Supported backends

Kaspersky Neuromorphic Platform supports the following types of backends:

• CPU backends:
  • Single-thread CPU backend. The backend works in one thread only. You can use a single-thread backend for debugging the platform and neural network operation. The interface to the backend is implemented in the SingleThreadedCPUBackend class.
  • Multi-thread CPU backend. The backend works in several threads simultaneously, thus increasing the speed of the neural network calculation. You can use a multi-thread backend to train and execute a spiking neural network. The interface to the backend is implemented in the MultiThreadedCPUBackend class.
    

    A neural network based on artificial neurons that perform temporal summation of input signals represented by spikes and weighted according to the weights of input connections, and generate an output pulse according to a certain threshold-based activation function.

• Backend for execution of spiking neural networks on an AltAI-1 neuromorphic processor (AltAI SNN backend):

  The interface to the backend is implemented in the AltAIBackend class. The AltAIBackend class provides the standard functions of the Backend base class.

  You can select the hardware processor of the AltAI-1 neuromorphic processor or a software emulator as the device type. The hardware processor unit uses a fixed core configuration. If selecting a software emulator as the device type, you can change the configuration settings of the AltAI-1 neuromorphic processing unit cores by using the load_core_grid_params() software emulator method, and set the location of the input and output ports in the core configuration by using the make_standart_north_port() method.

  Call load_core_grid_params method before initializing the backend.

  The figure below shows a diagram of the interaction of platform components when using the AltAIBackend backend. A spiking neural network is created with platform components. Train the neural network using the synaptic plasticity method. AltAIBackend currently supports neurons based on the AltAILIF model and synapses based on the DeltaSynapse model. Synapse weights must be integers in the range from -255 to 255.

  Training of spiking neural network is available on a central processor using the synaptic plasticity method. If needed, you can implement and use custom training methods and backends.

  The neural network is then validated against the constraints of the neuromorphic processor AltAI-1 and translated into a format that can be executed by the processing unit. You can execute a spiking neural network both on the software emulator or directly on the AltAI-1 hardware processor.

  

  Scheme of interaction between the platform and the AltAI SNN backend

• Backends for training and execution of artificial neural networks converted to spiking neural networks (ANN2SNN backends):
  

  A neural network based on artificial neurons that perform an instantaneous summation of input signals represented as numbers and weighted according to the weights of input connections, and generate an output numerical signal according to a given activation function.

  • AltAI-1 neuromorphic processor hardware backend
  • AltAI-1 neuromorphic processor software emulator backend

  The neural network is trained by backward propagation of the error with quantization of the weights.

  The figure below shows a diagram of the interaction of platform components when using AltAI ANN2SNN backends. The neural network is created using the ternary layers of the ANN2SNN package library, implemented as part of the Python framework, and is trained by back-propagating an error with quantization of weights. The neural network is then placed, using the placer utility, to be executed on the AltAI-1 neuromorphic processing unit. During placement, the neural network is validated against the constraints of the neuromorphic processing unit and translated into a format that can be executed by the processing unit.

  AltAI ANN2SNN backends are implemented in a separate package from CPU and AltAI SNN backends.

  

  Diagram of interaction between platform components when using the AltAI ANN2SNN package

Page top