Simon Thorneycroft and Jonathan Hodgson

Neural Networks and Bayesian Analysis

It is interesting to read about how neural networks are being used in the MSN Search engine via Adi Olteans's blog.

Looking further at the Microsoft Research papers by Chris Burges I saw one on Audio Fingerprinting.

You have an incoming stream of audio and you'd like to know what's playing.  Our system can identify any one of about 240,000 songs in real time using about 10% CPU on an 833 MHz PC.  On 36 hours of noisy test audio, it achieves 0.2% false positives at 4.10-6 false negative rate.  Confirmation fingerprints can be used to significantly further improve these error rates, with almost no extra CPU cost.  Audio fingerprinting has lots of applications: for example, it can be used to construct audio thumbnails for songs, or find duplicate clips on your PC.

When I first tried Shazam from my mobile phone I was impressed it successfully tagged, Frank Sinatra performing 'Bad, Bad Leroy Brown' and I wondered how the technology worked and could search millions of songs so quickly. All I had to do was play the song into my mobile phone for about 15 seconds and then it txt'd me back about 5 seconds later with the artist and track names!

On the subject of SMS txt'ing check out Google SMS UK.

The Microsoft Research project RARE: Robust Audio Recognition Engine papers discuss a similar technology; including a paper on the feature extraction algorithms used, bitvector matching algorithms for fast lookup and using fingerprints to find choruses in music.

Another cool company which licenced work done by Microsoft Research is Inrix for data aggregation of traffic-related content using Bayesian network technology.

What is a Bayesian Network? Bayesian analysis calculates the likelihood of something happening in the face of some particular piece (or pieces) of evidence. Since, traffic volume cannot be predicted using physical laws, we leverage sophisticated Bayesian analysis to create graphs for specific metropolitan areas that map the relationships among real-time traffic, historical traffic patterns, weather, time of day, events, and many other variables. These maps are examples of a Bayesian network.

Where most traffic services can only determine what traffic is like right now – and only in a limited number of cities that have road sensor networks deployed. What they are doing is literally predicting the future which, in addition to providing real-time and incident information, enables answers to questions never before possible such as how long before the congestion there now clears up? Is this congestion normal or an anomaly? This can then be displayed graphically on mobile devices - One to watch.