Zero-Shot Learning for IMU-Based Activity Recognition Using Video Embeddings

The Activity Recognition Chain generally precludes the challenging scenario of recognizing new activities that were unseen during training, despite this scenario being a practical and common one as users perform diverse activities at test time. A few prior works have adopted zero-shot learning methods for IMU-based activity recognition, which work by relating seen and unseen classes through an auxiliary semantic space. However, these methods usually rely heavily on a hand-crafted attribute space which is costly to define, or a learnt semantic space based on word embedding, which lacks motion-related information crucial for distinguishing IMU features. Instead, we propose a strategy to exploit videos of human activities to construct an informative semantic space. With our approach, knowledge from state-of-the-art video action recognition models is encoded into video embeddings to relate seen and unseen activity classes. Experiments on three public datasets find that our approach outperforms other learnt semantic spaces, with an additional desirable feature of scalability, as recognition performance is seen to scale with the amount of data used. More generally, our results indicate that exploiting information from the video domain for IMU-based tasks is a promising direction, with tangible returns in a zero-shot learning scenario.

An agent based force vector model of social influence that predicts strong polarization in a connected world

The model is based on a vector representation of each agent. The components of the vector are the key continuous “attributes” that determine the social behavior of the agent. A simple mathematical force vector model is used to predict the effect of each agent on all other agents. The force law used is motivated by gravitational force laws and electrical force laws for dipoles. It assumes that the force between two agents is proportional to the “similarity of attributes”, which is implemented mathematically as the dot product of the vectors representing the attributes of the agents, and the force goes as the inverse square of the difference in attributes, which is expressed as the Euclidean distance in attribute space between the two vectors. The force between the agents may be positive (attractive), zero, or negative (repulsive) depending on whether the angle between the corresponding vectors is less than, equal to, or greater than 90°. A positive force causes the attributes of the agents to become more similar and the corresponding vectors to become more nearly parallel. Interaction between all agents is allowed unless the distance between the attributes representing the agents exceeds a confidence limit (the Attribute Influence Bound) set in the simulation. Agents with similar attributes tend to form groups. For small values of the Attribute Influence Bound, numerous groups remain scattered throughout attribute space at the end of a simulation. As the Attribute Influence Bound is increased, and agents with increasingly different attributes can communicate, fewer groups remain at the end, and the remaining groups have increasingly different characteristic attributes and approximately equal sizes. With a large Attribute Influence Bound all agents are connected and extreme bi- or tri-polarization results. During the simulations, depending on the initial conditions, collective behaviors of grouping, consensus, fragmentation and polarization are observed as well as certain symmetries specific to the model, for example, the average of the attributes for all agents does not change significantly during a simulation.

Learning-Based Dissimilarity for Clustering Categorical Data

Comparing data objects is at the heart of machine learning. For continuous data, object dissimilarity is usually taken to be object distance; however, for categorical data, there is no universal agreement, for categories can be ordered in several different ways. Most existing category dissimilarity measures characterize the distance among the values an attribute may take using precisely the number of different values the attribute takes (the attribute space) and the frequency at which they occur. These kinds of measures overlook attribute interdependence, which may provide valuable information when capturing per-attribute object dissimilarity. In this paper, we introduce a novel object dissimilarity measure that we call Learning-Based Dissimilarity, for comparing categorical data. Our measure characterizes the distance between two categorical values of a given attribute in terms of how likely it is that such values are confused or not when all the dataset objects with the remaining attributes are used to predict them. To that end, we provide an algorithm that, given a target attribute, first learns a classification model in order to compute a confusion matrix for the attribute. Then, our method transforms the confusion matrix into a per-attribute dissimilarity measure. We have successfully tested our measure against 55 datasets gathered from the University of California, Irvine (UCI) Machine Learning Repository. Our results show that it surpasses, in terms of various performance indicators for data clustering, the most prominent distance relations put forward in the literature.

