The Surprising Relativism of the Brain’s GPS
1–5 of 5 posts
Re: The Surprising Relativism of the Brain’s GPS
#2Surely our current knowledge of artificial neural networks shows that a single cell is unlikely to encode for a simple disentangled concept like absolute position, or even predicted future position?
I know artificial and real neural networks are only loosely similar, but I would be surprised if real neural networks did not also have distributed representations.
Re: The Surprising Relativism of the Brain’s GPS
#3Surely our current knowledge of artificial neural networks shows that a single cell is unlikely to encode for a simple disentangled concept like absolute position, or even predicted future position? I know artificial and real neural networks are only loosely similar, but I would be surprised if real neural networks did not also have distributed representations.
There are many areas of the brain that have very different properties from each other. This makes sense when you consider that they likely solve different (sub) problems. While a distributed representation is very useful for things like semantics clustering, it actually hurts the ability to discriminate between similar things when you need to remember a specific relationship between things. This is the case in the hippocampus where we store memories of events, as well as the structure of the world (e.g. a map of sorts as mentioned in this article). So, an artificial neural network probably isn't the best model for this.
Re: The Surprising Relativism of the Brain’s GPS
#4Surely our current knowledge of artificial neural networks shows that a single cell is unlikely to encode for a simple disentangled concept like absolute position, or even predicted future position? I know artificial and real neural networks are only loosely similar, but I would be surprised if real neural networks did not also have distributed representations.
Triangulate a volume of space. Let neurons be vertices of that graph. Every time you are at a certain point of space three or more (or four, for 3D space) neurons will be spiking depending on how close you are to them.
That gives you absolute positioning. Also the ability to increase resolution for certain volumes of space.
Re: The Surprising Relativism of the Brain’s GPS
#5Surely our current knowledge of artificial neural networks shows that a single cell is unlikely to encode for a simple disentangled concept like absolute position, or even predicted future position? I know artificial and real neural networks are only loosely similar, but I would be surprised if real neural networks did not also have distributed representations.
There are many areas of the brain that have very different properties from each other. This makes sense when you consider that they likely solve different (sub) problems. While a distributed representation is very useful for things like semantics clustering, it actually hurts the ability to discriminate between similar things when you need to remember a specific relationship between things. This is the case in the hi…
Artificial neural networks also have different regions that have very different properties from each other. They have convolutional layers and fully connected layers, and some architectures have separate semantic labelling and visual processing parts. Take a look at Mask R-CNN for example:
https://medium.com/@jonathan_hui/image-segmentation-with-mas...
I don't think that is evidence that the brain doesn't use distributed representations.