop to compute the gradient of the local entropy before each update of the weights. We show that the new objective has a smoother energy landscape and show improved generalization over SGD using uniform stability, under certain assumptions. Our experiments on convolutional and recurrent networks demonstrate that Entropy-SGD compares favorably to state-of-the-art techniques in terms of generalization error and training time.

Theory of ultrafast phenomena in photoexcited semiconductors

Abstract: The authors review the physics of ultrafast dynamics in semiconductors and their heterostructures, including both the observed experimental phenomena and the theoretical description of the processes. These are probed by ultrafast optical excitation, generating nonequilibrium states that can be monitored by time-resolved spectroscopy. Light pulses create coherent superpositions of states, and the dynamics of the associated phase relationships can be directly investigated by means of many-pulse experiments. The commonly used experimental techniques are briefly reviewed. A variety of different phenomena can be described within a common theoretical framework based on the density-matrix formalism. The important interactions of the carriers included in the theoretical description are the phonon interactions, the interactions with classical and quantum light fields, and the Coulomb interaction among the carriers themselves. These interactions give rise to a strong interplay between phase coherence and relaxation, which strongly affects the non equilibrium dynamics. Based on the general theory, the authors review the physical phenomena in various semiconductor structures including superlattices, quantum wells, quantum wires, and bulk media. Particular results which have played a central role in understanding the microscopic origins of the relaxation processes are discussed in detail.

Centrality dependence of pi, K, and p production in Pb-Pb collisions at root s(NN)=2.76 TeV

Abstract: In this paper measurements are presented of pi(+/-), K-+/-, p, and (p) over bar production at midrapidity (vertical bar y vertical bar < 0.5), in Pb-Pb collisions at root s(NN) = 2.76 TeV as a function of centrality. The measurement covers the transverse-momentum (p(T)) range from 100, 200, and 300 MeV/c up to 3, 3, and 4.6 GeV/c for pi, K, and p, respectively. The measured p(T) distributions and yields are compared to expectations based on hydrodynamic, thermal and recombination models. The spectral shapes of central collisions show a stronger radial flow than measured at lower energies, which can be described in hydrodynamic models. In peripheral collisions, the p(T) distributions are not well reproduced by hydrodynamic models. Ratios of integrated particle yields are found to be nearly independent of centrality. The yield of protons normalized to pions is a factor similar to 1.5 lower than the expectation from thermal models.

Non-linear kinetics underlying generalized statistics

Abstract: The purpose of the present effort is threefold. Firstly, it is shown that there exists a principle, that we call kinetical interaction principle (KIP), underlying the non-linear kinetics in particle systems, independently on the picture (Kramers, Boltzmann) used to describe their time evolution. Secondly, the KIP imposes the form of the generalized entropy associated to the system and permits to obtain the particle statistical distribution, both as stationary solution of the non-linear evolution equation and as the state which maximizes the generalized entropy. Thirdly, the KIP allows, on one hand, to treat all the classical or quantum statistical distributions already known in the literature in a unifying scheme and, on the other hand, suggests how we can introduce naturally new distributions. Finally, as a working example of the approach to the non-linear kinetics here presented, a new non-extensive statistics is constructed and studied starting from a one-parameter deformation of the exponential function holding the relation f(−x)f(x)=1.

Comparisons among clustering techniques for electricity customer classification

Abstract: The recent evolution of the electricity business regulation has given new possibilities to the electricity providers for formulating dedicated tariff offers. A key aspect for building specific tariff structures is the identification of the consumption patterns of the customers, in order to form specific customer classes containing customers exhibiting similar patterns. This paper illustrates and compares the results obtained by using various unsupervised clustering algorithms (modified follow-the-leader, hierarchical clustering, K-means, fuzzy K-means) and the self-organizing maps to group together customers with similar electrical behavior. Furthermore, this paper discusses and compares various techniques-Sammon map, principal component analysis (PCA), and curvilinear component analysis (CCA)-able to reduce the size of the clustering input data set, in order to allow for storing a relatively small amount of data in the database of the distribution service provider for customer classification purposes. The effectiveness of the classifications obtained with the algorithms tested is compared in terms of a set of clustering validity indicators. Results obtained on a set of nonresidential customers are presented.