Attribute-Guided Network Sampling Mechanisms

This article introduces a novel task-independent sampler for attributed networks. The problem is important because while data mining tasks on network content are common, sampling on internet-scale networks is costly. Link-trace samplers such as Snowball sampling, Forest Fire, Random Walk, and Metropolis–Hastings Random Walk are widely used for sampling from networks. The design of these attribute-agnostic samplers focuses on preserving salient properties of network structure, and are not optimized for tasks on node content. This article has three contributions. First, we propose a task-independent, attribute aware link-trace sampler grounded in Information Theory. Our sampler greedily adds to the sample the node with the most informative (i.e., surprising) neighborhood. The sampler tends to rapidly explore the attribute space, maximally reducing the surprise of unseen nodes. Second, we prove that content sampling is an NP-hard problem. A well-known algorithm best approximates the optimization solution within 1 − 1/ e , but requires full access to the entire graph. Third, we show through empirical counterfactual analysis that in many real-world datasets, network structure does not hinder the performance of surprise based link-trace samplers. Experimental results over 18 real-world datasets reveal: surprise-based samplers are sample efficient and outperform the state-of-the-art attribute-agnostic samplers by a wide margin (e.g., 45% performance improvement in clustering tasks).

Falling through the cracks: Modeling the formation of social category boundaries

Social categorizations divide people into “us” and “them”, often along continuous attributes such as political ideology or skin color. This division results in both positive consequences, such as a sense of community, and negative ones, such as group conflict. Further, individuals in the middle of the spectrum can fall through the cracks of this categorization process and are seen as out-group by individuals on either side of the spectrum, becoming inbetweeners. Here, we propose a quantitative, dynamical-system model that studies the joint influence of cognitive and social processes. We model where two social groups draw the boundaries between “us” and ‘them” on a continuous attribute. Our model predicts that both groups tend to draw a more restrictive boundary than the middle of the spectrum. As a result, each group sees the individuals in the middle of the attribute space as an out-group. We test this prediction using U.S. political survey data on how political independents are perceived by registered party members as well as existing experiments on the perception of racially ambiguous faces, and find support.

Feature Level-Sets: Generalizing Iso-surfaces to Multi-variate Data

Iso-surfaces or level-sets provide an effective and frequently used means for feature visualization. However, they are restricted to simple features for uni-variate data. The approach does not scale when moving to multi-variate data or when considering more complex feature definitions. In this paper, we introduce the concept of traits and feature level-sets, which can be understood as a generalization of level-sets as it includes iso-surfaces, and fiber surfaces as special cases. The concept is applicable to a large class of traits defined as subsets in attribute space, which can be arbitrary combinations of points, lines, surfaces and volumes. It is implemented into a system that provides an interface to define traits in an interactive way and multiple rendering options. We demonstrate the effectiveness of the approach using multi-variate data sets of different nature, including vector and tensor data, from different application domains.

Setting up the Attribute Space of Periodic Time Dependencies for Dynamic Object Diagnostic Systems (Taking the Eye Retina as an Example)

The technology of testing dynamically and structurally similar aircraft models for flutter in subsonic wind tunnels using information and The article addresses techniques for setting up the attribute space of informative features of periodic signals recorded at the output of a dynamic object with an unknown structure in response to rectangular testing signals of different frequencies applied to the object input. The attribute space is used in developing expert systems for diagnosing the current state of an operating dynamic object. With a great variety of possible developing faults, the simplest practical techniques involving the use of characteristic points of change in the observed time dependencies yield a limited number of features with large mutual intersection domains. To expand the attribute space, it is proposed to use the expansion of input and output signals into a Fourier series for setting up a base of additional features. The proposed features characterize, depending on the testing conditions, the object’s transferring properties in the frequency domain from changes in its amplitude and phase characteristics. The test pulse frequency and duration serve as such conditions. For the convenience of comparing the object’s frequency responses variation pattern, two special procedures are used. The first procedure allows the observed time dependencies to be reduced to a single pseudo frequency of the test signals. The second procedure uses specially formed windows for subjecting individual fragments of the observed time dependencies to a spectral analysis. It is shown that, depending on the type of the frequency responses being analyzed, the techniques for their polynomial approximation, as well as integral estimates of frequency response individual domains can be useful. The polynomial approximation makes it possible to use the coefficients of the approximating polynomials as additional features, and the integration of individual characteristic domains of the frequency responses makes it possible to introduce dimensionless relative indicators that characterize the degree of change in the frequency responses depending on the experimental conditions. The considered techniques open the possibility to select additional features that can help distinguish both separate groups of faults and individual faults in operating objects. The study results are illustrated by the examples of analyzing the changes in electroretinograms that record changes in the eye retina biopotential in response to light flashes of different frequencies.