Silicon Updates: Indian scientists proposed a highly efficient Spiking Neural Networks (SNN) approach

A Neural Network (NN) is the system that “learns” to perform a task based upon the inputs given to it. Usually, Neural networks are observed in humans, in order to perform a specific task when activated. These are called Biological or Natural Neural Networks.

Illustration : iStockphoto

Scientists are working hard to recreate such “neural” system artificially, called Artificial Neural Networks (ANN). The tasks that these ANNs perform are represented in the form of mathematical functions, which are called Artificial Neurons. These artificial neurons are the basic building blocks of ANNs. Spiking Neural Networks (SNN) are another category of Artificial Neural Networks. The SNNs closely mimic the natural or biological neural networks.

The artificial neurons used in SNNs are different from that of the ones used in ANNs. The artificial neurons used in SNNs have the ability to “fire” like the biological neurons i.e., these neurons release bursts of electricity. These impulses help them to connect with other neuron surrounding them and form a neural network.

A group of researchers from Indian Institute of Bombay (IITB) are working on the third generation of these SNNs. Compared to the previous generations of SNNs, the third generation systems are far efficient in firing the impulses and forming networks. This allows more “neurons” to be placed onto a computer chip. This mimicry provides us a better approach to understand the very nature of human brain and its functionality.

Working of artificial neurons in ANNs:

As mentioned earlier, the neurons in human brain communicate with each other by transmitting electrical spikes among them. Such electrical impulses in the SNNs are produced by leaky capacitors. If a leaky capacitor reaches a threshold charge, the voltage or current “spikes” out affecting the neighbor capacitor. This phenomenon of spiking is called Quantum mechanical tunneling. A recursive spiking among the capacitors form a network, which can work together in performing a give task without any training inputs.

Contribution of research team from IITB:

The research team from IITB has modified the artificial neurons in the ANNs with silicon-based devices called Metal-Oxide Semiconductor Field-Effect Transistors (MOSFET). When compared to the traditional transistors, MOSFTEs produce more efficient tunneling. Efficient tunneling results in strong SNNs with better capabilities of learning and adapting.

Advantage of artificial neurons used in SNNs:

Replacing the conventional artificial neurons with MOSFETs adds another mode of working of these SNNs – off-current mode. This mode allows the capacitors to be 10000 times smaller than their size, when current is to be passed through them. This makes these neurons more energy efficient that the conventional neurons. Following are the words of Tanmay Chavan, a member of IITB's research team.

The use of quantum mechanical tunneling provides incredible control, which is a huge advantage. Given the fantastic performance at a unit neuron level, we plan to demonstrate networks of such neurons to understand how models of networks of neurons behave on silicon. This will enable us to understand the robustness and systems-level efficiency of the technology.

You can find the simulations to visualize SNNs here and here.