Neuron Decay Issue in MNIST Tutorial Implementation

[qianhuiliu]

Description:
I am currently working on the MNIST tutorial https://github.com/lava-nc/lava/blob/main/tutorials/end_to_end/tutorial01_mnist_digit_classification.ipynb and attempting to develop my own model. However, I have encountered an issue related to the decay of neuron current. In the tutorial https://github.com/lava-nc/lava/blob/main/tutorials/end_to_end/tutorial00_tour_through_lava.ipynb, the LIF (Leaky Integrate-and-Fire) dynamics abstraction is presented as follows:

u[t] = u[t-1] * (1-du) + a_in # Neuron current
v[t] = v[t-1] * (1-dv) + u[t] + bias # Neuron voltage
s_out = v[t] > vth # Spike if threshold is exceeded
v[t] = 0

I made the following modifications to implement my spiking neuron:

u[t] = a_in # Neuron current
v[t] = v[t-1] * (1-dv) + u[t] + bias # Neuron voltage
s_out = v[t] > vth # Spike if threshold is exceeded
v[t] = 0

To achieve my neural dynamics, I set du = 1. However, when observing the "u" values of the first fully connected (FC) layer using the monitor in Lava, I noticed that they continue to accumulate at each timestep. This behavior is abnormal because "u" should be equal to "a_in" at each timestep. Strangely, when I set du to 4095 (the same value used in the tutorial (https://github.com/lava-nc/lava/blob/main/tutorials/end_to_end/tutorial01_mnist_digit_classification.ipynb), the "u" values of the first FC layer appear normal and remain unchanged, consistently matching "a_in" at each timestep.

Steps to Reproduce:

Follow the MNIST tutorial for implementing the LIF dynamics abstraction.
Modify the dynamics as described above to set du = 1.
Observe the "u" values of the first FC layer using a monitor.
Notice that the "u" values keep accumulating, but when du is set to 4095, the "u" values appear normal and unchanged.

Expected Behavior:
The "u" values of the first FC layer should remain stable and not accumulate regardless of the value of du when du=1.

I would greatly appreciate guidance on how to correctly set the "du" values in this context.

[Andy-Urbano]

The decay variables get multiplied by 4096 (2**12) before being used inside of the model. For a given value of du you want to implement, you should set the value in the neuron to int(du*4095). By setting this value to 1, you are actually setting the du value of the neuron to be du = 1/4096.
It's set up this way because that is the value the Loihi and the bit accurate models use to actually do the set by step iteration inside the neuron, although I agree it's a bit annoying

[qianhuiliu]

Thank you very much for your help. It quite helps! It seems the tutorial or other related documents do not mention this. I spent a lot of time debugging this problem.
Do you know is there any other special settings or considerations that I should be aware of in order to ensure the correct implementation of the model? Very appreciated!