Regarding AI, I am very doubtful that it can help much, although I would love to see it happen (since it's been my primary field.)
But regardless of my pessimism, its well worth brainstorming to come up with possible approaches to try, and I'd love to hear people's ideas, even if they are just very general conceptually ... even nascent ideas may spawn something clever to try.
When we refer to AI, we really mean the sub-technique of "Machine Learning". That is the basis of all the successful AI you've hear about recently. And ML is in essence nothing more than "pattern recognition", although it can be remarkably good at it -- It can recognize deep patterns in the data that are beyond our abilities as humans to detect. But it requires first having many ... MANY!... examples with which to train it. It's not as simple as just feeding the Voynichese into some existing model (i.e. AI system) and then letting it come up with some kind of answer for what it says, or what language it's based on, or whatever.
One needs to first train the system with a lot of examples from which it can "learn" how to recognize the underlying patterns (like the encryption rules, for example.)
But the challenge is that we have only one example -- the manuscript itself. When it comes to using AI/ML in cryptology, it's what I refer to as the "single sample quandary". We need thousands or even millions of examples, depending on just how deep and complex the patterns are that we are trying to find. But an unknown cipher often involves just one case (e.g. Voynich, Zodiac cipher, etc.)
So you have to come up with some clever way to use the power of the pattern recognition that ML offers -- some framing of the problem for which you DO have a large amount of training data. (Or can generate lots of data in a way that accurately reproduces the deep patterns even though you haven't yet identified them.) Then you can use the AI/ML to solve at least a piece of the problem.
That might all sound a bit hopeless, but it isn't. Like many techniques though, it's only 10% about actually applying the technique, and 90% about framing the problem in a way that can leverage the technique.
But regardless of my pessimism, its well worth brainstorming to come up with possible approaches to try, and I'd love to hear people's ideas, even if they are just very general conceptually ... even nascent ideas may spawn something clever to try.
When we refer to AI, we really mean the sub-technique of "Machine Learning". That is the basis of all the successful AI you've hear about recently. And ML is in essence nothing more than "pattern recognition", although it can be remarkably good at it -- It can recognize deep patterns in the data that are beyond our abilities as humans to detect. But it requires first having many ... MANY!... examples with which to train it. It's not as simple as just feeding the Voynichese into some existing model (i.e. AI system) and then letting it come up with some kind of answer for what it says, or what language it's based on, or whatever.
One needs to first train the system with a lot of examples from which it can "learn" how to recognize the underlying patterns (like the encryption rules, for example.)
But the challenge is that we have only one example -- the manuscript itself. When it comes to using AI/ML in cryptology, it's what I refer to as the "single sample quandary". We need thousands or even millions of examples, depending on just how deep and complex the patterns are that we are trying to find. But an unknown cipher often involves just one case (e.g. Voynich, Zodiac cipher, etc.)
So you have to come up with some clever way to use the power of the pattern recognition that ML offers -- some framing of the problem for which you DO have a large amount of training data. (Or can generate lots of data in a way that accurately reproduces the deep patterns even though you haven't yet identified them.) Then you can use the AI/ML to solve at least a piece of the problem.
That might all sound a bit hopeless, but it isn't. Like many techniques though, it's only 10% about actually applying the technique, and 90% about framing the problem in a way that can leverage the technique.
![[+]](https://www.voynich.ninja/images/collapse_collapsed.png)