Multivariate Density Estimation by Neural Networks.

We propose nonparametric methods to obtain the Probability Density Function (PDF) to assess the properties of the underlying data generating process (DGP) without imposing any assumptions on the DGP, using neural networks (NNs). The proposed NN has advantages compared to well-known parametric and nonparametric density estimators. Our approach builds on literature on cumulative distribution function (CDF) estimation using NN. We extend this literature by providing analytical derivatives of this obtained CDF. Our approach hence removes the numerical approximation error in differentiating the CDF output, leading to more accurate PDF estimates. The proposed solution applies to any NN model, i.e., for any number of hidden layers or hidden neurons in the multilayer perceptron (MLP) structure. The proposed solution applies the PDF estimation by NN to continuous distributions as well as discrete distributions. We also show that the proposed solution to obtain the PDF leads to good approximations when applied to correlated variables in a multivariate setting. We test the performance of our method in a large Monte Carlo simulation using various complex distributions. Subsequently, we apply our method to estimate the density of the number of vehicle counts per minute measured with road sensors for a time window of 24 h.

Using the Personal Health Train for Automated and Privacy-Preserving Analytics on Vertically Partitioned Data.

Conventional data mining algorithms are unable to satisfy the current requirements on analyzing big data in some fields such as medicine, policy making, judicial, and tax records. However, applying diverse datasets from different institutes (both healthcare and non-healthcare related) can enrich information and insights. So far, analyzing this data in an automated, privacy-preserving manner does not exist to our knowledge. In this work, we propose an infrastructure, and proof-of-concept for privacy-preserving analytics on vertically partitioned data.

Audiophile hardware in vision science; the soundcard as a digital to analog converter.

The design objective was to develop an inexpensive digital to analog (D/A) converter for use in vision science. Soundcards are hardware units that can be integral or can be added to a computer to add sound capability. A soundcard contains D/A converters designed to work in the audio frequency range, typically 20-20,000 Hz. Soundcard outputs are high-pass filtered and thus do not convey sub-audio frequency or dc information. It is possible to circumvent this design feature by programming the desired output waveform as an amplitude modulation of a high frequency carrier, and then demodulating the soundcard output. The circuit, using a 20 kHz carrier, provides precise D/A conversion for the frequency range relevant for vision experiments, dc to 100 Hz, using inexpensive readily available components. The specific application was for 8 channels of D/A conversion using a Macintosh computer running under OS X. The software needed to program stimuli was created using CoreAudio, a library for programming sounds in OS X. Using soundcards on other platforms would not be a problem, as long as there exists a low level library that would enable the wave table to be filled.

Inferred retinal mechanisms mediating illusory distortions.

The Zoellner illusion is a geometric distortion occurring when nonorthogonal inducing lines appear to tilt veridically parallel bars. The retinal pathways contributing to such illusions are unknown. The goal of this experiment was to investigate the retinal origin of the illusion. This was accomplished by determining the contrast gain for illusion thresholds. The magnocellular (MC-) and parvocellular (PC-) pathways exhibit different contrast gains, and this difference can be used psychophysically to identify the pathway. The stimulus pattern was four vertical bars with a series of inducing lines. The bars were always 5% higher in contrast than the inducing bars. The pattern was presented on a larger pedestal. Two paradigms were used. In the pulsed-pedestal paradigm, the observer adapted to the background and the pedestal and pattern were presented together as a brief pulse. In the steady-pedestal paradigm, the observer adapted to the continuously presented pedestal and the pattern appeared as a brief pulse. The contrast between the pedestal and the pattern was varied to obtain thresholds for two criteria: perceiving the directions of the inner inducing lines, and perceiving the distortion of the bars. The results for both criteria were similar in shape, but displaced in sensitivity. Detection of the directions of the inner inducing lines was 0.16-0.29 log unit more sensitive than perception of the illusion. The data for the pulsed-pedestal paradigm depended on the contrast between the pedestal and the pattern and produced a shallow V-shape. These results were associated with mediation in the PC-pathway. The data for the steady-pedestal paradigm depended on the pedestal luminance in a linear relation and showed similar sensitivity to the data for the pulsed-pedestal paradigm. Perception of the illusion required 10-15% Weber contrast.