Nakul Verma

Learning hierarchical similarity metrics

Categories in multi-class data are often part of an underlying semantic taxonomy. Recent work in object classification has found interesting ways to use this taxonomy structure to develop better recognition algorithms. Here we propose a novel framework to learn similarity metrics using the class taxonomy. We show that a nearest neighbor classifier using the learned metrics gets improved performance over the best discriminative methods. Moreover, by incorporating the taxonomy, our learned metrics can also help in some taxonomy specific applications. We show that the metrics can help determine the correct placement of a new category that was not part of the original taxonomy, and can provide effective classification amongst categories local to specific subtrees of the taxonomy.

CitiSense: improving geospatial environmental assessment of air quality using a wireless personal exposure monitoring system

Environmental exposures are a critical component in the development of chronic conditions such as asthma and cancer. Yet, medical and public health practitioners typically must depend on sparse regional measurements of the environment that provide macro-scale summaries. Recent projects have begun to measure an individuals exposure to these factors, often utilizing body-worn sensors and mobile phones to visualize the data. Such data, collected from many individuals and analyzed across an entire geographic region, holds the potential to revolutionize the practice of public health. We present CitiSense, a participatory air quality sensing system that bridges the gap between personal sensing and regional measurement to provide micro-level detail at a regional scale. In a user study of 16 commuters using CitiSense, measurements were found to vary significantly from those provided by official regional pollution monitoring stations. Moreover, applying geostatistical kriging techniques to our data allows CitiSense to infer a regional map that contains considerably greater detail than official regional summaries. These results suggest that the cumulative impact of many individuals using personal sensing devices may have an important role to play in the future of environmental measurement for public health.

Citisense: mobile air quality sensing for individuals and communities

Individual and community health can be greatly impacted by poor air quality. Unfortunately air quality metrics are hard for individuals to obtain and are often not precise enough for people to make the inferences they need to construct positive personal health choices. Through the Citisense mobile air quality system we enable users to track their personal air quality exposure for discovery, self-reflection, and sharing within their local communities and online social networks.

Efficient energy management and data recovery in sensor networks using latent variables based tensor factorization

A key factor in a successful sensor network deployment is finding a good balance between maximizing the number of measurements taken (to maintain a good sampling rate) and minimizing the overall energy consumption (to extend the network lifetime). In this work, we present a data-driven statistical model to optimize this tradeoff. Our approach takes advantage of the multivariate nature of the data collected by a heterogeneous sensor network to learn spatio-temporal patterns. These patterns enable us to employ an aggressive duty cycling policy on the individual sensor nodes, thereby reducing the overall energy consumption. Our experiments with the OMNeT++ network simulator using realistic wireless channel conditions, on data collected from two real-world sensor networks, show that we can sample just 20% of the data and can reconstruct the remaining 80% of the data with less than 9% mean error, outperforming similar techniques such is distributed compressive sampling. In addition, energy savings ranging up to 76%, depending on the sampling rate and the hardware configuration of the node.

Latent variables based data estimation for sensing applications

Recovering missing sensor data is a critical problem for sensor networks, especially when nodes duty cycle their activity or may experience periodic downtimes due to limited energy. Fortunately, sensor readings are often correlated across different nodes and sensor types. Among state-of-the-art statistical data estimation techniques, latent variable based factor models have emerged as a powerful framework for recovering missing data. In this paper we propose the use of latent variable models to estimate missing data in heterogeneous sensor networks. Our model not only correlates data across different sensor locations and types, but also takes advantage of the temporal structure that is often present in sensor readings. We analyze how this model can effectively reconstruct missing sensor data when the individual sensor nodes have to duty-cycle their activity in order to extend network lifetime. We evaluate our model on a real life sensor network consisting of 122 environmental monitoring stations that periodically collect data from 13 different sensors. Results show that our proposed model can effectively reconstruct over 50% of missing data with less than 10% error.

The cortical dynamics orchestrating skilled prehension

Skillful control of movement is central to our ability to sense and manipulate the world. Dexterous acts depend on cerebral cortex[1-10], and the activity of cortical neurons is correlated with movement[11-15]. By isolating the neural dynamics that command skilled movements from those that reflect other processes (such as planning and deciding to move), we were able to characterize and manipulate the motor commands underlying prehension. We showed that in mice trained to perform a reach / grab / supination / bring-to-mouth sequence (volitional prehension), multiple forms of optogenetic stimuli in sensorimotor cortex resulted in an involuntary, complete movement (opto-prehension). This result suggested that the trained brain could robustly transform a variety of aberrant stimuli into the dynamics sufficient for prehension. We measured the electrical activity of cortical populations and detailed limb kinematics during volitional prehension and optoprehension. During volitional prehension, neurons fired before and during specific stages of the movement, and the population collectively tiled the entire behavioral sequence. During opto-prehension, most neurons recapitulated their volitional prehension activity patterns, but a physiologically distinct subset did not. On trials where the liminal optogenetic stimulus failed to produce these dynamics, movement did not occur, providing further evidence that a specific pattern of neural activity was causally coupled to prehension. Having identified these dynamics, we next tested their robustness to brief, closed-loop perturbation. Regardless of where along the reach we optogenetically halted cortical activity and the movement, relief of suppression resulted in cortical dynamics that immediately recapitulated all steps of the prehension program, and the animal completed the behavior. By combining electrophysiology and optogenetic perturbations, we have identified and characterized the cortical motor program driving a learned, dexterous movement sequence.