Author Archive

Posted by Anthony Francis

Maybe your computer just needs a hug
I'm Anthony Francis, an artificial intelligence researcher working in Google's Search Quality group.  One of the things I like about Google is that we give back to the world in ways that make sense for our business, both as a company and as individuals.  Google's search engine runs on electrical power, and so Google as a company invests in renewable energy and encourages Googlers as individuals to conserve energy.  Google's search engine runs on open source software, and so Google as a company open sources its software and encourages Googlers to participate in existing open source projects.

What may be less obvious is that Google's search engine runs on ideas.  Google as a company has a great Research department, but it also encourages Googlers to participate in the research community - both in visible, obvious ways like writing papers and attending conferences, but also in less visible but still necessary chores that keep the research community running, like reviewing papers and organizing conferences.  What I'd like to tell you about today is how Google helped me give back to the research community by letting me take time to write a paper on work I did before I came to Google.

The Personal Pet Project
Ashwin Ram's Cognitive Computing Lab at Georgia Tech has been working on emotional agents for over ten years, first with robots, and more recently in computer games and virtual environments.  I worked on one of our earliest efforts in this area, the PEPE (Personal Pet) project, a joint effort with Yamaha to develop an robotic pet.  Dr. Ram noticed that many consumer electronics require configuration steps that are frustrating, even baffling: everything from VCRs to toasters has blinking clocks and unused features.  Pets, on the other hand, don't have to be "configured": they understand us on emotional level, noticing what makes us happy or angry, remembering what brings reward and punishment, and learning to respond appropriately without us reading a manual or pushing a button.

PEPE emulated this intuitive understanding with an emotional long term memory module, which I developed.  This module had three parts: basic emotions, emotional memory, and emotional reminding.  Basic emotions gave the robot realistic behavior "out of the box": for example, the robot interpreted being petted on the head as being pleasant, which made it want to socialize; it interpreted a kick to the rear as unpleasant, which made it want to hide.  The emotional memory associated these primitive emotions with the people and objects the robot saw in its environment.  Emotional reminding closed the loop: when the robot saw a person or object again, it was reminded of the past emotion, which influenced but did not determine the robot's ultimate emotional expression.  So the robot naturally wanted to play with people who petted it and fled people who had kicked it in the past, but it would be possible to overcome that past learning by ignoring the robot (if you didn't want to play with it as much anymore) or by petting it (if you had kicked it but no longer wanted it to be scared of you).

I worked with a team of engineers at Yamaha in Japan to implement this emotional long term memory on a small robot they were building.  Our job was easier because PEPE's basic architecture enabled creating many different kinds of behaviors that could easily interact with each other, in parallel or in sequence.  On top of this architecture, we added basic emotions, emotional learning and emotional reminding in stages.  Our initial tests were positive, and upon my return to the United States, we began porting this software back to Georgia Tech's PEPE robot.  PEPE even had a brief moment of fame, appearing on a local TV station; ultimately, however, Georgia Tech and Yamaha decided to move on to new projects.

Characters with Personality
But the dream of emotional agents didn't die.  Manish Mehta, a graduate student at Georgia Tech researching interactive games, began working with Ram to try to develop more sophisticated models of emotional change.  Mehta realized that emotional events can act as a trigger for behavioral change: if we humans try something that works really poorly, or spectacularly well, the emotions generated by that experience can prompt us to make changes to our behavior in the future --- for example, remembering to bring that umbrella in the future so we no longer have to walk home wet in the rain.

Mehta used ABL, an agent language developed by Michael Mateas and Andrew Stern for the Facade project.  Like the fundamental architecture of PEPE, ABL enabled the author of an agent to design complicated behaviors that could interact in sequence, in parallel, or through complex triggering.  Mehta extended ABL to allow an agent to rewrite its own behaviors, and triggered that rewriting based on charged emotional events.  This worked: using two agents playing tag, Mehta was able to make them revise their behaviors based on how well they played the game. However, it had an unintended consequence: one of the agents decided it was less stressed out when it was "IT".  When it was "IT", it wasn't being chased and wasn't feeling stressed, so it just let itself get tagged, enabling it to "chase" the other player at a leisurely pace.

While this result was surprising, even amazing, it failed to achieved the original objective.   Mehta realized this had exposed one more property of human behavioral change: emotional events can prompt change, but we also have a self image - a model of what we think our personality should be like.  As our personalities change, we audit the changes to make sure they fit our self image and apply further corrections.  For a computer game character, that "personality" is really the designer's intent, so Mehta augmented his emotional learning system to check potential behavioral changes and make sure that they did not violate the original intended design.  With this change, both characters were able to learn from their good and bad experiences playing tag, and neither "quit playing the game" just because they didn't want the stress of being chased.

