Bjørn Sand Jensen

Estimating human predictability from mobile sensor data

Quantification of human behavior is of prime interest in many applications ranging from behavioral science to practical applications like GSM resource planning and context-aware services. As proxies for humans, we apply multiple mobile phone sensors all conveying information about human behavior. Using a recent, information theoretic approach it is demonstrated that the trajectories of individual sensors are highly predictable given complete knowledge of the infinite past. We suggest using a new approach to time scale selection which demonstrates that participants have even higher predictability of non-trivial behavior on smaller timer scale than previously considered.

Learning Combinations of Multiple Feature Representations for Music Emotion Prediction

Music consists of several structures and patterns evolving through time which greatly influences the human decoding of higher-level cognitive aspects of music like the emotions expressed in music. For tasks, such as genre, tag and emotion recognition, these structures have often been identified and used as individual and non-temporal features and representations. In this work, we address the hypothesis whether using multiple temporal and non-temporal representations of different features is beneficial for modeling music structure with the aim to predict the emotions expressed in music. We test this hypothesis by representing temporal and non-temporal structures using generative models of multiple audio features. The representations are used in a discriminative setting via the Product Probability Kernel and the Gaussian Process model enabling Multiple Kernel Learning, finding optimized combinations of both features and temporal/ non-temporal representations. We show the increased predictive performance using the combination of different features and representations along with the great interpretive prospects of this approach.

A predictive model of music preference using pairwise comparisons

Music recommendation is an important aspect of many streaming services and multi-media systems, however, it is typically based on so-called collaborative filtering methods. In this paper we consider the recommendation task from a personal viewpoint and examine to which degree music preference can be elicited and predicted using simple and robust queries such as pairwise comparisons. We propose to model - and in turn predict - the pairwise music preference using a very flexible model based on Gaussian Process priors for which we describe the required inference. We further propose a specific covariance function and evaluate the predictive performance on a novel dataset. In a recommendation style setting we obtain a leave-one-out accuracy of 74% compared to 50% with random predictions, showing potential for further refinement and evaluation.

Efficient preference learning with pairwise continuous observations and Gaussian Processes

Human preferences can effectively be elicited using pairwise comparisons and in this paper current state-of-the-art based on binary decisions is extended by a new paradigm which allows subjects to convey their degree of preference as a continuous but bounded response. For this purpose, a novel Beta-type likelihood is proposed and applied in a Bayesian regression framework using Gaussian Process priors. Posterior estimation and inference is performed using a Laplace approximation. The potential of the paradigm is demonstrated and discussed in terms of learning rates and robustness by evaluating the predictive performance under various noise conditions on a synthetic dataset. It is demonstrated that the learning rate of the novel paradigm is not only faster under ideal conditions, where continuous responses are naturally more informative than binary decisions, but also under adverse conditions where it seemingly preserves the robustness of the binary paradigm, suggesting that the new paradigm is robust to human inconsistency.

A service based estimation method for MPSoC performance modelling

This paper presents an abstract service based estimation method for MPSoC performance modelling which allows fast, cycle accurate design space exploration of complex architectures including multi processor configurations at a very early stage in the design phase. The modelling method uses a service oriented model of computation based on Hierarchical Colored Petri Nets and allows the modelling of both software and hardware in one unified model. To illustrate the potential of the method, a small MPSoC system, developed at Bang & Olufsen ICEpower a/s, is modelled and performance estimates are produced for various configurations of the system in order to explore the best possible implementation.

Bounded Gaussian process regression

We extend the Gaussian process (GP) framework for bounded regression by introducing two bounded likelihood functions that model the noise on the dependent variable explicitly. This is fundamentally different from the implicit noise assumption in the previously suggested warped GP framework. We approximate the intractable posterior distributions by the Laplace approximation and expectation propagation and show the properties of the models on an artificial example. We finally consider two real-world data sets originating from perceptual rating experiments which indicate a significant gain obtained with the proposed explicit noise-model extension.

Pseudo inputs for pairwise learning with Gaussian processes

We consider learning and prediction of pairwise comparisons between instances. The problem is motivated from a perceptual view point, where pairwise comparisons serve as an effective and extensively used paradigm. A state-of-the-art method for modeling pairwise data in high dimensional domains is based on a classical pairwise probit likelihood imposed with a Gaussian process prior. While extremely flexible, this non-parametric method struggles with an inconvenient O(n3) scaling in terms of the n input instances which limits the method only to smaller problems. To overcome this, we derive a specific sparse extension of the classical pairwise likelihood using the pseudo-input formulation. The behavior of the proposed extension is demonstrated on a toy example and on two real-world data sets which outlines the potential gain and pitfalls of the approach. Finally, we discuss the relation to other similar approximations that have been applied in standard Gaussian process regression and classification problems such as FI(T)C and PI(T)C.

Haussdorff and hellinger for colorimetric sensor array classification

Development of sensors and systems for detection of chemical compounds is an important challenge with applications in areas such as anti-terrorism, demining, and environmental monitoring. A newly developed colorimetric sensor array is able to detect explosives and volatile organic compounds; however, each sensor reading consists of hundreds of pixel values, and methods for combining these readings from multiple sensors must be developed to make a classification system. In this work we examine two distance based classification methods, K-Nearest Neighbor (KNN) and Gaussian process (GP) classification, which both rely on a suitable distance metric. We evaluate a range of different distance measures and propose a method for sensor fusion in the GP classifier. Our results indicate that the best choice of distance measure depends on the sensor and the chemical of interest.