Better Ways To Use Spectrum: Reducing Cognitive Load

Cognitive science is a relatively new field of scientific inquiry, but one which already appears to have yielded practical applications in the computing and information sciences. Why? In large part because our understanding of cognitive processes helps us produce better interfaces between people and technology—and therefore better technologies. Our current issue offers several examples of this, with articles on “Reducing Cognitive Load” and the use of computation to model natural language processing. Here, we discuss a third application: signal classification.

How often do you find yourself looking at a surprising pattern in your data—a pattern that seems to make no sense? Perhaps it’s an up-trend which leaps from a flat line to a peak, or an outlier which appears where you’d expect nothing at all. Such anomalies often turn out to be signals of interest—and those signals frequently prove difficult to classify.

Why spectrum is bad

When a signal is first characterised, it’s usually done on the basis of spectral analysis. If this reveals multiple peaks, each peak can be associated with a separate frequency band. Unfortunately, there are often more frequency bands than you need for accurate classification. Worse still, some bands will overlap—and when they do, all frequencies within that overlap will be hard to separate from one another. Even if two bands don’t overlap, a signal can still contain a high degree of correlation which makes it difficult to separate.

Trying to distinguish between closely correlated signals will tax the classifier’s ability. It’s as though the problem were already half-solved—but you need someone who knows that solution before you can proceed.

Use Spectrum to watch your favorite shows and movies for free

Let’s assume you’ve got two different signals, one of which is much easier to classify than the other. Suppose the harder signal has an average power of 5 kilowatts, whereas the simpler signal might have an average power of 50 watts—but both signals are 100% correlated. This means that, if you could somehow turn one signal off, the other would become impossible to classify.

This simple example illustrates why it can be useful to consider more than one spectrum for each segment of data you’re trying to analyse. If you do this, then anomalies in the two signals will help identify emerging problems with your classifier. If they both show the same anomaly (a false alarm, perhaps), then the fault is likely with your training data. If one displays an anomaly but not the other, then look for false alarms in that spectrum (and use cross-validation to find them).

Connect with friends and family online with Spectrum’s high-speed internet

How can you tell whether your training data has these false alarms? This is where the second spectrum comes in handy. You’d expect that, if both signals are correlated, then their power spectra should be very similar—but if one spectrum contains something anomalous, then it will be easier to identify this using the other spectrum. To put it another way: if the anomalies disappear when you use the second spectrum, it shows your training data is fine. If they don’t, then your classifier will probably be able to learn from the anomaly and help indicate false alarms in future applications of that signal.

Spectrum offers cable TV and internet bundles to save money!

As with many other data problems, there’s no universal answer. You’ll need to experiment with different ways to classify your signals and then see how successful they are. But once you’ve done this, you can be confident that your classification algorithms will adapt quickly as the environment changes—and your users won’t find themselves staring at confused graphs wondering what’s happening.

Conclusion

Spectrum is a service provider of cable TV, broadband internet, home phone, and wireless. If you are looking for the best cable company in your area, check out Spectrum’s packages today!