https://github.com/Neuromorphic-Computing-X-Wearables/EnergySNNs

N3 Networks

Settings:

Weights randomly initiated, runs = 100 (mean and std are indicated in opacity spread).

Timesteps (T) = [125, 250, 500, 1000, 2000], dt = 0.001

Documentation: Number of spikes of the neuron is proportional to firing_rate*timesteps_per_inference*dt

Spiking Activity vs Input Rate (different N3 networks, with varying timesteps T)

Network: 10-256-8

Network: 20-512-16

Network: 40-1024-32

Network: 80-2046-64

Network: 160-4092-128

<aside> 💡

Increase in input poisson spikes, logarithmically increases the total spiking activity in SNNs, so input encoding sparsity will effect the power consumption in SNNs.

</aside>

<aside> 💡

The standard deviation of total spikes count increases as the number of neurons in the network increases (though the error bars looks smaller in the larger network, the scale of y-axis is large)

</aside>

Total Energy vs Spiking Activity (different N3 networks, with varying timesteps T)

Network: 10-256-8

Network: 20-512-16

Network: 40-1024-32

Network: 80-2046-64

Network: 160-4092-128

<aside> 💡

Increase in the spike activity, increases the total energy for all networks. With zero spike activity, the total energy is not zero!

</aside>