When to use Artificial Intelligence and when to use Algorithms?

The main strength of Artificial Intelligence is it’s easy to understand by anybody. This results in new applications in all industries at a rapid pace. Are there new possibilities generated or have the possibilities always been possible? The answer is both.

In case of totally unknown input, AI is better capable of adapting. A clearly defined algorithm has much more unpredictable behavior in such cases.

If the input is known well, the tables are turned. AI could give unpredictable results and with that introducing hard-to-solve bugs. Algorithms do exactly what it is designed for, also often at a much higher performance.

Making the wrong choice here results often in a much more expensive solution.

Why algorithms are not that trusted as AI?

It’s all about trust. We trust friends, colleagues and employees that they are capable of doing new things. Trust in AI is very comparable – AI tells us a story we can believe.

Algorithms are abstract and we can only indirectly trust them – we first have to trust the mathematician and computer scientist who co-developed it. Indirect trust does not have an easy story to listen to.

Both AI and algorithms are math in the end. Difference is that AI feels more adaptable. If AI has something wrong, then we trust it will learn to fit the problem. We have different trust in adaptability of algorithms, while this may be totally untrue.

AI is faster and more precise in all cases where algorithms cannot solve the problem best.

Both AI and algorithms have different (hard) limits. Do you know these limits in your project? Did your AI-project got delayed a lot, but you kept trust in the solution?

It’s not trust we should have, but control

Trusting solutions has less value than being in control. The questions we should answer should be like:

  • Do we have control over all possible input?
  • Can boundary cases be determined and detected?
  • Do we understand the limitations of the chosen direction?

Especially the last question is to avoid the answer to be “I trust it will be solved by my smart employees”.

Discussing control forces you to fully understand the problem, before trying to solve it. The good news here is that by fully understanding the problem, you can much better beat the competition.

Combing the two

Understanding the field the software should work on, often gives solutions that are a combination of both. A simple example: do you want to teach the AI to do blurring on the image?

Like we can give better conclusions when input-data is enriched, so can the AI. This is not fixing the AI, but this is understanding limitations.

Where we use AI

We program GPUs, and with the push of AI by GPUs we automatically rolled in. I personally have a specialisation in Neural networks and Fuzzy Logic from my university, which also helped.

For example, we use AI for the “guess-work”. Say there are 5 algorithms, which to use? For the the input can be assessed by the AI, but also the best result can be picked. These are not clearly defined problems, and scoring-algorithms perform less.

And as you probably already concluded, we focus on getting full understanding of the problem before applying a solution. Get in touch to get solutions that help you outperform the competition.

 

Related Posts

Jaap van de Loosdrecht

Guest-blog: Accelerating sequential machine vision algorithms with OpenMP and OpenCL

...  my thesis “Acceleration sequential machine vision algorithms using commodity parallel hardware” is available at ...  how to ...

visitekaartje-achter-2013-V

The 13 application areas where OpenCL can be used

...  algorithms map is best to GPUs and other vector-processors? In other words: What kind of algorithms are faster when using accelerators ...

AMD-machine intelligence area

AMD gets into Machine Intelligence with “MI” range of hardware and software

...  targeting machine learning with the MI ( = Machine Intelligence) range of hardware and software.With all the news around AMD's ...

ordners1

When Big Data needs OpenCL

...  But interest for historical data is highest when new algorithms have been developed. Fully testing the algorithm's best parameters ...