Virginia Smith

MLI: An API for Distributed Machine Learning

MLI is an Application Programming Interface designed to address the challenges of building Machine Learning algorithms in a distributed setting based on data-centric computing. Its primary goal is to simplify the development of high-performance, scalable, distributed algorithms. Our initial results show that, relative to existing systems, this interface can be used to build distributed implementations of a wide variety of common Machine Learning algorithms with minimal complexity and highly competitive performance and scalability.

Identifying models of HVAC systems using semiparametric regression

Heating, ventilation, and air-conditioning (HVAC) systems use a large amount of energy, and so they are an interesting area for efficiency improvements. The focus here is on the use of semiparametric regression to identify models, which are amenable to analysis and control system design, of HVAC systems. This paper briefly describes two testbeds that we have built on the Berkeley campus for modeling and efficient control of HVAC systems, and we use these testbeds as case studies for system identification. The main contribution of this work is that the use of semiparametric regression allows for the estimation of the heating load from occupancy, equipment, and solar heating using only temperature measurements. These estimates are important for building accurate models as well as designing efficient control schemes, and in our other work we have been able to achieve a reduction in energy consumption on a single room testbed using heating load estimation in conjunction with the learning-based model predictive control (LBMPC) technique. Furthermore, this framework is not restrictive to modeling nonlinear HVAC behavior, because we have been able to use this methodology to create hybrid system models that incorporate such nonlinearities.

Distributed optimization with arbitrary local solvers

With the growth of data and necessity for distributed optimization methods, solvers that work well on a single machine must be re-designed to leverage distributed computation. Recent work in this area has been limited by focusing heavily on developing highly specific methods for the distributed environment. These special-purpose methods are often unable to fully leverage the competitive performance of their well-tuned and customized single machine counterparts. Further, they are unable to easily integrate improvements that continue to be made to single machine methods. To this end, we present a framework for distributed optimization that both allows the flexibility of arbitrary solvers to be used on each (single) machine locally and yet maintains competitive performance against other state-of-the-art special-purpose distributed methods. We give strong primal–dual convergence rate guarantees for our framework that hold for arbitrary local solvers. We demonstrate the impact of local solver selection both theoretically and in an extensive experimental comparison. Finally, we provide thorough implementation details for our framework, highlighting areas for practical performance gains.

Modeling building thermal response to HVAC zoning

HVAC systems account for 38% of building energy usage. Studies have indicated at least 5-15% waste due to unoccupied spaces being conditioned. Our goal is to minimize this waste by retrofitting HVAC systems to enable room-level zoning where each room is conditioned individually based on its occupancy. This will allow only occupied rooms to be conditioned while saving the energy used to condition unoccupied rooms. In order to achieve this goal, the effect of opening or closing air vent registers on room temperatures has to be predicted. Making such a prediction is complicated by the fact that weather has a larger effect on room temperatures than the settings of air vent registers, making it hard to isolate the influence of the HVAC system. We present a technique for dynamically estimating the heat load due to weather on room temperatures and subtracting it out in order to predict the effect of the HVAC system more directly.

A comparative study of high renewables penetration electricity grids

Electricity grids are transforming as renewables proliferate, yet operational concerns due to fluctuations in renewables sources could limit the ultimate potential for high penetrations of renewables. In this paper, we compare three electricity grids - California, Germany, and Ontario - studying the effects of relative cost of solar and wind generation on the selection of the renewables mix, and examine the resulting excess generation. We then observe the effects of the renewables mix and the use of baseload energy generation on the limits to renewables penetration, quantifying what proportion of delivered energy can be provided by renewables. Our study shows that the optimal renewables mix, from the perspective of minimizing total cost of generation, is highly dependent on the relative costs of technology, and that above a certain penetration rate, different for each grid, the optimal mix must contain both solar and wind generation.

Classification of sidewalks in street view images

Mapping sidewalks in urban environments is key in the creation of pedestrian-friendly, sustainable cities. Currently, urban planners are hindered by a lack of information available in a format suitable for the large-scale analysis of sidewalk design. To demonstrate the impact that information technology could have in this area, we leverage techniques from machine learning and computer vision to gather information about the presence and quality of sidewalks in map images. In particular, we identify sidewalk segments in street view images using a random forest classifier, utilizing a set of local and global features that include geometric context, presence of lanes, pixel color, and location. Our results illustrate that this approach is effective in classifying sidewalk segments in a large set of street view images. This algorithm can be easily extended to other datasets, and can be automated to gather complete, fine-grained details about sidewalks for arbitrarily large urban environments.

Going In-Depth: Finding Longform on the Web

tl;dr: Longform articles are extended, in-depth pieces that often serve as feature stories in newspapers and magazines. In this work, we develop a system to automatically identify longform content across the web. Our novel classifier is highly accurate despite huge variation within longform in terms of topic, voice, and editorial taste. It is also scalable and interpretable, requiring a surprisingly small set of features based only on language and parse structures, length, and document interest. We implement our system at scale and use it to identify a corpus of several million longform documents. Using this corpus, we provide the first web-scale study with quantifiable and measurable information on longform, giving new insight into questions posed by the media on the past and current state of this famed literary medium.