Search

How to build machine learning models from few annotations is an open research question. This article shows an application of a meta-learning algorithm (REPTILE) to solve the problem of object segmentation. We evaluate how using REPTILE during a pre-training phase accelerates the learning process without loosing performance of the resulting segmentation in poor labeling conditions, and compare these results against training the detectors using basic transfer learning. Two scenarios are tested: (i) how segmentation performance evolves through training epochs with a fixed amount of labels and (ii) how segmentation performance improves with an increasing amount of labels after a fixed amount of epochs. The results suggest that REPTILE is useful making learning faster in both cases.

This paper introduces a method for the generation of images that activate any target neuron or group of neurons of a trained convolutional neural network (CNN). These images are created in such a way that they contain attributes of natural images such as color patterns or textures. The main idea of the method is to pre-train a deep generative network on a dataset of natural images and then use this network to generate images for the target CNN. The analysis of the generated images allows for a better understanding of the CNN internal representations, the detection of otherwise unseen biases, or the creation of explanations through feature localization and description.

The current availability of the humanoid robots opens up a wide range of applications, for instance, in the domain of hospitality the humanoids can be programmed to behave autonomous ways to provide help to people. The aspect of the humanoids and the humanness of interaction are key components of success. We developed a system to endow the humanoid robot Pepper, from SoftBank Robotics, with the capability of both: identify the emotion state of the humans and exhibiting emotional states via gestures and postures generated using a DCGAN model to learn from the human body language and to create originals body expressions to exhibit like-human movements.

Automatic recognition of user context is essential for a variety of emerging applications, such as context-dependent content delivery, telemonitoring of medical patients, or quantified life-logging. Although not explicitly observable as, e.g., activities, an important aspect towards understanding user context lies in the affective state of mood.While significant work has been done to assess mood, most approaches require the use of customized sensors and controlled laboratory settings.In this work, we engineer a recognition pipeline to recognize daily activities from commercially popularized wearable electronics. In turn, we use predicted activities to learn a regression model capable of assessing user mood.Using only commercially popularized wearable devices, we enable the potential for seamless deployment to the general public.Conducting a real-world study with a prototype system, we collect and evaluate data from 18 users, who provide over 93 user-days of labelled activity data.Regressing for mood based on predicted daily activities, we are able to infer mood angles with a mean absolute error of $0.24\pi$ radians on the Circumplex Model of Affect.Comparing with benchmark approaches, our approach outperforms with statistical significance and is validated for robustness against noise from activity misclassification.

Mountain regions worldwide are particularly sensitive to on-going climate change. Specifically in the Alps in Switzerland, the temperature has increased twice as fast than in the rest of the Northern hemisphere. Water temperature closely follows the annual air temperature cycle, severely impacting streams and freshwater ecosystems. In the last 20 years, brown trout ( Salmo trutta L . ) catch has declined by approximately 40–50% in many rivers in Switzerland. Increasing water temperature has been suggested as one of the most likely cause of this decline. Temperature has a direct effect on trout population dynamics through developmental and disease control but can also indirectly impact dynamics via food-web interactions such as resource availability. We developed a spatially explicit modelling framework that allows spatial and temporal projections of trout biomass using the Aare river catchment as a model system, in order to assess the spatial and seasonal patterns of trout biomass variation. Given that biomass has a seasonal variation depending on trout life history stage, we developed seasonal biomass variation models for three periods of the year (Autumn–Winter, Spring and Summer). Because stream water temperature is a critical parameter for brown trout development, we first calibrated a model to predict water temperature as a function of air temperature to be able to further apply climate change scenarios. We then built a model of trout biomass variation by linking water temperature to trout biomass measurements collected by electro-fishing in 21 stations from 2009 to 2011. The different modelling components of our framework had overall a good predictive ability and we could show a seasonal effect of water temperature affecting trout biomass variation. Our statistical framework uses a minimum set of input variables that make it easily transferable to other study areas or fish species but could be improved by including effects of the biotic environment and the evolution of demographical parameters over time. However, our framework still remains informative to spatially highlight where potential changes of water temperature could affect trout biomass.

