Building AI: Classifying errors

Error

Editor’s note: This is a series of blog posts on the topic of “Demystifying the creation of intelligent machines: How does one create AI?” You are now reading part 5. For the list of all, see here: 1, 2, 3, 4, 5, 6, 7.

As I discussed in my last posts, I have been working with colleagues at DXC to build an artificially intelligent fan, one that can monitor its operations, report issues and even, sometimes, fix problems itself.

Classifying errors is the second component necessary to build an intelligent fan. In this process, we introduce errors while running the fan, then record the data and train a machine learning system to identify those errors. That way, we have errors labeled.

This is a somewhat easier problem than detecting general anomalies. In our autoassociative network, we used as inputs the entire power spectrum and the machine needed to detect an error without ever being trained on what errors look like. But the second component uses classified known errors and was trained to use only a selected part of the spectrum.

The first question that may pop in one’s mind is, “Why not classify errors as the first component? Why not skip anomaly detection and go directly to classification of errors?”

Learning to detect any (and all) errors

There are at least two good reasons for introducing two stages. One is that we want our AI to detect any errors, not just the ones that we know about and can simulate. By all likelihood, the type of errors we simulated present only a small fraction of all possible errors that could occur.

Second, and equally importantly, we want to minimize false positives while making it relatively easy for errors to be classified. A classifier has to decide whether certain inputs belong to a category. For that, the classifier has to judge against all other possible categories it has been trained for.

Let’s say we have three types of errors: Type 1, Type 2 and Type 3. If a classifier is trained to identify Type 1, this classifier has to be able to distinguish the features against Type 2 and Type 3. It has to be trained to minimize the confusion matrix between those.

I have described in a previous post that existing AI technologies, such as deep learning, do not scale well when the number of categories that needs to be distinguished increases. When a category is added, more resources need to be added to keep the performance at the same level.

We do not want this fundamental limitation to affect our fan’s key decision of whether there is an error or not. This decision is more important than any other decision related to the classification of the error. Hence, we want to dedicate full machine learning power to that error-or-not conclusion. This choice is necessary to minimize both false positives and misses.

Another issue is that of imbalanced sizes of training sets. It is easier to collect data for normal operations than for errors. In addition, the harder the problem, the more data we need. So, if we train everything in bulk, we would need to be collecting much larger amounts of data, and this could get complicated from a practical viewpoint.

Therefore, we have solved at least two problems with the architecture we decided on: We are able to detect unknown errors and make a reliable decision on whether the fan experiences malfunction in its operations.

 

RELATED LINKS

Is the fusion of AI and cloud the future?

The merging of man and machine: Is your workplace prepared?

A primer on personal AI assistants

Trackbacks

  1. […] Editor’s note: This is a series of blog posts on the topic of “Demystifying the creation of intelligent machines: How does one create AI?” You are now reading part 6.  Previous posts: 1, 2, 3, 4, 5 […]

    Like

  2. […] How does one create AI?” You are now reading part 1. For the list of all, see here: 1, 2, 3, 4, 5, 6, […]

    Like

  3. […] How does one create AI?” You are now reading part 2. For the list of all, see here: 1, 2, 3, 4, 5, 6, […]

    Like

  4. […] How does one create AI?” You are now reading part 3. For the list of all, see here: 1, 2, 3, 4, 5, 6, […]

    Like

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: