LabVIEW
Open soruce code of Neural Network Model of the Cerebellum
I would like to share with the community the code of a Neural Network Model of the Cerebellum (CNN). I have been using the CNN for studying the cerebellum and for adaptive robot control. The CNN was developed using Object Oriented Programming (OOP) and a customized Address Event Representation (AER) format. Using OOP and AER allows the construction and evaluation of CNN with more than 80 k neurons, and more than 400 k synaptic connections in real time. The code includes the tools for creating the network, connecting synapses, create the AER format, and a demo for controlling a Virtual Model of a FAN.
The link to the Cerebellar Network: https://bitbucket.org/rdpinzonm/the-bicnn-model
Some details of the architecture of the cerebellar model:
In comparison with traditional ANN or RNN, the CNN has a very peculiar architecture with at least three layers (see below, Fig. 1). Inputs from the external world such as the position of the arms, legs, or sensors from a robot, are carried to the cerebellum via mossy fibers (mf). mfs are then processed in the input layer that includes Golgi (Go) and Granule cells (Gr). The ratio of Gr to mf is around 1000:1, whereas Go to Gr is 15000:1. Because of these numbers it has been proposed that the input layer of the cerebellum transform the input mfs into a sparse representation easing the work of the other layers. The second layer, the molecular layer that could be regarded as a hidden layer, includes Ba/St, Basket and Stellate cells. Their numbers are similar to Go, and their role is still a matter of debate. The last layer, the output layer, includes Purkinje cells (Pk). There are around 150.000 Gr per one Pk. This is a remarkable feature because the Pk is the only output of the cerebellar cortex. The output of the cerebellar cortex will eventualy reach the motor centers to correct movements. The CNN includes a plausible learning rule of the cerebellum at synapses between Gr and Pk. It works a an supervised anti-Hebbian rule or a anti-correlation rule in the following way: the teaching signal carrying the information about erroneous motions of the leg, arm, robot, etc, is conveyed by the climbing fiber (cf) to a single Pk. Then, the synaptic weights og Gr-Pk are decreased if there is both cf and Gr activity, whereas if there is not cf (i.e., error) the weights are increased. What this rule means, is that those Gr producing errors have their weight decreased, while those decreasing the error are promoted by increasing their weight.
Fig. 1. Neural Network Model of the Cerebellum. mf, Mossy fibers (inputs); Go, Golgi Cells; Gr, Granule cells; Ba/St, Basket and Stellate cells; Pk, Purkinje Cell (Sole output of the cerebellar cortex); cf, climbing fiber (teaching signal); pf, parallel fibers (synapses of pf-Pk are the only adjustable weights in this model, and main loci in the cerebellum); and IO, inferior olivary nucleus.
Cheers,
As you can see, the CNN has a very interesting and simple architecture with huge potential for adaptive controller. Do not hessitate in using the model, explore its code, adn post any thought, question, comment, issue. The labview project includes a demo for constructing a CNN and employ it in a classical fedback control of a DC FAN. Fig. 2-3 are some pictures of the application:
Fig 2. 3D construction of the CNN in LabVIEW representing a cube of the cerebellar cortex with edge length 100 um. Red mf, cyan Gr, green Go, yellow Ba/St, purple Pk.
Fig 3. Screen capture of the demo application in LabVIEW for the CNN used for controlling a Virtual Model of a DC FAN.
Thanks,
