A recent post by babaganoush discussed the issue of privacy in keyword-based advertising. As touched on in class, maintaining privacy is one of the chief concerns moving forward as we develop more sophisticated methods of targeted advertising. A recent paper by Chopra and White discusses various implications for privacy as we begin to deal with highly advanced automated agents. Consider the following scenario where Google’s technology becomes advanced enough such that it can discern that terrorists are using Gmail to plot an attack on a major city. One can imagine the government’s desire to create laws to leverage this technology (analog of a human failing to report a crime being itself a criminal act). As agents become more sophisticated, companies such as Google will likely face increasing pressure when striving to protect the privacy of their users.

Another post by sithswine182 touched on the topic of multimedia search (e.g., searching for pictures or videos). One way to make advancements in this area is to use pre-labeled data (such as images annotated with keywords) to detect statistically meaningful features to feed back into the ranking algorithms. Unfortunately, our collection of annotated multimedia data is very noisy. To address this deficiency, Google developed its Image Labeler site. This site is played like a game where you and another anonymous user try to annotate a random image with as many keywords as possible. All common keywords submitted by both you and your partner are scored. As an incentive, Google keeps track of the top scoring users and user pairs.

Machine learning is an area of computer science which tries to learn salient features from pre-labeled data. One of the most popular machine learning methods is what is known as an Artificial Neural Network (ANN). ANNs are inspired by our understanding of our own neural networks, which was discussed in shadow’s recent post. In an ANN, data features are interpreted as signals and passed through layers of artificial perceptrons. Each perceptron emits a response signal whenever it receives a set of input signals. A perceptron can also receive a feedback signal which it uses to adjust its responses. Individually, the function of each perceptron is difficult to interpret (much like our brains). But collectively, the ANN usually provides good predictive ability on future data similar to the pre-labeled data it was trained on. For those interested in learning more, there are a number of good tutorials available on the internet, such as this one.