4 ms·

did they actually provide any evidence for this "teaching signal" in the brain? I'm sure neuroscience community would be very interested in such discovery.

did they actually provide any evidence for this "teaching signal" in the brain? I'm sure neuroscience community would be very interested in such discovery. Otherwise this seems like another model looking for the proverbial nail.

mach1ne 5y ago
Yeah, the problem with models of cognition is that networks are actually a very low-level framework to store information, meaning that both the possible methods to learn and to store are numerous. The model would have to strongly conform to existing data we have of neurology in order to be even worthy of consider to be _the_ cognitive model of the brain.
- throwaway05112 5y ago
  well.. isn't that how it supposed to work? science, i mean start with data -> build a model -> test your model on new data and if doesn't strongly conform to data, than throw it away and start again now, it doesn't mean the model has to match neurophysiological data at the very low level, e.g. neural spikes trains and all the related biochemistry and regulatory pathways and whatnot, no need to go to the rabbit hole in order to come up with a working theory at some level of abstraction. Just like Newton didn't need to go to the atom level to come up with macro view of natural laws, and express them in mathematical form and they hold value even after the special theory of relativity was discovered and generalized his findings. But he did start from actual data, may be the famous "falling apple" is just a children story, but there were experiments that he used to both come up with and validate the theory. I strongly believe that this line of research, trying to take arbitrary (but popular) ML contraptions, and somehow link them to "the brain", without any neurophysiological evidence for delta rule optimization, the bedrock of ML, it's cargo cult science par excellence. It's also not very new but falls under the old "behaviorist" branch of cognitive pseudosciences popularized by BF Skinner about a hundred years ago. Animal learning and memory doesn't rely on anything remotely similar to reinforcement learning, gradient descent, backprop and all that stuff from calculus book chapter on function minimization, simply because of energy and time constraints. This effort lacks the constraints of evidence.

mach1ne 5y ago

Yeah, the problem with models of cognition is that networks are actually a very low-level framework to store information, meaning that both the possible methods to learn and to store are numerous. The model would have to strongly conform to existing data we have of neurology in order to be even worthy of consider to be _the_ cognitive model of the brain.

throwaway05112 5y ago
well.. isn't that how it supposed to work? science, i mean start with data -> build a model -> test your model on new data and if doesn't strongly conform to data, than throw it away and start again now, it doesn't mean the model has to match neurophysiological data at the very low level, e.g. neural spikes trains and all the related biochemistry and regulatory pathways and whatnot, no need to go to the rabbit hole in order to come up with a working theory at some level of abstraction. Just like Newton didn't need to go to the atom level to come up with macro view of natural laws, and express them in mathematical form and they hold value even after the special theory of relativity was discovered and generalized his findings. But he did start from actual data, may be the famous "falling apple" is just a children story, but there were experiments that he used to both come up with and validate the theory. I strongly believe that this line of research, trying to take arbitrary (but popular) ML contraptions, and somehow link them to "the brain", without any neurophysiological evidence for delta rule optimization, the bedrock of ML, it's cargo cult science par excellence. It's also not very new but falls under the old "behaviorist" branch of cognitive pseudosciences popularized by BF Skinner about a hundred years ago. Animal learning and memory doesn't rely on anything remotely similar to reinforcement learning, gradient descent, backprop and all that stuff from calculus book chapter on function minimization, simply because of energy and time constraints. This effort lacks the constraints of evidence.