Device localization is a common issue on connected objects and persons. Relative localization is a particular case where one don't care about an absolute position, but about knowing how far and in which direction an object is from a reference point in an approximate manner. This work makes part of a larger project which aims to track groups of children for preventing hazards. In this paper we present a relative localization system in the form of a distributed multi-hop routing algorithm deployed on a set of wearable Bluetooth Low Energy devices. This paper introduces the distributive algorithm, describes its implementation, and presents performance results which permit to validate the approach and identify the system limitations.

Activity recognition is gaining a lot of interest given its direct use in applications like ambient assisted living and has been empowered by the increasing ubiquity of sensors (e.g., clothes, smartphones, watches). The machine learning approach to activity recognition consists on finding the signatures characterizing the activities to be recognized, with the hope of identifying them (pattern matching) within the stream of sensor data. The finding of those signatures can be very complex, thus many approaches deal with the streams of sensor data by segmenting them into sections or “time-windows”, before processing them by a feature extraction procedure. The problem then concerns the association of features to class labels. In this paper, we propose the use of the Gamma Growing Neural Gas algorithm to unsupervisely discover templates in a recording containing gestures performed by a person in a home environment. The system is able to do vector quantization from the time-series of data coming from one accelerometer, and finds salient patterns (e.g., templates) in the signal. These templates integrate information not only from single time-windows but do consider the recent history of the incoming signal (e.g., multiple time-windows). Those templates are then associated to activity classes by supervised learning. Our experiments show that the resulting performance is better than previous benchmarks of the same database.

We present the co-design of a gaming scenario between an Artificial Evolution algorithm and a human designer. Such co-design is twofold, consisting of an initial stage in which a genetic algorithm is used to evolve the control parameters that define the behavior of a group of virtual agents. This produces interesting and unexpected results not only creating differentiated behaviors but also increasing the flexibility of the character to adapt to a given objective. In the second stage of the game design a human interplay was introduced in adding other elements to the game, such as other characters and new game dynamics. In this paper we introduce a game that integrates virtual and physical characters while taking advantage of such co-design approach. The physical character consists of a robot which controlled through a Natural User Interface can be part of the game by interacting with other characters in the virtual environment.

ActiDote —activity as an antidote— is a system for manual wheelchair users that uses wireless sensors to recognize activities of various intensity levels in order to allow self-tracking while providing motivation. In this paper, we describe both the hardware setup and the software pipeline that enable our system to operate. Laboratory tests using multi-modal fusion and machine learning reveal promising results attaining a F1-score classification performance of 0.97 on five different wheelchair-based activities belonging to four intensity levels. Finally, we show that such a low cost system can be used for an easy self-monitoring of physical activity levels among manual wheelchair users.

Every time a farmer plants and harvests a crop represents a unique event or experiment. Our premise is that if it were possible to characterize the production system in terms of management and the environmental conditions, and if information on the harvested product were collected from a large number of harvesting events under varied conditions, it should be possible to develop data-driven models that describe the production system. These models can then be used to identify appropriate growing conditions and improved management practices for crops that have received little attention from researchers. The analysis and interpretation of commercial production data in the context of naturally occurring variation in environmental and management, as opposed to controlled experimental data, requires novel approaches. Information was available on both variation in commercial production of the tropical fruit, lulo (Solanum quitoense), and the associated environmental conditions in Colombia. This information was used to develop and evaluate procedures for the interpretation of the variation in commercial production of lulo. The most effective procedures depended on expert guidance: it was not possible to develop a simple effective one step procedure, but rather an iterative approach was required. The most effective procedure was based on the following steps. First, highly correlated independent variables were evaluated and those that were effectively duplicates were eliminated. Second, regression models identified those environmental factors most closely associated with the dependent variable of fruit yield. The environmental factors associated with variation in fruit yield were then used for more in depth analysis, and those environmental variables not associated with yield were excluded from further analysis. Linear regression and multilayer perceptron regression models explained 65–70% of the total variation in yield. Both models identified three of the same factors but the multilayer perceptron based on a neural network identified one location as an additional factor. Third, the three environmental factors common to both regression models were used to define three Homogeneous Environmental Conditions (HECs) using Self-Organizing Maps (SOM). Fourth, yield was analyzed with a mixed model with the categorical variables of HEC, location, as a proxy for cultural factors associated with a geographic region, and farm as proxy for management skills. The mixed model explained more than 80% of the total variation in yield with 61% associated with the HECs and 19% with farm. Location had minimal effects. The results of this model can be used to determine the appropriate environmental conditions for obtaining high yields for crops where only commercial data are available, and also to identify those farms that have superior management practices for given environmental conditions.

Climate change has created the need for new strategies in conservation planning that account for the dynamics of factors threatening endangered species. Here we assessed climate change threat to the European otter, a flagship species for freshwater ecosystems, considering how current conservation areas will perform in preserving the species in a climatically changed future. We used an ensemble forecasting approach considering six modelling techniques applied to eleven subsets of otter occurrences across Europe. We performed a pseudo-independent and an internal evaluation of predictions. Future projections of species distribution were made considering the A2 and B2 scenarios for 2080 across three climate models: CCCMA-CGCM2, CSIRO-MK2 and HCCPR HADCM3. The current and the predicted otter distributions were used to identify priority areas for the conservation of the species, and overlapped to existing network of protected areas. Our projections show that climate change may profoundly reshuffle the otter’s potential distribution in Europe, with important differences between the two scenarios we considered. Overall, the priority areas for conservation of the otter in Europe appear to be unevenly covered by the existing network of protected areas, with the current conservation efforts being insufficient in most cases. For a better conservation, the existing protected areas should be integrated within a more general conservation and management strategy incorporating climate change projections. Due to the important role that the otter plays for freshwater habitats, our study further highlights the potential sensitivity of freshwater habitats in Europe to climate change.

The Andean blackberry (Rubus glaucus) is an important source of income in hillside regions of Colombia. However, growers have little reliable information on the factors that affect the development and yield of the crop, and therefore there is a dearth of information on how to effectively manage the crop. Site specific information recorded by small-scale producers of the Andean blackberry on their production systems and soils coupled with publicly available meteorological data was used to develop models of such production systems. Multilayer perceptrons and Self-Organizing Maps were used as computational models in the identification and visualization of the most important variables for modeling the production of Andean blackberry. Artificial neural networks were trained with information from 20 sites in Colombia where the Andean blackberry is cultivated. Multilayer perceptrons predicted with a reasonable degree of accuracy the production response of the crop. The soil depth, the average temperature, external drainage, and the accumulated precipitation of the first month before harvest were critical determinants of productivity. A proxy variable of location was used to describe overall differences in management between farmers groups. The use of this proxy indicated that, even under essentially similar environmental conditions, large differences in production could be assigned to management effects. The information obtained can be used to determine sites that are suitable for Andean blackberry production, and to transfer of management practices from sites of high productivity to sites with similar environmental conditions which currently have lower levels of productivity.

Activity recognition has recently gained a lot of interest and there already exist several methods to detect human activites based on wearable sensors. Most of the existing methods rely on a database of labelled activities that is used to train an offline activity recognition system. This paper presents an approach to build an online activity recognition system that do not require any a priori labelled data. The system incrementally learns activities by actively querying the user for labels. To choose when the user should be queried, we compare a method based on random sampling and another that uses a Growing Neural Gas (GNG). The use of GNG helps reducing the number of user queries by 20% to 30%.

Small farmers in developing countries face the problem of deciding where to cultivate and how to manage their crops. In under researched crops, they base many of their decisions on traditional knowledge and personal experience. We surmised that their decision making processes could be enriched by inductive or data-driven models which should provide a means to improve crop management practices. Bio-inspired machine learning techniques like artificial neural networks are promising modelling tools for accomplishing the aforementioned task due to their proven capabilities when dealing with noisy, incomplete, and heterogeneous data. Moreover, bio-inspired techniques appear to perform quite well without strong assumptions on the data. Last but not least, they provide innovative ways to process and visualize highly-dimensional information. In this chapter, we illustrate the benefits of this methodology by presenting two case studies on fruit crops in Colombia. The studies reported here are associated with two related but separate problems: First the association of crop productivity with growing conditions and management and; Secondly the identification of similar or analogue sites between which technology can readily be transferred.

