Our proof-of-concept implementation was primarily aimed at demonstrating the validity of our approach for building a Cognitive Architecture, capable of learning like humans do. Because of the possible magnitude of such a scientific breakthrough, processing speed was not a concern yet. Based on our decades of experience in software development, we were obviously aware of all the possibilities for scaling software to higher processing speeds later on.

Our proof-of-concept system took about two minutes to semantically analyze a sixteen-word sentence, containing several words that were not yet in the system's 'brain'. The sixteen-word test-sentence is our benchmark value for a common trigger that needs a cognitive response. Most instructions and single-challenge requests can be formulated with sixteen words or fewer.

Because PHP is an interpreted language, mostly running on an Apache web-server, we were considering that a certain part of these two minutes were due to start-up delays. However, the start-up delays would account for maybe ten or twenty seconds, meaning that the processing would, roughly speaking, still require close to the two-minute mark.

Roughly speaking, it is not much of a benchmark, when you are trying to shave-off some percentages of this processing time.

So, when we were ready to measure the processing speed for the same sixteen-word test-sentence, we were hoping to see a 4x speed increase. If we could get the processing under thirty seconds, that would be great. Mind you, this test was still on minimal hardware (one virtualized quad-core machine with 16 Gigabytes of RAM), with no parallel processing optimizations. If we could get our 4x speed increase, that would give us confidence to make it even faster in the (near) future.

After running the test, and looking at the resulting measurement, it went silent for a moment as we couldn't believe the results. So we ran the test again and checked if we didn't make a mistake somewhere. We didn't make a mistake, the system ran the test-sentence in just under 400 milliseconds. Of course, we were hoping to see something better than just under thirty seconds, but we weren't expecting this: under 400 milliseconds.

Going from here, we have many options to increase the speed several times over, probably at least ten-fold. This means the ASTRID-system will be capable of doing real-time inference fast enough for high-speed applications like self-driving cars or even (very fast) airplanes and similar applications.

NOTE: This news-article is presented here for historical perspective only.

This article is more than two years old. Therefore, information in this article might have changed, become incomplete, or even completely invalid since its publication date. Included weblinks (if present in this article) might point to pages that no longer exist, have been moved over time, or now contain unrelated or insufficient information. No expectations or conclusions should be derived from this article or any forward-looking statements therein.