|OCCaM is a new lab at Olin College. My name is Paul Ruvolo, and I am the director of the lab as well as a new faculty member in Computing. If any of the information below sparks your interest, please send me an e-mail.|
We are interested in the creation of intelligent machines. By intelligent we mean capable of flexibly and adaptively solving a wide array of tasks that do not have obvious solutions. By machines we mean a variety of things such as robots, mobile computing devices, and software programs. By we, I currently mean I (since I am still in the process recruiting my first group of students). While the quest for creating artificial intelligence has long been a goal of computer science, a number of recent trends (both hardware and software) have made the present a time of nearly unlimited promise for the advancement of AI research. The trends driving our optimism include:
- New developments in machine learning: machine learning is a new field of computer science concerned with developing, analyzing, and applying algorithms for learning from data. These algorithms have significantly broadened the abilities of machines to solve tasks in such diverse fields as computer vision, data mining, robotics, and control.
- New computing hardware: the availability of cheap embedded systems (such as Arduino, Raspberry PI), along with the proliferation of smart phones, cheap video sensors, as well as cloud-based computing platforms such as Amazon's EC2 provide many new opportunities for computer science researchers.
- New ways to learn from humans: projects such as the ESP game and Amazon's mechanical Turk have shown that internet technology can be used to harvest vast amounts of human intelligence to solve interesting computational problems.
These trends enable motivated students to contribute to the goal of developing intelligent machines through research in areas such as algorithms and analysis, software development, electrical and computer engineering, design, and data analysis.
Crowdsourcing Sensorimotor Intelligence for Assistive Machines
This project starts from the following premise: there is a huge pool of people that want to volunteer to help others, but don't due to a variety of factors including: the difficulty of investigating where their talents can be utilized, physically commuting to the location for volunteering, and finding times where they are free and their services are needed. Wouldn't it be great if we could overcome all of these limitations and thus free up this huge untapped pool of goodwill? We are working to harness cutting edge computer science to do just that by leveraging human volunteers to power intelligent machines for helping disabled populations.
Picture from: http://grozi.calit2.net/
For more information, check out our Assistive Technology Project Page
Connections to Industry:
There are several computer vision companies focusing on technology such as logo recognition which pertain to this project. One such company is Orpix
who develop logo recognition, vehicle recognition, and other computer vision solutions. We plan on collaborating with such companies when appropriate.
another major class of agents (besides people with sensory impairments) that require sensorimotor assistance is robots
. A major obstacle to deploying robots as assistants in the home is that modules for object recognition, grasping, and manipulation do not work well in the cluttered, unconstrained environments, and highly variable environment of the home. The next phase of this project will be to partner with the folks at Olin Robotics to leverage this crowdsourced approach to sensorimotor intelligence to provide help to an assistive robot rather than to a human.
Other Research Interests:
here at OCCaM, we try to keep an open mind. In my experience this is necessary to do research. In this spirit, if you have a particular research direction you want to explore that is related to my interests I would love to talk. To get your creative juices flowing, below is a list of some other projects of mine. To view information on the papers cited below, cross-reference the citation number (e.g. ) with the publications section of the About Paul Ruvolo
Lifelong Machine Learning
Traditional machine learning algorithms learn tasks in isolation. While this works well given sufficient data computational power, in the case where these conditions are not met we want to be able to learn tasks efficiently in sequence while both building upon and refining previously learned knowledge. In previous work, , we have shown that we can get near identical performance to competing methods and achieving about a 1000x speedup while providing various theoretical guarantees.
Quality Control Methods for Crowdsourced Data
How to best combine input from multiple people performing some crowdsourced task is an active area of research. Issues include how to automatically assess participant expertise, instance difficult, and participant specialization. In previous work, , I have derived various machine learning based methods for automatic quality control of crowdsourced data.
Inverse Optimal Control and Goal-Based Imitation
A very hot topic in machine learning is inferring the goals of an agent from instances of their behavior. These goals are typically formalized as particular performance objectives that an agent may be trying to achieve. It has been shown that goal-inference can serve as a very effective way to allow a robot to learn skills from a human. As part of my thesis, , I developed several new techniques to infer goals both as a means of skill transfer and as a means of explaining the behavior of people. My aim is to publish this in a good conference, but I need to run some good experiments first (perhaps you are interested in helping?).
I have worked on machine learning methods for: recognizing auditory categories such as crying and emotion in human speech , localizing audio , fusing auditory and visual information , and recognizing facial expressions .
At UC San Diego I did research on a project to try to program a robotic infant that would learn in a manner similar to how human infants learn -- through experience interacting with its environment. In this project we investigated various tools from optimal control as a means to allow a robot to autonomously learn to more effectively interact with the physical and social world. My relevant publications on this topic are: .
Computational Analysis of Behavior
I collaborate with Daniel Messinger's developmental psychology lab at the University of Miami. In our research, , we apply cutting edge machine learning techniques to the task of characterizing and understanding infant social and motor development. Projects based on our work are still flourishing at the University of Miami. For this reason, Olin students have ample opportunity to work as part of an interdisciplinary team in concert with the scientists in Daniel Messinger's lab.
Preschool has been shown to be one of the most crucial times for children's learning. In the past, , I have worked on creating a robotic teaching aide to help 18-24 month old children learn new vocabulary.
Lab Meetings: once we have a critical mass of members, we will setup a weekly lab meeting.
The Olin Machine Learning Reading Group
will meet weekly to discuss machine learning opportunities at Olin College. In terms of personnel, this group may have significant overlap with OCCaM.