Inside MemScore: Model Overview

MemScore Model

MemScore uses ResNet50, a deep convolutional neural network architecture with 50 layers, specifically designed to improve image classification by addressing the vanishing gradient problem in deep networks. It achieves this through the use of residual connections, which allow the network to learn better by passing information from one layer to another more directly, bypassing some layers if needed.

The ResNet50 model is pre-trained on the ImageNet dataset, which contains millions of images across a thousand classes, so the network has learned to extract highly effective features for a wide variety of visual tasks.

Model Architecture

• Base Model: ResNet50 Model with final Fully Connected (FC) layer discarded
• Loss Function: Huber Loss; hybrid of Mean Square Error (MSE) and Mean Absolute Error (MAE) to better handle outliers
• Error Metric: Root Mean Squared Error (RMSE)
• Optimizer: Adam (Adaptive Moment Estimation); learning rate is kept low to avoid risk of overshooting