M.S. AAI Capstone Chronicles 2024

Another challenge was volatility of the validation curves, which may reflect the underlying nature of the data, but could be mitigated by alternative learning rate schedules and enhanced data augmentation techniques to stabilize training and improve model convergence. Addressing these challenges would enhance this model as a practical and scalable tool to manage insulin dosing precision. The prediction error distribution between predicted carbohydrate and actual value is shown in Figure 3.4. The second method framed this problem as multi labeled classification, where each dish could contain multiple different ingredients. A dictionary was created to associate dish images with the metadata of ingredients of each dish, while target labels were binarized. The MobileNetV3Large CNN was trained to recognize the ingredients in each image, and then predict on a validation dataset (Figure 4). Model performance over epochs is depicted in Figure 5. The nutritional information of each of the predicted ingredients was obtained from a separate dictionary, and the glycemic index, and glycemic load were calculated and displayed. Results Standard accuracy was high (0.974) due to high true-negative prediction, which was likely an outcome of class imbalance. The more strict subset accuracy (where even one ingredient mismatch flags a prediction error) was 0.302. The more forgiving Hamming score, which simply performs an exclusive OR operation on the y_true and y_pred data, was 0.987, and the area under the precision-recall curve was 0.72.

14

248