4 ms·
The out of band pleasure/pain signal is a super interesting idea, but it's still fundamentally materialist, which is in direct contradiction with from the autho
by danans 12y ago
The out of band pleasure/pain signal is a super interesting idea, but it's still fundamentally materialist, which is in direct contradiction with from the author's (IMO silly) central claim that consciousness is a phenomenon based in some non-physical medium.
- ChuckMcM 12y agoI believe it depends on how you articulate it. The author is taking the traditional 'inner life' definition of consciousness, which is to say that entirely internal thinking is occurring, and arguing that AI based on external stimuli will never include that. The pleasure/pain internal path is basically adding additional stimuli sources within the network itself. So you can modify the training just by re-running the previous training. A made up example of how that might work is lets say you had a 'funny' neuron. Basically anytime its weight was adjusted by the network you got an immediate feedback on that association as 'funny'. Now you re-run a previous association that shares weights with the one that was tagged as 'funny' and that weight back propagates through the network giving a previously 'not associated with funny' association a proportional association. What I got out of the paper was building neural networks that were pre-disposed, though an internally hard wired linkage of some neurons, to learning a particular form of association. The idea that an identical set of training values would train differently on nets based on this sort of internally forced messaging for some neurons, is the closest thing I've seen to the 'inner life' notion of consciousness.
- danans 12y agoIt's fascinating to consider that an AI should have a mechanism of internal stimulation in addition to its external stimulus. Maybe this could be simulated by concurrently running the AI's existing models in feedback mode on internally generated stimuli, themselves sampled from those models, with some variations, but with limited ability to control the AI's actuators. The biological analogy for this might be dreaming, where our neurons seem to engage in activity (hence experiences) which are similar to but also different than our waking mental models, and does so while we are in a state of self-induced paralysis - so we don't actually jump off a bridge.