Representational power of restricted boltzmann machines and deep belief networks

Deep belief networks (DBN) are generative neural network models with many layers of hidden explanatory factors, recently introduced by Hinton, Osindero, and Teh (2006) along with a greedy layer-wise unsupervised learning algorithm. The building block of a DBN is a probabilistic model called a restricted Boltzmann machine (RBM), used to represent one layer of the model. Restricted Boltzmann machines are interesting because inference is easy in them and because they have been successfully used as building blocks for training deeper models. We first prove that adding hidden units yields strictly improved modeling power, while a second theorem shows that RBMs are universal approximators of discrete distributions. We then study the question of whether DBNs with more layers are strictly more powerful in terms of representational power. This suggests a new and less greedy criterion for training RBMs within DBNs.     

Monotony properties of connected visible graph searching

Search games are attractive for their correspondence with classical width parameters. For instance, the invisible search number (a.k.a. node search number) of a graph is equal to its pathwidth plus 1, and the visible search number of a graph is equal to its treewidth plus 1. The connected variants of these games ask for search strategies that are connected, i.e., at every step of the strategy, the searched part of the graph induces a connected subgraph. We focus on monotone search strategies, i.e., strategies for which every node is searched exactly once. The monotone connected visible search number of an n-node graph is at most O(logn) times its visible search number. First, we prove that this logarithmic bound is tight. Precisely, we prove that there is an infinite family of graphs for which the ratio monotone connected visible search number over visible search number is Ω(logn). Second, we prove that, as opposed to the non-connected variant of visible graph searching, “recontamination helps” for connected visible search. Precisely, we prove that, for any k4, there exists a graph with connected visible search number at most k, and monotone connected visible search number >k     

Mixtures of robust probabilistic principal component analyzers

Mixtures of probabilistic principal component analyzers model high-dimensional nonlinear data by combining local linear models. Each mixture component is specifically designed to extract the local principal orientations in the data. An important issue with this generative model is its sensitivity to data lying off the low-dimensional manifold. In order to address this problem, the mixtures of robust probabilistic principal component analyzers are introduced. They take care of atypical points by means of a long tail distribution, the Student-t. It is shown that the resulting mixture model is an extension of the mixture of Gaussians, suitable for both robust clustering and dimensionality reduction. Finally, we briefly discuss how to construct a robust version of the closely related mixture of factor analyzers.     

A wireless embedded tongue tactile biofeedback system for balance control

We describe the architecture of an original biofeedback system for balance improvement for fall prevention and present results of a feasibility study. The underlying principle of this biofeedback consists of providing supplementary information related to foot sole pressure distribution through a wireless embedded tongue-placed tactile output device.Twelve young healthy adults voluntarily participated in this experiment. They were asked to stand as immobile as possible with their eyes closed in two conditions of no-biofeedback and biofeedback. Centre of foot pressure (CoP) displacements were recorded using a force platform. Results showed reduced CoP displacements in the biofeedback relative to the no-biofeedback condition.On the whole, the present findings evidence the effectiveness of this system in improving postural control on young healthy adults. Further investigations are needed to strengthen the potential clinical value of this device.     

Hierarchical learning from human preferences and curiosity

Applied Intelligence - Recent success in scaling deep reinforcement algorithms (DRL) to complex problems has been driven by well-designed extrinsic rewards, which limits their applicability to many...     

Writing with Fluid: Structuring Hydrogels with Micrometer Precision by AFM in Combination with Nanofluidics

Hydrogels have many applications in biomedical surface modification and tissue engineering. However, the structuring of hydrogels after their formation represents still a major challenge, in particular due to their softness. Here, a novel approach is presented that is based on the combination of atomic force microscopy (AFM) and nanofluidics, also referred to as FluidFM technology. Its applicability is demonstrated for supramolecular hydrogel films that are prepared from low‐molecular weight hydrogelators, such as derivates of 1,3,5‐benzene tricarboxamides (BTAs). BTA films can be dissolved selectively by ejecting alkaline solution through the aperture of a hollow AFM‐cantilever connected to a nanofluidic controller. The AFM‐based force control is essential in preventing mechanical destruction of the hydrogels. The resulting “chemical writing” process is studied in detail and the influence of various parameters, such as applied pressure and time, is validated. It is demonstrated that the achievable structuring precision is primarily limited by diffusion and the aperture dimensions. Recently, various additive techniques have been presented to pattern hydrogels. The here‐presented subtractive approach can not only be applied to structure hydrogels from the large class of reversibly formed gels with superior resolution but would also allow for the selective loading of the hydrogels with active substances or nanoparticles.     

Anticipating Cutoff Diameters in Deterministic Lateral Displacement (DLD) Microfluidic Devices for an Optimized Particle Separation

Deterministic lateral displacement (DLD) devices enable to separate nanometer to micrometer‐sized particles around a cutoff diameter, during their transport through a microfluidic channel with slanted rows of pillars. In order to design appropriate DLD geometries for specific separation sizes, robust models are required to anticipate the value of the cutoff diameter. So far, the proposed models result in a single cutoff diameter for a given DLD geometry. This paper shows that the cutoff diameter actually varies along the DLD channel, especially in narrow pillar arrays. Experimental and numerical results reveal that the variation of the cutoff diameter is induced by boundary effects at the channel side walls, called the wall effect. The wall effect generates unexpected particle trajectories that may compromise the separation efficiency. In order to anticipate the wall effect when designing DLD devices, a predictive model is proposed in this work and has been validated experimentally. In addition to the usual geometrical parameters, a new parameter, the number of pillars in the channel cross dimension, is considered in this model to investigate its influence on the particle trajectories.     

Tunable Diode Laser Absorption Spectroscopy of Carbon Monoxide Around 2.35 microm

Novel GaInSbAs/GaSb multiple-quantum-well lasers operating near room temperature have been successfully used for tunable diode laser absorption spectroscopy in the vicinity of 2.35 mum. Continuous current tuning over a more than 150-GHz frequency range has been realized. Direct absorption measurements have been carried out on the R, R, R, and R lines of carbon monoxide. Traces of carbon monoxide at the level of 0.3 part in 10(6) in volume at 100 Torr could be detected by the low-frequency wavelength-modulation technique and an astigmatic multipass cell. These results show a potential use of these diode lasers in portable low-cost trace-pollutant sensors.     

Indexing and querying segmented web pages: the BlockWeb Model

We present in this paper a model for indexing and querying web pages, based on the hierarchical decomposition of pages into blocks. Splitting up a page into blocks has several advantages in terms of page design, indexing and querying such as (i) blocks of a page most similar to a query may be returned instead of the page as a whole (ii) the importance of a block can be taken into account, as well as (iii) the permeability of the blocks to neighbor blocks: a block b is said to be permeable to a block b?? in the same page if b?? content (text, image, etc.) can be (partially) inherited by b upon indexing. An engine implementing this model is described including: the transformation of web pages into blocks hierarchies, the definition of a dedicated language to express indexing rules and the storage of indexed blocks into an XML repository. The model is assessed on a dataset of electronic news, and a dataset drawn from web pages of the ImagEval campaign where it improves by 16% the mean average precision of the baseline.