Idea for Using Hardware to Inform ML Training

[badeaa3]

Core of the idea -- use the physical chip during the training of the network or at a fine tuning step

During training:

• Push the weights to the chip, do many injections, get the DNN result from chip
• Add term into loss function that encourages network to find weights that yield correct results on chip

[badeaa3]

Model accuracy on the test dataset with no reweighting to a physical pT distribution:

• Full precision Tensorflow:
• Reduced precision QKeras: 0.7549742394504416 (From Jennet)
• C++:
• RTL:
• On chip: