to model the known anatomy and physiology of the basal ganglia in a more detailed and faithful manner (. I am currently a PhD student at the cnrg after completing my MMath (also at the cnrg) in 2011. We can improve its performance by changing the training regime to alternate more quickly. In the next 40 trials, it then had to change that association because it was suddenly being rewarded more often for turning left instead. References Aragona,., Day,., Roitman,., Cleaveland,., Wightman,. Up and down states in striatal medium spiny neurons simultaneously recorded with spontaneous activity in fast-spiking interneurons studied in cortexstriatumsubstantia nigra. 7 this result allows us to find synaptic connection weights that will approximate any desired function, given a set of neurons with varying tuning curves.
We can send input into this model by driving current into the striatal and STN neuron populations as per. We replace each variable in the original model with a population of 40 LIF neurons with randomly selected background currents ( I bias) and gains producing a wide variety of tuning curves (as per Figure 2 ). Striatal D2 neurons project primarily to GPe, which then connects to both the STN and the GPi, forming the indirect pathway. Most importantly for our purposes, this study couples behavioral data with measurements of spiking activity in the striatum. 11, we can start with random connections between the cortex and basal ganglia, and over time the system should learn connection weights that make correct estimates of utility. Figure 6 Figure. First, it provides precise information about the values being stored in the different areas of the basal ganglia, and the computations that are needed. In Figure 8, we compare the activity seen in the rats ( Kim., 2009 ) to that seen in the model for decisions involving turning to the left. Model performance when learning is turned off after the first 80 trials. For our model, we do not consider where the reward signal r comes from; rather we directly inject the appropriate current into the ventral striatum neurons using.
This is a difficult quantity to make available to a system that operates continuously in time, as ours does. Incorporating this may come as a result of the network dynamics resulting from incorporating linking together sequentially occurring states, or it may require a more reasoned approach to detecting the time elapsed since a stimulus. As the number of neurons used increases, the error decreases (Figure 4 ).