The development of neural networks in OCAS Cortiphica is carried out in a window of the module "Master of Networks" with a graphical interface. The system has created numerous convenient tools to quickly facilitate numerous connections between randomly located soma neurons. The system also uses nested sheets to model the three-dimensional volume of neural networks created. Since networks can consist of hundreds of millions of elements that cannot be accommodated in the computer's RAM, a mechanism is created for combining network parts into groups — loci. Within each locus, work is carried out independently of the rest of the network. The network hierarchy becomes recursive — loci may contain other loci. For quick access to network fragments, a mechanism for generating a visual representation of the network at different scales is implemented.
A memory trace forms in a nerve cell through a series of stages, from the appearance of electrical potentials to the formation of new protein receptors and structural reorganization as a result of the inclusion of genes in the cell nucleus. The base models of artificial neurons used in OCAS Cortiphica simulate the process of consolidation and reconsolidation of the memory trace, which includes the short, medium and long-term phases of memory formation.
Shorter memory phases serve as the basis for the formation of more long-term ones. The models of compartments (dendrites, catfish, axons) of OCAS neurons use parameters such as the frequency of action potentials, axonal facilitation, the release of a mediator, the events of a mediator-receptor connection, postsynaptic potentials and the synthesis of new receptors of a mediator during training. Important cognitive abilities of neural networks are associated with each of the phases of memory formation in artificial neurons. Editing models of neuron compartments is carried out in the module "Master of Neuron Models". The created neuron compartment models are stored in the model library as a template for use in creating various neurons.
To control the growth and development of large neural networks in OCAS Cortiphica, "cybergenomics" is used - software algorithms that ensure the formation and modification of the network structure by simulating the effects of genetic expression in living organisms. Elements of artificial neurons (dendrites, soma, axons, and also synapses) are created as sets of cybergens.
Cybergens provide for the deployment and structural changes of neural networks. Cybergens are understood to be structurally functional units of software code that control the development of a particular trait, properties, procedures, or expression of other cybergens within a cybergenetic system, such as a neurogenetic network. Cybergens determine the development, growth and functioning of a cyberorganism. There are different types of cybergens such as structural, coordinate, functiona, etc. Cybergens can be edited in the software module "Cybergen Master".
Signal events in a neural network can cause its structural reorganization, realizing the effects of epigenetic regulation.
Signal Input and Output Interfaces
The system provides input to the neural network and output from the signals of various modalities: text, sound, images and other sensory signals, including signals from external devices and drives. The input video signal can be stratified by wavelength to activate various types of artificial photoreceptors. In addition, the system has virtual cameras used within the virtual world of the system.
For the development of neural control networks for simulated cyber-physical systems in the framework of NIDE OCAS Cortiphica, a virtual world of polygons and controlled 3D-models is created.
Performance and Hardware
NIDE OCAS Cortiphica is designed as a parallel high-performance software package for supercomputers. The system is implemented in a low-level C ++ language with a set of cross-platform Qt class libraries.
As a hardware platform, nodes based on high-performance servers with Intel Xeon Processor E5-2650V3 processors and Xeon PHI co-processors (or similar) are recommended. The reason for choosing the Intel PHI platform is a combination of high-powered processors and a large amount of RAM (256 gigabytes) in the servers.
The OpenMP library is used for the efficient parallel execution of computation on systems with shared memory. The system allows you to create neural networks with 30 million artificial synapses per computer node.