Building AI: How well did we do?

Suitcase Hardware BW

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 the final post, part 7.  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. So how did our AI fan perform?

Let me put it this way: It became a captivating toy.

People could not resist playing with it, and demos easily kept users’ attention for an hour or so. People would try out various things, ask question, perform small experiments and so on. We even had a case of someone not being able to resist sticking a physical object into the fan to experiment how it will react, resulting in a broken fin and the need to purchase a replacement part.

Overall, when judged from the perspective of our initial goal to create an intelligent fan, we were quite satisfied with the result.

The ability to detect new errors worked well, too. For example, someone had an idea to place a paperclip on one of the fins to put the rotation of the fan out of balance. The fan immediately detected it and shut off.

Another interesting capability of the fan was to detect errors that we did not intentionally introduce and did not plan for. One such event happened when we made a presentation for one of the directors of DXC. Everything worked nice and smooth before the visitor came to the lab. But when he arrived, suddenly the fan started reporting errors when, to the best of our knowledge, everything should have worked just fine. Yet, the LED kept alerting us that something was wrong.

As we tried again and again, the error received a more serious status and the fan decided to shut off. Oops. How embarrassing. It appeared that our AI malfunctioned and even deteriorated over time.

However, soon it became clear what happened. The fan had an option to run from a power bank, and we played for too long without recharging the power. As a result, the fan could not reach its maximum speed. The fan was warning us about an anomaly, which took us too long to identify. The subsequent tinkering drained even more power, which made the error more severe and the machine shut off. We then connected the fan back to the power supply and everything was once again fine.

It took us some time to realize that the fan exhibited exactly the type of intelligent behavior for which we designed it, even exceeding our own expectations.

Of course, it wouldn’t be fair to write this blog if we did not also touch on the limitations of our creation. Remember, this machine operates in a real world, not within a simplified world of a computer simulation. The machine has real physical parts and has to deal with the uncertainties of real life.

The parameters of the physical world drift. The temperature and pressure of the room changes. The physical position of the fan within the suitcase changes, which then affects the vibrations of the fan. All those variations induce changes in the types of signals that should be considered normal operations. However, we did not collect sufficient training data for the anomaly detection component to cover all those variations. Data collection should have taken place over a much longer period of time and a much bigger variety of situations in order for the system to give fewer false alarms.

This is still on our to-do list.

The team

There are a number of people who contributed to that project who should be acknowledged here. Foremost, my thanks go to Zilong Zhao who was the main person behind the code implementation and the physical assembly of the hardware. Also, I would also like to thank Davor Andric for creative thinking and several excellent suggestions. In addition, a number of people supported with advice, ideas, coding and all kinds of other forms of support: Lukas Ott, Christian Kaupa, Günter Koch, David Knussmann and Jochen Thäder. Without fully dedicated assistance from the entire team, there would be no great result to blog about.

Conclusion

In place of conclusions, I want to share a list of the most important take-home messages we learned from the project:

  •  Fail fast: If a project is doomed to failure, it is better to do it quickly.
  • Acquire domain knowledge: It is difficult to create AI about X, if you do not involve human experts on X.
  • Address critical problems first: Begin solving issues that are most likely to fail.
  • Perform the necessary data science: analyzing data and than understanding it can save countless hours of trying by trials and errors.
  • Understand how various theorems apply to what you are doing: Besides the data, another side of understanding is at the deep knowledge of the basic principles underlying machine learning, AI and cybernetics.

 

RELATED LINKS

Putting machine learning into context

DIY your own AI assistant

Automating AI to make enterprises smarter, faster

 

Comments

  1. geosupergirl says:

    What next? A deep Machine learning project 🙂 I hope so. Keep the experiments coming.

    Like

  2. There is no ”AI about X”. Such an ”AI” will be in fact digitized human expertize – DHE or pseudo-intelligence – PI. An Intelligence, A or H, is personalised, meaning a personal entity, such a human person. Without a self-consciousness, there is no intelligence, no free will (animals do not have free will, they are prisoners of their instincts, that allows humans to tame them). The importance of free will is capital – its source is the consciousness that the respective entity is independent of the entire universe, even if it/he/she needs to harmonize with the universe; this decision is made entirely by it/him/herself.

    Like

Trackbacks

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

    Like

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

    Like

  3. […] 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 3. For the list of all, see here: 1, 2, 3, 4, 5, 6, 7. […]

    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: