Neural networks are often employed to model the brain. One such model developed by Numenta’s research group is the hierarchical temporal memory (HTM) model. A comparison of the HTM model and more conventional neural networks such as the network architectures available in the Stuttgart Neural Network Simulator (SNNS) is missing from the literature. This study aims to report on the performance on machine learning problems of HTM neural networks as compared to the performance of other common network architectures. The HTM neural network will either be built on the JAVA platform using the descriptions in Numenta’s publications or be used as provided by Numenta for public use. The SNNS network architectures will be used as provided by the University of Stuttgart. Possible learning problems considered will include: motion detection, character recognition, recognition and tracking of humanoid forms, object identification, and general pattern recognition. Possible SNNS network architectures include: backpropagation through time (for recurrent networks), quickprop through time (for recurrent networks), self-organizing maps (Kohonen maps), TDNN (time-delay networks) with backpropagation, Elman networks and extended hierarchical Elman networks. Evaluation of a network’s performance on each problem will be based on accuracy of identification, and time required for computation, while holding learning resources constant (same time exposure and same number of examples during learning stage). A qualitative comparison of the similarities of the learning problems to cortical learning algorithms and brain architectures may be attempted depending on time constraints. These results will be useful for evaluating both the usefulness of HTM neural networks and their connection to the cortical learning algorithm.

People with severe visual impairments cannot drive, so they often rely on public transit. The usability of the public transit system is, therefore, of crucial importance to blind people. One major challenge blind people experience when using public transportation is finding the exact location of bus stops. We have developed StopFinder, a system that uses crowdsourcing to provide information about non-visual landmarks around bus stops to enable blind people to find them more easily when navigating with a cane or a guide dog. It is developed in iPhone, and blind people can use our system using the access technology called VoiceOver, the built-in feature of iPhone that enables them to hear the information in the screen and navigate screen through simple hand gestures like single or double finger flicks, taps etc. The crowdsourced information includes which street/avenue has the bus stop, which direction to walk to reach the bus stop once you arrive at the intersection and what to expect once you arrive at the bus stop like the presence of shelters, garbage cans, benches etc. Since we are using crowsourcing for getting information, we provide an option to rate the entry. In addition to accessing and rating information about a bus stop, blind people can also contribute information through a non-graphical interface. The interface consists of mostly multiple-choice questions that are convenient to fill in a short time while waiting for a bus using VoiceOver. Since one of our primary foci has been the ease of using Stopfinder, we conducted interviews with blind people that informed the design of the application.

Given a very fast calculator and a digital camera, is it possible to relate what the camera “sees” to everyday life as we know it? Impossibly fast electronic signals pass from our optic nerves to our brains to our muscles to give us the power of object recognition and the ability to articulate that recognition. Algorithmically, this project attempts to take the wizardry of that recognition and “dumb it down” for computer consumption. Using the Python coding language, through a mathematical approach, an attempt is made to add rudimentary letter and word recognition to a Scribbler robot. The input to be analyzed is gathered by a hardware add-on, which provides a 220 x 220 pixel still-image camera to the Scribbler robot. Through an established letter recognition scheme, the Python code scans for matching values and assigns letters accordingly. The letters are then stored into string values and fed to the Scribbler's built-in text-to-speech command. Through this experiment, the bridge between human and computer understanding will be linked, if only for a crude, brief second.

OneBusAway is an application providing real-time arrival information to tens of thousands of transit riders daily. As a system built on providing reliable information, one of the cornerstones of OneBusAway is threatened when it does not provide accurate information to its users. Our surveys and interviews of current OneBusAway users confirmed that inconsistencies in the application have a significant impact on riders’ transit experience by making them miss their bus and lose trust in the application. In addition, users of the application have limited options for reporting issues with inaccurate information and lack the means to easily receive targeted feedback about their reports. Furthermore, reports received by OneBusAway are not currently categorized, verified for validity, or sent to one central repository, leading to a significant amount of manual work for OneBusAway developers.To address this problem, we designed an issue-reporting system that relies on a network of users to submit, approve, and resolve issues encountered in OneBusAway. Using OneBusAway to establish communication between transit riders and agencies, our system will direct issue reports to the source for resolution and provide feedback to users when the problems are resolved. Our application will empower users with the tools to get actively involved in their transit community and improve the transit experience.

A modern computer control system contains mainly four parts, a process (e.g., a water tank where the water temperature is adjustable), an actuator (e.g., a heater to warm up the water), a sensor (e.g., the thermometer to get the real temperature of the water), and the controller. The controller compares the set-point of the interested signal (e.g., the water temperature) with the real measurement from the sensor, hence closed control loop; it also computes the adjustment signal and sends it to the process actuator so that the process will generate an output as close as possible to the set-point. Traditional computer control systems use wired sensors and actuators and have served well in the past century. With the development of wireless sensors and wireless communication technologies, there is a trend to make wireless a viable replacement for hard-wired control, due to low cost, mobility and ability to operate in a harsh or ultra clean environment. However, challenging issues arise; one of them is reliability, which will be investigated in our research. We use a water tank temperature control system as the test-bed; the controller is the MCF52233DEMO ColdFire board with a Real Time Operating System (RTOS) kernel; the control algorithm is the widely applied Proportional Integral Derivative (PID); and the wireless standard is the IEEE 802.11b (Wi-Fi). By comparing the performance between the hard-wired control system and the wireless control system using the same PID controller, we expect to see poor performance of the wireless control system due to wireless network issues such as packet loss, time delays, and so on. We will modify the classical PID algorithm for wireless control so that similar performance to hard-wired control could be achieved. In the future, we will investigate the co-design of communication protocols and control strategies for wireless control systems.

Cameras have become a fundamental part of our daily lives because of the powerful ability of images to communicate, preserve, and educate. Computational photography is an emerging field of study that considers sensory information from a scene to be intermediate data, and uses computation to form the final photo. The goal is to improve images by combining computing power with digital imaging to create new images that no digital camera on the market today could capture. Techniques that represent computational photography include high-dynamic-range (HDR) imaging, panoramic stitching, and light-field imaging. Last year, Andrew Adams, a leading Stanford researcher behind computational photography, released a fully programmable digital camera system to aid other researchers in computational photography. Named FCam, it is an open-source C++ application programmable interface (API) for easy and precise control of digital cameras. My research will focus on using the FCam API to develop faster, smarter algorithms to capture images for accessibility purposes. Computational photography can be directly applied to several accessibility research projects. The project I am currently working on improves the visibility of street signs so that another person or a computer can recognize the text. By quickly capturing multiple images with different levels of luminosity, an optical character recognition engine can provide better textual results to the user. This research demonstrates the importance of using computational photography to make information more accessible to people with disabilities.

Given a boolean function, we can construct a Free Branching Decision Diagram (FBDD) which we can then use to calculate the function. Any query on a database that returns true or false creates a boolean function, and the existence of FBDDs whose sizes are polynomial in the size of the database for these derived expressions can be used to evaluate these queries quickly. In a recent paper, Suciu and Jha proved that one query in particular, which they called h_1, creates boolean functions which cannot have polynomial sized FBDDs. We are currently attempting to generalize this result to a class of queries h_k which are intimately related to h₁. While we are fairly certain that it is true that the h_ks do not have polynomial sized FBDDs because they cannot be evaluated in polynomial time, many of the properties used to prove the hardness of h₁ are lost when k is greater than one. Hence we are currently attempting to generalize and apply ideas first formulated by Simon and Szegedy in their paper "A New Lower Bound Theorem for Read Only Once Branching Programs" to our boolean formulae and their resulting FBDDs. While we currently have no positive results, we have definitively ruled out a number of approaches and we believe that a positive result is forthcoming. Such a result would prove a number of conjectures made by Jha and ultimately we hope that it will lead to a complete characterization of queries which have a polynomial sized FBDD.