Measuring Impact and Updating AI Model for a banking/financial services AI product use case
As a product manager, I am constantly looking to improve machine learning models and products by applying proven strategies to mitigate bias, scale a product, and continuously update models.
The problem the capstone project is trying to solve in the banking/financial services industry is to reduce processing time it takes to enroll business clients for Cash Management Services, and make the service accessible and available to clients in a faster time frame.
Capstone Project focused on:
• Traditional Machine Learning approach (which works better with smaller datasets) instead of Deep Learning approach (which outperforms with large datasets)
• Scanned images as input data
Some key points to consider when training a model, is to train it with all types of data that the model is likely to encounter in the future in order to build a robust model; and while there is no clear-cut rule as to how much data is required, it is generally advised to start with a few hundred examples of each target class and then scale up the amount of training data until a desired accuracy is reached’.
On the otherhand if this project was executed through Deep Learning approach, the number of parameters in any Convolutional Neural Network (CNN) are very large and without sufficient amount of training data, the model will not learn anything; so the model may need up to 5,000 images to get a relatively solid confidence level and accuracy.