иконка меню

Kaspersky
Neuromorphic
Platform

An open-source machine learning platform designed for research
in spiking neural networks (SNNs) and cognitive architectures,
as well as the development of commercial products based on them.

Advantages and Applications of neuromorphic AI

Similar to biological neural networks, SNNs are asynchronous, local, and event-driven: only neurons and synapses that receive spikes are active and consume energy.

This type of computation is most efficient when run on specialized devices: neuromorphic processors. Neuromorphic processors are based on the same biologically inspired principles as SNNs and use near-memory computing.

The combination of SNNs and neuromorphic processors enables the development of AI solutions that offer high performance while consuming significantly less energy – typically by several orders of magnitude – compared to similar AI solutions that use conventional neural networks and run on CPUs or GPUs.

This high energy efficiency and performance make neuromorphic AI ideal for the rapidly growing segment of industrial and consumer edge AI devices, including:

  • Internet of Things (IoT) sensors, detectors, and controllers;
  • Autonomous (mobile) robots, including self-learning and neuro-robots;
  • Smart home devices;
  • Wearables: medical devices, smartphones, tablets, smartwatches, fitness trackers, headphones, virtual and augmented reality glasses.

Key information about Kaspersky
Neuromorphic Platform (KNP):

Two technologies for creating spiking neural networks:
ANN2SNN technology

Conversion of deep learning artificial neural networks (ANNs) into spiking neural networks (SNNs) that involves training a neural network using the backpropagation method with weight quantization and a step activation (Heaviside) function for neurons....

ANN2SNN technology

The ANN2SNN package is used to create neural networks that can be converted into SNNs and deployed on an AltAI neuromorphic chip.

To build neural networks, special ternary layers must be used that take into account AltAI’s constraints. These layers are similar to those in the Keras library, with the following differences:

  • Input and output data of each neural network layer are binary tensors.
  • All neural network layer weights are within the {–1, 0, 1} range.
  • Bias values for neural network layers range from –32 768 to 32 767.

See the KNP Online Help the KNP Online Help for detailed technical information about the ANN2SNN package.
SNN technology

Development of topologies/architectures and training of neural networks in a spiking domain, using local learning principles based on a biologically plausible model of spike-timing-dependent plasticity (STDP)....

SNN technology

KNP implements the STDP-based local learning principle using various types of delta synapses.

A delta synapse is a synapse that sends a single synaptic input to the postsynaptic neuron each time it receives a spike from the presynaptic neuron, to change the membrane potential.

KNP supports the following synapse models for STDP-based local learning:

  • Delta synapse with spike-timing-dependent plasticity: If a spike from the presynaptic neuron arrives at the synapse before the postsynaptic neuron generates a spike, the synaptic weight increases. If a spike arrives after the postsynaptic neuron generates a spike, the synaptic weight decreases. The synaptic weight determines the synaptic input sent to the postsynaptic neuron.
  • Delta synapse with synaptic resource-dependent plasticity: The synaptic resource is determined by the specific plasticity model being used. KNP implements Hebbian, unconditional, and dopamine plasticity models.

See the KNP Online Help for technical details on KNP components for spiking neural networks.
Ability to create SNNs with flexible structural organization by minimizing hard-coded decisions in the software architecture:

1. Frontend - network infrastructure

2. Backend - network operation

3. Between them - messages

A neuron population

is a group of neurons with the same set of parameters

A synaptic projection

is a group of synapses also with the same set of parameters, that connect two neuron populations.

A message

is a unit of information exchange in an SNN model. Examples include spikes emitted by neurons and synaptic inputs generated by synapses.

Integration with the AltAI neuromorphic chip

KNP supports the AltAI neuromorphic chip as a hardware backend. AltAI is a joint development by Motiv-Neuromorphic Technologies and Kaspersky.

The first - and current - generation of the AltAI chip supports inference of spiking neural networks.

The next generation of AltAI, scheduled for release in 2026, will support both inferencing and training of SNNs directly on the chip.

Modular architecture
Main software components of KNP:
  • Python and C++ development frameworks;
  • CPU backend (training and inference);
  • AltAI neuromorphic chip backend (inference);
  • GPU backend (under development);
  • AltAI software emulator (the "golden model").

Contact

To discuss collaboration opportunities, please email us at neuro@kaspersky.com,
or use the contact form below