Analogical hardware is meant to be a relevant alternative to currently dominant digital hardware in AI computing systems. Analogical hardware can perform most of the operations involved in AI algorithms using less energy, less time and with a simpler architecture. But analogical operations are not accurate in the strict way digital operations are. And the effects of this are particularly clear when a real number must be compared against a threshold value. This feature can lead to a situation in which AI algorithms implemented for complex functionalities could return similar but not exactly equal valid results when applied different times to the same set of input data.
If algorithmic repeatability is a requirement for a Trustworthy AI, will analogical computing be discarded when ethics compliance is necessary?
- Anmelden, um Kommentare zu posten.
Kommentare
User account was deleted
- Anmelden, um Kommentare zu posten.
Antwort auf Well...repeatability stands von Anonymous (nicht überprüft)
User account was deleted
- Anmelden, um Kommentare zu posten.
Antwort auf actually some algorithms like von Anonymous (nicht überprüft)
Thanks, good point. The problem of algorithmic repeatability could raise not only due to the use of analogical hardware but also due to the involvement of random generation functions in the algorithm. I forgot this contingency.
- Anmelden, um Kommentare zu posten.
Antwort auf Well...repeatability stands von Anonymous (nicht überprüft)
The truth is that the document “Draft Ethics Guidelines for Trustworthy AI” doesn't develop the concept of “algorithmic repeatability”. Being the goal the trustworthy, I can imagine that concept stands for "Exactly the same result". I suppose the repeatability will be necessary in case of controversy, in order to debug the full process and find out why a specific result has been obtained. I agree that different hardware and different software may drive to different results even if using the same input dataset. For this reason the repeatability could only be guaranteed for the same hardware, the same software and the same input dataset.
In your example the result of the algorithm is a probability, it is to say, a number. In a digital computing environment using the same hardware, the same software and the same input dataset, the resulting probability should always be the same regardless of the algorithm complexity.
But you are right in that we need a definition of algorithmic repeatability before speculating about using it as a requirement.
- Anmelden, um Kommentare zu posten.
Antwort auf The truth is that the von Juan Andrés Hu…
User account was deleted
- Anmelden, um Kommentare zu posten.
Antwort auf Another concern regarding von Anonymous (nicht überprüft)
I suppose you are speaking about AI systems involving some kind of machine learning. Those AI systems evolve over time thanks to the new data they consume. Although you are right in pointing this contingency, I think that in this case the solution is pretty obvious: the system must record the software version used for each “transaction” in order to be able to exactly reproduce the operations when necessary. Anyway it is good to take it into account.
- Anmelden, um Kommentare zu posten.
User account was deleted
- Anmelden, um Kommentare zu posten.
Antwort auf Dear Colleagues - I have been von Anonymous (nicht überprüft)
User account was deleted
- Anmelden, um Kommentare zu posten.
Antwort auf From a strictly philosophical von Anonymous (nicht überprüft)
Thanks Juan
- Anmelden, um Kommentare zu posten.