This paper presents the final hardware platform developed in the Perplexus project. This platform is composed of a reconfigurable device called the ubichip, which is embedded on a pervasive platform called the ubidule, and can also be integrated on the marXbot robotic platform. The whole platform is intended to provide a hardware platform for the simulation of complex systems, and some examples of them are presented at the end of the paper.

Vector quantization of large datasets can be carried out by means of an incremental modelling approach where the modelling task is transformed into an incremental task by partitioning or sampling the data, and the resulting datasets are processed by means of an incremental learner. Growing Neural Gas is an incremental vector quantization algorithm with the capabilities of topology-preserving and distribution-matching. Distribution matching can produce overpopulation of prototypes in zones with high density of data. In order to tackle this drawback, we introduce some modifications to the original Growing Neural Gas algorithm by adding three new parameters, one of them controlling the distribution of the codebook and the other two controlling the quantization error and the amount of units in the network. The resulting learning algorithm is capable of efficiently quantizing large datasets presenting high and low density regions while solving the prototype proliferation problem.

Growing Neural Gas is an incremental vector quantization algorithm with the capabilities of topology-preserving and distribution-matching. Distribution matching can produce overpopulation of prototypes in zones with high density of data. In order to tackle this drawback, we introduce some modifications to the original Growing Neural Gas algorithm by adding three new parameters, one of them controlling the distribution of the codebook and the other two controlling the quantization error and the amount of units in the network. The resulting learning algorithm is capable of efficiently quantizing datasets presenting high and low density regions while solving the prototype proliferation problem.

Data incompleteness and data scarcity are common problems in agroecological modelling. Moreover, agroecological processes depend on historical data that could be fed into a model in a vast number of ways. This work shows a case study of modelling in agroecology using artificial neural networks. The variable to be modelled is sugar cane yield and for this purpose we used climate, soil, and other environmental variables. Regarding the data precision issue, we trained different neural models using monthly and weekly data in order to compare their performance. Furthermore, we studied the influence of using incomplete observations in the training process in order to include them and thus use a larger quantity of input patterns. Our results show that the gain in observations due to the inclusion of incomplete data is preferable in this application.

Upper limb impairment is one of the most common problems for people with neurological disabilities, affecting their activity, quality of life (QOL), and independence. Objective assessment of upper limb performance is a promising way to help patients with neurological upper limb disorders. By using wearable sensors, such as an egocentric camera, it is possible to monitor and objectively assess patients’ actual performance in activities of daily life (ADLs). We analyzed the possibility of using Deep Learning models for depth estimation based on a single RGB image to allow the monitoring of patients with 2D (RGB) cameras. We conducted experiments placing objects at different distances from the camera and varying the lighting conditions to evaluate the performance of the depth estimation provided by two deep learning models (MiDaS & Alhashim). Finally, we integrated the best performing model for depth-estimation (MiDaS) with other Deep Learning models for hand (MediaPipe) and object detection (YOLO) and evaluated the system in a task of hand-object interaction. Our tests showed that our final system has a 78% performance in detecting interactions, while the reference performance using a 3D (depth) camera is 84%.

Foraging is a common benchmark problem in collective robotics in which a robot (the forager) explores a given environment while collecting items for further deposition at specific locations. A typical real-world application of foraging is garbage collection where robots collect garbage for further disposal in pre-defined locations. This work proposes a method to cooperatively perform the task of finding such locations: instead of using local or global localization strategies relying on pre-installed infrastructure, the proposed approach takes advantage of the knowledge gathered by a population about the localization of the targets. In our approach, robots communicate in an intrinsic way the estimation about how near they are from a target; these estimations are used by neighbour robots for estimating their proximity, and for guiding the navigation of the whole population when looking for these specific areas. We performed several tests in a simulator, and we validated our approach on a population of real robots. For the validation tests we used a mobile robot called marXbot. In both cases (i.e., simulation and implementation on real robots), we found that the proposed approach efficiently guides the robots towards the pre-specified targets while allowing the modulation of their speed.