UCB annual state of R & D in EECS

February 17, 2011, University of California, Berkeley–The Berkeley EECS Annual Research Symposium (BEARS) offered a quick look into the leading edge of research at Berkeley. The four presentations (Powerpoint or PDF available on-line) looked at data creation and use across a number of areas.

This event was just after IBM’s Watson computer thumped its human adversaries on Jeopardy, so many of the speakers noted the fact. The first talk, by Ion Stoica entitled “A Berkeley View of Big Data in the Cloud.” considered what to do with the growing volumes of data in the world. Because terabytes don’t even make a least significant figure in this environment, the data world needs new algorithms plus better hardware and easier ways for people to be involved in the analysis.

In addition to the new paradigms for data evaluation, they also announced a new research initiative at the AMPLab. Working at the intersection of three massive trends: crowdsourcing, cloud computing, and powerful statistical machine learning, the AMPLab stands for Algorithms, Machines, and People. The AMPLab will focus on next-generation scalable machine learning algorithms, data management tools for large-scale and heterogeneous datasets, datacenter-friendly programming models, and an improved computational infrastructure.

The Founding Sponsors of the AMPLab are Google and SAP. The AMP Lab Sponsors are Amazon Web Services, ebay, Huawei, IBM, Intel, Microsoft, NEC, NetApp, and VMware. Cloudera is an AMP Lab Affiliate. The collective support of these companies is more than $2M per year. The AMPLab is expected to be a five-year project.

Next, Dan Klein talked about natural language processing. Despite Watson’s triumph, it used a brute force to language understanding. Although it represents the state-of-the-art in finding information using keywords and names. It is a knowledge-free “structural system that can only use superficial patterns to query search engines for its database. The next generation of natural language search will be able to synthesize information and recognize entities to relax the rigid requirements for highly structured queries.

By 2020, the tools will talk to embedded devices and have a deep understanding of patterns and inherent information. If a number of major breakthroughs come about, the search engines will be able to infer information and understand concepts. Some of its knowledge will come from grounded concepts, e.g. a chair is solid, a loveseat is a chair for two people. The tools will be able to talk to mobile robots and the interactions will change from a monolog to a dialog. Simple algorithms and large data have gotten us amazingly far! To go further, we need algorithms that work with deeper structure, systems that are contextualized, and learning methods that turn data into knowledge.

All of these data have to be personalized to be of greater value. Avideh Zakhor described augmenting reality with cloud-based data. This area reflects the growing adoption of smart phones with lots of sensors to feed data to the cloud. The combination of localization through GPS and cameras plus algorithms to do signal processing and image recognition can now look up metadata associated with the user’s environment via the stored data in the cloud.

The challenges are that the data from the phones are not high enough resolution. To be of value, the data must be pixel accurate, versus 10’s of meters from GPS. In addition, the user’s orientation is necessary–if you are looking at something, the sensor needs to know which way you are looking in addition to where you are. Once again, new algorithms and search techniques are necessary to add tags to the converted images to provide augmented reality for everyday use.

Looking ahead, almost certainly we will see many more AR apps on cell phones to include mobile advertising and location based services. Most likely to appear in this list are 3D AR apps with compelling user experience such as gaming and entertainment. The ultimate goal will be to blur the line between real and virtual environments.

And how will all these data get generated? Kris Pister suggested that ubiquitous wireless connectivity and pervasive sensors will enable new levels of visualization. The data cloud will be connected to a swarm of connected devices that can connect the cyber or the biological and physical worlds. The Immersed Human will be in an always-available augmented real-life interaction between humans and cyberspace, enabled by enriched input (sensory) and output (actuation, stimulation) devices on (and in) the body and in the surrounding environment.

This will not be easy. The challenges include communications of large numbers of multimedia data streams combined with critical sensing and control data with varying degrees of availability, mobility, latency, reliability, security, and privacy. The analysis and activation of features will require tremendous computational power to generate true real-time enhanced reality which is mostly provided by the “cloud” –but latency issues dictate locality. These capabilities will require distributed storage and all of the functions will have to operate within limited energy budgets.

To further their research, they are starting a new Swarm Lab. The Swarm Lab creates an environment for the creation of the most advanced “swarm nodes”, exploring post-Moore technologies and manufacturing strategies combined with ultra-low power implementation fabrics and architectures for computation, communication, and energy provision. It also contains an experimental playground for the exploration of innovative and disruptive swarm applications. Multi-disciplinary in nature, the lab combines researchers from diverse backgrounds covering the complete spectrum from application to technology and materials. The Swarm Lab has been launched with major support from Qualcomm Incorporated.