Begins to sound like a science fiction novel where a bunch of sleep researchers discover that far from being 'down time' sleep is actually a period where our neural hardware is being used by beings from another dimension to perform some nefarious calculation. Or if Douglas Adams were still alive a P2P version of Deep Thought.
Probably related. Mammals dream, and we humans experience time in dreams much faster than the actual passage of time. Dreaming is likely the same mechanism as 'experience replay' used in AI reinforcement learning. We're just training our neutral networks using minibatches of our experiences from our waking hours.
I've heard it before and it's not their wild guess but a seriously discussed idea. It's just that it's hard to set up experiments to test these sort of ideas.
I read somewhere a conjecture that sleep for human brains is analogous to regularization schemes (such as dropout) for neural networks. Dropout helps the net avoid overfitting; nets that are overfitted have a difficult time distinguishing noise from signal. Sleep for mammals might serve a similar purpose, to help the brain avoid overfitting to noise (imagine if you had difficulty distinguishing dreams from reality).
Yea, I would suspect this is true as well.
Upenn had some sleep studies that were looking this idea, I'n not sure how mature the research was/is though.
Couldn't it go the other way, too? With weak AIs, we get a sandbox of rough sketches of consciousnesses with which to experiment. Some optimisation here or there which looks uncannily like sleep may shine light on the biological version.
I think you guys are interepreting carsongross a little too literally. The bigger point is we understand very little about sleep - we understand even less about how neurons learn and how the brain functions.
My hypothesis is that all RNNs (and in general complex dynamical systems) need to be reset periodically. If run for too long without resetting, they tend to get stuck in strange states, blow up, or cease activating. You can see this effect by running a generative RNN model for a long time - eventually the output is garbage.
Under this model the next obvious question is why it takes so long to reset the brain's state. Maybe it can be done faster.