Re: [neat] Pattern Association and Consolidation Emerges from Connectivity Properties between Cortex and Hippocampus
- Jan 19 Expand MessagesView SourceCongrats Martin! It’s interesting work in a nice journal. Thanks also for highlighting the relevance to GDS. Very interesting!On Jan 14, 2014, at 6:05 AM, martin.pyka@... wrote:
I would like to share with you my recent paper that deals with hippocampus-related functions that emerge from abstract connectivity properties between cortex and hippocampus.
As I already discussed with some of you, I think that our research (I work in computational neuroscience) toward a better understanding of natural nervous systems could greatly benefit from insights and methods that have been developed in the GDS community. I hope, my paper is a ministep in order to bring both disciplines closer.
I think it wouldn't make sense to use evolutionary searches and indirect encodings directly to develop biological plausible networks with biological plausible functions. This would simply run into known problems like local-optima etc. However, what I somewhat learned through several of your papers is that
- natural encodings describe spatial global patterns which are used to place neurons with their projections cones in space which, through many side-effects (e.g. STDP, synaptic plasticity, conduction delays), result in the concrete wiring of the network which in turn implements the function
- the encoding space is highly low-dimensional compared to the phenotype space
- the relationship between encoding properties and network function is highly indirect and intuitively not understandable by humans
So what I did in this study is, I played around with some abstract connectivity properties in a network with heterogenous conduction delays and STDP and found some functions that I would not have expected. The discussion highlights a bit the connection to GDS literature:
"Our study could provide a link between form and function of biological networks that is mostly neglected in computational neuroscience. In contrast to most artificial encodings, genes in biological organisms generate gradients of protein concentration , , which in turn serve as guide for the placement of neurons and their axonal and dendritic growth cones –. The network connectivity properties so important in our model are a direct result of spatial structures , . In our opinion, the parameters of our model follow biologically plausible dimensions, such as the distance between two networks and the branching degree of neurons, that may proof useful to connect studies analyzing the impact of genes on the spatial structure of networks with the function that emerges from them.
Computational approaches along these ideas mainly focus on the spatial organization of networks involved in spatial processing, like vision , , navigation and locomotion –. The study at hand highlights that also sequential problems, like learning temporal associations, can emerge from networks designed in spatial categories. From this perspective, differing conduction delays that directly result from distances between neurons , , , could be a crucial property of neural networks to understand how episodic memories are formed."
I hope this is also inspirational for people in the GDS world.