Putting It All Together
The paper that Ram, and Mehta and I wrote for the Handbook of Synthetic Emotions and Sociable Robotics reports this work and ties together all of these threads.  We were not alone in challenges we faced in developing intelligent agents: others faced them too, in robotics, computer games and academic AI.  Many people use flexible behavior systems like PEPE's architecture or like ABL, but building large systems out of these can be challenging and extremely labor intensive.  What Mehta and I found, however, is that the use of emotion models acted like a force multiplier: adding basic emotional responses to an existing behavior system radically increased the flexibility and apparent realism of an intelligent agent, whether it was a robotic puppy or a character in a computer game.

Moreover, the more components of the emotion stack you add to an agent, the easier it becomes to extend.  While adding emotion made PEPE extended the behaviors it could perform; modifying those emotions using memories changed the character of its behavior, making it seem more lifelike.  Mehta's work in extending ABL took this further, enabling creative behavior changes in response to agent frustration.  Adding personality models makes these changes stable, enabling the agent to adapt to new situations while staying consistent with their designer's intent.  Often, just adding more behaviors to an existing agent can make it more brittle; we found in contrast adding new layers inspired by research in human and animal emotion made the agent more sophisticated and robust.

Giving Something Back
Writing this paper did not take much time from Google.  Most of it was done in the evening on my own time, with the occasional email exchange with my coauthors and editors during the day while I was waiting on compiles.  And that is how it should be: it may be a long time before this work directly benefits Google, since we do not currently develop robots or computer games.  But Google still benefited because the advances that we need to improve our search engine often start in academia.

By encouraging our staff to follow up on their research work and to contribute to the research community, Google supports the growth of the next big idea.  Knowing that we can follow up on our past work makes researchers at Google feel more empowered to pursue new ideas and continually exposes us to sources of inspiration.  And that inspiration, which makes Google a better place to work, is exactly what I need to keep trying new ideas, which in the end will help make Google a better place to search.

Posted by Tomáš Ižo (Software Engineer) and Jay Yagnik (Head of Computer Vision Research)

Posted by Juan Vargas, University Relations

Posted by Vahab S. Mirrokni and Muthu Muthukrishnan

Google auctions ads, and enables a market with millions of advertisers and users.  This market presents a unique opportunity to test and refine economic principles as applied to a very large number of interacting, self-interested parties with a myriad of objectives. Researchers in economics, computer science, operations research, marketing and business are increasingly involved in defining, understanding and influencing this market. 

On January 7th, a group of computer science researchers from Google and various universities formed a workshop to discuss key research directions in this area, specifically "Market Algorithms and Optimization." Corinna Cortes (Head, Google Research, NY)  and Alfred Spector (VP of Research, Google) gave short talks about research groups in Google, academic collaborations, and research awards. The morning session comprised talks by Google researchers including Noam Nisan, Jon Feldman, Vahab Mirrokni, Yishay Mansour, and Muthu Muthukrishnan. These talks explored the following topics and key issues: 

• Role of budgets. Identify suitable properties of mechanisms for repeated auctions in the presence of budgets. Truthfulness is impossible, and may not be the suitable property.  
  
• Advertisers' bidding strategies. While bidding equally on all keywords has certain desirable properties,  identify other bidding strategies if advertisers want to maximize different utility functions and  use rich bidding features.
  
• Dynamics of ad games. Understand dynamics of sophisticated advertiser bidder strategies under various ad allocation rules. For proportional allocation rules, the dynamics of simple bidder strategies are described here.
• Online Reservations. Design suitable mechanisms for selling ad inventory in the future rather than on the spot. Models and methods for what happens when reservations can not be honored are explored in WWW08, WINE08, and SODA09 papers.

The research presented in these talks can be found at the Google publications site. Some general research directions and open problems in ad auctions are here.

The post-lunch (sushi included of course) session comprised talks by researchers from academia.  Bobby Kleinberg (Cornell) went first and took the audience far into projective geometry in an attempt to understand equilibria resulting from a sequence of selfish behavior of players using specific learning algorithms. Silvio Micali (MIT) described how difficult it was to model or propose mechanisms in presence of collusions, and then went on to show how to correctly design such mechanisms for combinatorial auctions. Kamesh Munagala (Duke Univ.) and Anna Karlin ( U. Wash) discussed algorithmic problems related to ad auctions, including how to run auctions in presence of advertisers with mixed utilities and tight remnant budgets. Other participants -- Richard Cole (NYU), Amos Fiat (Tel Aviv), Uri Feige (Weizmann), Michel Goemans (MIT), Anupam Gupta (CMU),   Nicole Immorlica (Northwestern), Michael Rabin (Harvard), and Eva Tardos (Cornell) -- kept it lively with questions and discussions.

