Tree Parity Machines (neural cryptography) advantages

It seems that neural cryptography is an interesting research topic and there are some important contributions published in the last years.

The proposed algorithms are suitable for symmetric key generation starting from the common hidden weights, thus representing a possible Diffie-Hellman substitute.

As far as I understood, the main advantages are that Tree Parity Machines are quantum resistant and their hardware implementation is simpler when compared to DH.

So my question is: is this enough to switch to neural cryptography in practical use cases? Or does the scientific community consider TPMs only a trend, because of perceptrons? Or is it unsafe to deploy TPMs because more search is required in this topic? And about common public inputs: in practical implementations is it a threat or it is required to keep them secret to yield authentication?

Thanks in advance.

After like an hour of search on the subject of Tree Parity Machines as an alternative to Diffie-Hellman key exchange, I found that:

• It started with this article: Ido Kanter, Wolfgang Kinzel and Eran Kanter, Secure exchange of information by synchronization of neural networks, in Europhysics Letters (2002). It describes a new, interesting setup, which rightly attracted attention as shown by many citations, and is refreshingly honest:

  (…) we were not able to find a mathematical proof (…) that our key exchange protocol is secure.

• Shortly afterwards, that was analyzed by cryptographers, and found completely insecure: Alexander Klimov, Anton Mityagin and Adi Shamir, Analysis of Neural Cryptography, in proceedings of Asiacrypt 2002.

For followups, I recommend reading the free abstract of this paywalled 2012 survey article.

Count me as skeptical on first exposure to the whole idea, and also about the quality of some of the followup articles which cite the original but not it's refutation.