--- a
+++ b/R&D/046 44 Class Classification.md
@@ -0,0 +1,11 @@
+[44 Class Classification Quantum TL vs. TL Transformer vs. 2 Classical TL](https://www.chemicalqdevice.com/44-class-classification-quantum-tl-vs-tl-transformer-vs-2-classical-tl) PDF 7/4/23.
+
+Quantum Machine Learning (QML) analysis is a promising field for Healthcare Data through the use of quantum algorithms in conjunction with classical deep learning networks. Likely one of the largest qml classifications with medical images was performed based on 1) 44 Class 4400+ neuroimage dataset, 2) Quantum transfer learning model improved to accommodate classes, and 3) Three other classical models serving as standards for direct comparison. Return on Investment (ROI) will likely be experienced once the physics behind qubits are better understood in regards to processing larger datasets in machine learning models; with GPU quantum simulators likely recovering lost speed. 
+
+Quantum algorithm deep learning pair: Model A uses “frozen” trainable parameters for most of the deep learning “ResNet18” network in order to evaluate quantum algorithms, at higher performance than similar Model B. Model A/B lack performance experienced in Model C and D because ResNet18 was not allowed to improve its weights over time. The current dataset for training is likely sufficient based on Model C and D results, but performance would likely be improved with larger validation sets.
+
+The QTL model by Andrea Mari 9 is based on the classical Transfer Learning fixed feature extractor model by Sasank Chilamkurthy.10 Efforts were made to maximum performance of all four models, with the majority of the runs focusing on optimizing Model A and troubleshooting a working quantum-hybrid Model D. Previous ChemicalQDevice 10 Class classifications experienced better results when the qubit:classes 2:1 ratio was reached with 20 qubits for 10 classes of brain images.11 The current 44 Class set results would likely improve for the quantum model when 88 qubits become available. Hybrid Model D code, data extraction from ML results, and confusion matrices were assisted by OpenAI.8 
+
+To further understand quantum machine learning, related papers continue to be studied, such as Larocca, M., et al. 2021 where overparametrizing models may assist in finding highest landscape peaks to prevent training stallout.12 In addition, Peters, E., et al. 2022 reported that benign overfitting can occur and still achieve generalization.13 In addition Lloyd, S., et al. 2020 reported that quantum embeddings of images can be appropriately distributed by the surface of the qubit “sphere” for desirable classifications.14 
+
+In addition, measuring a quantum bit to yield classical bits loses a lot of information, and future methods may evolve to minimize loss of quantum benefits before relaying information back into bits. Interesting runs have also occurred by the Startup where validation data is distributed in localized ways over confusion matrices; and also an increase in performance for certain runs was observed at earlier epochs. Download full [PDF](https://www.chemicalqdevice.com/44-class-classification-quantum-tl-vs-tl-transformer-vs-2-classical-tl).