Part of what makes Google ad systems exciting is the challenges they pose and overcome, and yet others in the horizon. It is remarkable how some of the fundamental problems Google ad systems grapple with are also some of the hardest research problems in the community of Market Algorithms.  Joint Google and Academia meetings like this help researchers begin to attack these problems, and may be a model for  research collaborations in the future.

Posted by T.V Raman, Research Scientist

A little over a year ago, I gave a lightning talk at the W3C Technical Plenary in Boston where I looked forward to what came after Web-2.0. The key insight underlying that talk was that the Web was now mature enough for us to build Web technologies purely out of Web parts. Web-2.0 is a result of applying the Web to itself and is therefore better thought of as Web(Web()) or more concisely, Web².

Today, we can build new web artifacts out of existing ones by aggregation (web mashups) and by projection (filtered views), and publish the resulting artifacts on the web by assigning them a URL. This leads to the insight that this web that is to come potentially consists of the power-set of all web content. These ideas, and their logical consequences are detailed in article entitled Toward 2^W --- Beyond Web 2.0 in the February edition of the Communications Of The ACM. You can also find a slightly more extensive blog I posted here.

Posted by Fernando Pereira, Research Director

Corinna Cortes, Head of Google Research in New York, has just been awarded the ACM Paris Kanellakis Theory and Practice Award jointly with Vladimir Vapnik (Royal Holloway College and NEC Research). The award recognizes their invention in the early 1990s of the soft-margin support vector machine, which has become the supervised machine learning method of choice for applications ranging from image analysis to document classification to bioinformatics.

What is so important about this invention? In supervised machine learning, we create algorithms that can learn a rule to accurately classify new examples based on a set of training examples (e.g. spam or non-spam). There is no single attribute of an email message that tells us with certainty that it is spam. Instead, many attributes have to be considered, forming a vector of very high dimension. The same situation arises in many other machine practical learning tasks, including many that we work on at Google.

To learn accurate classifiers, we need to solve several big problems. First, the rule learned from the training data should be accurate on new test examples, even though it has not seen those examples. In other words, the rule must generalize well. Second, we must be able to find the optimal rule efficiently. Both of these problems are especially daunting for very high dimensional data. Third, the method for computing the rule should be able to accommodate errors in the training data, such as messages that are given conflicting labels by different people (my spam may be your ham).

Soft-margin support vector machines wrap these three problems together into an elegant mathematical package. The crucial insight is that classification problems of this kind can be expressed as finding in very high dimension (or even infinite dimension) the hyperplane that best separates the positive examples (ham) from the negative ones (spam).

Remarkably, the solution of this problem does not depend on the dimensionality of the data, it depends only on the pairwise similarities between the training examples determined by the agreement or disagreement between corresponding attributes. Furthermore, a hyperplane that separates the training data well can be shown to generalize well to unseen data with the same statistical properties.

Now, you might be asking how could this be done if the training data is inconsistently labeled. After all, you cannot have the same example on both sides of the separating hyperplane. That's where the soft margin idea comes in: the quadratic optimization program that finds the optimal separating hyperplane can be cleverly modified to "give up" on a fraction of the training examples that cannot be classified correctly.

With this crucial improvement, support vector machines became really practical, while the core ideas have had huge influence in the development of further learning algorithms for an ever wider range of tasks.

Congratulations to Corinna (and Vladimir) on the well-deserved award.

Posted by Jeff Walz, University Relations and Anne Bray, Head of Agency, WPP

Google and the WPP Group have teamed up to create a new research program with the goal of improving industry understanding of digital marketing. Eleven research awards have been given to universities through the Google and WPP Marketing Research Awards Program, announced by both companies in the fall of 2008. The academic studies will harness WPP client data to explore how online media influences consumer behavior, attitudes, and decision making. The research provides an opportunity for very innovative thinking in an area that is at the crossroads of marketing, computer science, economics, and various mathematical disciplines.

More than 120 entries were received by the deadline for proposals. The awards represent the first round of grants in the three-year program towards which WPP and Google will commit up to $4.6 million in an effort to support research around digital marketing. Hal Varian, Google's Chief Economist, participated on the decision committee.

The winning projects offer convincing designs for exploring how online and offline marketing influence consumer attitudes, decisions, and purchase behavior. As marketing continues to become more digital and more measurable, the results of these studies will also advance our understanding of how advertising investment should be allocated among media channels.

