How to correctly dequantize the CUBA LIF neurons voltage (neuron.v) - When using neuronal voltages output layer as logits for classification and bounding box regression

[gwgknudayanga]

Hi,

I am using the output layer's neuronal voltage values to decide the bounding box coordinates of a detection algorithm(DVS camera based). When i trained and did inference in slayer, it gave me 41% map@50 accuracy. (I am using mean only batch-norm and conv layers only and I trained with delay_shift = False. All the layers are CUBA-LIF. Output layer's neurons voltage threshold are 2048. For all the other layer neurons the voltage threshold is 1.). Then i run the trained algorithm on Loihi2 (both loihi2 board and simulation). I reset the network every "32" time steps. Also i read the output layer neuron voltages with "len(network) + 1" time-step offset. But the detection performance and classification accuracies are very low in loihi board and loihi cpu simulation though in slayer it is good ( i did quantization aware training). In this case, to dequantize the output layer neuron voltages and get the fix-point values from floating points values, i divide the output layers neuron values by (32*64) as they were used in pilotnet snn. But the resulting values are different than the corresponding fix-point values that can be seen in slayer. Is this the correct way to dequantize?

• Are these huge performance drop is due to that dequantization or any other reason? Could you please help to resolve these issues and get the same results as in slayer?

Thanks and Rgds,
Udayanga