|
PureBytes Links
Trading Reference Links
|
>>all of the images without tanks had shadows and all of the images with tanks did not. So the system they had trained was looking for images with shadows, not tanks. <<
Here's another story...
Cruise missiles have multiple guidance mechanisms, including optical, which enables it to identify buildings, streets, etc. During the research phase years ago, a neural net was trained to infer building shape and location from the different brightness values of a building's various flat sides. Training images were obtained from satellite photos of San Francisco. But it performed miserably on its intended targets. (not Frisco) Why? The NN's preprocessing helped detect sides by first finding sharp corners or edges of high visible contrast and then scanned along those edges to find corners. Corners define a building's shape and size. In some locations, buildings had rounded sides and no corners, driving the preprocessor into the equivalent of digital epilepsy. Moral of the story: 1) You can't anticipate everything that may go wrong. 2) Live in a dome, it's safer. :-)
- mark jurik
|