The researchers and affiliated academic institutions participating in this first round of awarded projects are:

• “Effect of Online Exposure on Offline Buying: How Online Exposure
Aids or Hurts Offline Buying by Increasing the Impact of Offline
Attributes”; Amitav Chakravarti, New York University, Stern School of
Business, Department of Marketing

• “The Interaction Between Digital Marketing Tactics and Sales
Performance Online and Offline”; Elie Ofek, Associate Professor
Marketing, Harvard Business School and Zsolt Katona, Associate
Professor of Marketing, UC Berkeley, Haas School of Business

• ”Are Brand Attitudes Contagious? Consumer Response to Organic
Search Trends”; Donna L. Hoffman, Professor, A. Gary Anderson
Graduate School of Management, University of California Riverside and
Thomas P. Novak, A. Gary Anderson Graduate School of Management,
University of California Riverside

• “Does internet advertising help established brands or niche ("long
tail") brands more? Catherine Tucker, Assistant Professor of
Marketing, MIT Sloan School of Marketing and Avi Goldfarb, Associate
Professor of Marketing, Joseph L. Rotman School of Management
University of Toronto

• “Marketing on the Map: Visual Search and Consumer Decision Making”;
Nicolas Lurie, Assistant Professor of Marketing, College of
Management, Georgia Institute of Technology, College of Management and
Sam Ransbotham, Assistant Professor of Information Systems, Carroll
School of Management, Boston College

• “Methods for multivariate metric analysis; identifying change
drivers”; Trevor J. Hastie, Professor, Department of Statistics,
Stanford University

• “Unpuzzling the Synergy of Display and Search Advertising: Insights
from Data Mining of Chinese Internet Users”; Hairong Li, Department of
Advertising, Public Relations, and Retailing, Michigan State
University and Shuguang Zhao, Media Survey Lab, Tsinghua University

• “Optimal Allocation of Offline and Online Media Budget”; Sunil
Gupta, Professor of Business Administration, Harvard Business School;
Anita Elberse, Associate Professor, Harvard Business School; and
Kenneth C. Wilbur, Assistant Professor of Marketing, Marshall School
of Business, University of Southern California

• “Targeting Ads to Match Individual Cognitive Styles: A Market
Test”; Glen Urban, Professor, MIT Sloan School of Management

• “How do consumers determine what is relevant? A psychometric and
neuroscientific study of online search and advertising effectiveness”;
Antoine Bechara, Professor of Psychology and Neuroscience, Department
of Psychology/Brain & Creativity Institute, University of Southern
California and Martin Reimann, Fellow, Department of Psychology/Brain
& Creativity, University of Southern California

• “A Comprehensive Model of the Effects of Brand-Generated and
Consumer-Generated Communications on Brand Perceptions, Sales and
Share”; Douglas Bowman and Manish Tripathi, Professors of Marketing,
Goizueta Business School, Emory University.

You can find more information about the Google and WPP Marketing Research Awards Program on the website.

Posted by Fernando Pereira, Google Research

Alon Halevy, Peter Norvig, and I argue that we should stop acting as if our goal is to author extremely elegant theories, and instead embrace complexity and make use of the best ally we have: the unreasonable effectiveness of data. See the full article here (IEEE Intelligent Systems, March/April 2009).

Posted by Hal Varian, Chief Economist and Hyunyoung Choi, Decision Support Engineering Analyst

Can Google queries help predict economic activity?

The answer depends on what you mean by "predict." Google Trends and Google Insights for Search provide a real time report on query volume, while economic data is typically released several days after the close of the month. Given this time lag, it is not implausible that Google queries in a category like "Automotive/Vehicle Shopping" during the first few weeks of March may help predict what actual March automotive sales will be like when the official data is released halfway through April.

That famous economist Yogi Berra once said "It's tough to make predictions, especially about the future." This inspired our approach: let us lower the bar and just try to predict the present.

Our work to date is summarized in a paper called Predicting the Present with Google Trends. We find that Google Trends data can help improve forecasts of the current level of activity for a number of different economic time series, including automobile sales, home sales, retail sales, and travel behavior.

Even predicting the present is useful, since it may help identify "turning points" in economic time series. If people start doing significantly more searches for "Real Estate Agents" in a certain location, it is tempting to think that house sales might increase in that area in the near future.

Our paper outlines one approach to short-term economic prediction, but we expect that there are several other interesting ideas out there. So we suggest that forecasting wannabes download some Google Trends data and try to relate it to other economic time series. If you find an interesting pattern, post your findings on a website and send a link to econ-forecast@google.com. We'll report on the most interesting results in a later blog post.

It has been said that if you put a million monkeys in front of a million computers, you would eventually produce an accurate economic forecast. Let's see how well that theory works ...

Posted by Ben Bayer, Google Research

Alfred Spector, Google's VP of Research, tells COMPUTERWORLD the ins and outs of Research at Google and where it's headed for the future. Read the complete interview here.