How Smart Software Goes Off the Rails

June 23, 2019

Navigate to “How Feature Extraction Can Be Improved With Denoising.” The write up seems like a straight forward analytics explanation. Lots of jargon, buzzwords, and hippy dippy references to length squared sampling in matrices. The concept is not defined in the article. And if you remember statistics 101, you know that there are five types of sampling: Convenience, cluster, random, systematic, and stratified. Each has its strengths and weaknesses. How does one avoid the issues? Use length squared sampling obviously: Just sample rows with probability proportional to the square of their Euclidean norms. Got it?

However, the math is not the problem. Math is a method. The glitch is in defining “noise.” Like love, there are many ways to define love. The write up points out:

Autoencoders with more hidden layers than inputs run the risk of learning the identity function – where the output simply equals the input – thereby becoming useless. In order to overcome this, Denoising Autoencoders(DAE) was developed. In this technique, the input is randomly induced by noise. This will force the autoencoder to reconstruct the input or denoise. Denoising is recommended as a training criterion for learning to extract useful features that will constitute a better higher level representation.

Can you spot the flaw in approach? Consider what happens if the training set is skewed for some reason. The system will learn based on the inputs smoothed by statistical sanding. When the system encounters real world data, the system will, by golly, convert the “real” inputs in terms of the flawed denoising method. As one wit observed, “So s?c^2 p gives us a better estimation than the zero matrix.” Yep.

To sum up, the system just generates “drifting” outputs. The fix? Retraining. This is expensive and time consuming. Not good when the method is applied to real time flows of data.

In a more colloquial turn of phrase, the denoiser may not be denoising correctly.

A more complex numerical recipes are embedded in “smart” systems, there will be some interesting consequences. Does the phrase “chain of failure”? What about “good enough”?

Stephen E Arnold, June 23, 2019

Comments

Got something to say?





  • Archives

  • Recent Posts

  • Meta