Tony Vladusich on Nostr: It’s incredible to me that deep convolutional neural networks and transformers all ...
It’s incredible to me that deep convolutional neural networks and transformers all fail at the outset as models of human vision: they assume translation invariance, which is absent in human vision due to space variance!
Published at
2024-03-20 18:49:16Event JSON
{
"id": "aab24aafa1def5fba6da6e24c8d62c735b60f340cbf957af88da227f859a9db0",
"pubkey": "f44c5df3773df373cda9861fe92904fb9662d1ec25385c94fe8bed045577fb82",
"created_at": 1710960556,
"kind": 1,
"tags": [
[
"proxy",
"https://mathstodon.xyz/users/TonyVladusich/statuses/112129511040027835",
"activitypub"
]
],
"content": "It’s incredible to me that deep convolutional neural networks and transformers all fail at the outset as models of human vision: they assume translation invariance, which is absent in human vision due to space variance!",
"sig": "90b2689b29626cb1022b309b08bd67cd8de04decfa97abdc76b71e5ab0cfd297d13d89697366faae49d043231f31e659d57abd2b1aa9462c976abce7f4083b20"
}
