Gary Marcus

Exactly what you would expect in neural network with winner-take-all localist output units...

No, I disagree. Given the sampling issues I discussed (small % of stimuli and cells) it’s actually a guarantee of a distributed code.

but how would the data look different? given the sampling that we can currently do, there is no pattern of data that you would accept

If we it was hard to find neurons in higher-order areas that responded to stimuli, and when they did they only responded to one (note: cells in these studies usually responded to multiple stimuli), that would at least not falsify a localist account. The current data does.

there are many connectionist models with localist output schemes (eg a node cat, anode dog etc) w variable activity levels that are thresholded by winner take all that behave exactly like this. localism is about what nodes stand for, not whether they have firing rates or real #s

Those localist output schemes are kludgy, and NN modellers know that. A human’s output for ‘cat’ is very high D! And I never said it’s about firing rates: it’s the fact that cells respond to multiple distinct stimuli that disproves localism.

every node in alexnet etc is localist and responds to different degrees to multiple inputs. you are disregarding the whole field.

?? I'm sorry, I feel like we're speaking a different language now... AlexNet is not localist, per the definition I know. From Geoff's intro lectures (http://www.cs.toronto.edu/~bonner/courses/2014s/csc321/lectures/lec5.pdf …): "Localist architectures dedicate one neuron to each thing". Neither AlexNet nor the brain do that.