#PhysicsJournalClub
"Blending Optimal Control and Biologically Plausible Learning for Noise-Robust Physical Neural Networks"
by S. Sunada et. al.
Phys. Rev. Lett. 134, 017301 (2025)
I am not a big fan of generative AI, especially in the form it is sold and advertised today, bvut I think that there is a lot of cool and potentially useful research been done in the field of Machine Learning. In particular I am very interested in approaches where most of the heavy lifting is left to Physics, i.e. the Physical system naturally does stuff, and you exploit it for Machine learning, instead of spending a ton of resources and energy to create and manipulate the whole neural network.
In this paper the authors borrow ideas from different sub-fields to put together a training method that doesn't really need to know much about what the Physical system is actually doing, and doesn't need an accurate control of the Physical system either.
I admit I understand no more than 10% of what they say (Machine Learning is definitively not my field), but it looks interesting and promising!
https://doi.org/10.1103/PhysRevLett.134.017301
#MachineLearning #Physics
j_bertolotti /
npub1a4…razsc
2025-01-09 11:21:49
Author Public Key
npub1a4ywuu668ac6558wk53ng6nsykluuag9zpwm64nhgmjpeucc5n6skrazscPublished at
2025-01-09 11:21:49Event JSON
{
"id": "120c06b957928424f15ee9891e97c16313263e319e1d71c892e98828ff8be8b1",
"pubkey": "ed48ee735a3f71aa50eeb523346a7025bfce7505105dbd567746e41cf318a4f5",
"created_at": 1736421709,
"kind": 1,
"tags": [
[
"t",
"physics"
],
[
"t",
"machinelearning"
],
[
"t",
"physicsjournalclub"
],
[
"proxy",
"https://mathstodon.xyz/users/j_bertolotti/statuses/113798133161008003",
"activitypub"
]
],
"content": "#PhysicsJournalClub \n\"Blending Optimal Control and Biologically Plausible Learning for Noise-Robust Physical Neural Networks\"\nby S. Sunada et. al.\nPhys. Rev. Lett. 134, 017301 (2025)\n\nI am not a big fan of generative AI, especially in the form it is sold and advertised today, bvut I think that there is a lot of cool and potentially useful research been done in the field of Machine Learning. In particular I am very interested in approaches where most of the heavy lifting is left to Physics, i.e. the Physical system naturally does stuff, and you exploit it for Machine learning, instead of spending a ton of resources and energy to create and manipulate the whole neural network.\nIn this paper the authors borrow ideas from different sub-fields to put together a training method that doesn't really need to know much about what the Physical system is actually doing, and doesn't need an accurate control of the Physical system either.\nI admit I understand no more than 10% of what they say (Machine Learning is definitively not my field), but it looks interesting and promising!\nhttps://doi.org/10.1103/PhysRevLett.134.017301\n#MachineLearning #Physics",
"sig": "b54ddd2785a104d9e156b164c96c4a83b5438dd575d82e8fb5489862342c326c04276e739aa84638e0d571d56240c5eb44eea3907f513101a6fd433bcb713507"
}