--- a
+++ b/R&D/056 Advanced PyTorch, Keras.md
@@ -0,0 +1,11 @@
+[Advanced PyTorch, Keras Deep Learning with QML/QiML](https://drive.google.com/file/d/1haT2_R3Ghkb1QlCCRA9WNx7yYCzfj-PB/view?usp=drive_link) PDF, [Mobile Version](https://www.chemicalqdevice.com/advanced-pytorch-keras-deep-learning-with-qmlqiml) 12/7/23. 
+
+Incorporating many small quantum algorithms in parallel using PyTorch or Keras neural networks is a promising method to distribute high dimensional data from classical nodes across circuits. This is done to avoid exponential increases in RAM and Runtime associated with adding qubits to algorithms. 
+
+Although Hilbert spaces are restricted for circuits with limited numbers of qubits, their small scale makes them easier to study, and the combined quantum effects may make parallelization a viable application for larger classifications vs. less exact quantum tensor networks.
+
+In this study, a PyTorch neural network with 64 2-qubit quantum algorithms decreased model loss for 100% accuracy. Also, it was observed that having more parallel layers yielded faster processing times per layer than with less quantum algorithms. (Slide 09)
+
+In addition, a Keras framework with 8 4-qubit algorithms achieved 100% accuracy on a more difficult dataset, making it the best overall tested neural network. (Slide 12) Future studies will focus on using more circuits in parallel, different algorithm combinations, and alternative architectures.
+
+Many benchmarkings for Qiskit and PennyLane models revealed the need to fix high performance devices such as NVIDIA GPUs used with QiML libraries and PyTorch or Keras. A discussion is available through the startup video channel.