29 Oct 2024 - Technical Papers

We've written a technical paper! It's been a long time since I've written research. I've been published six times in my life, specifically for construction-related research. This experience has been hugely beneficial to my desire to share knowledge and learn more.

Our paper is about a novel approach to AI model training. This should help others to improve their model training results. It's mutually beneficial.

We have submitted the paper to a conference in Q1. More on that if it comes together as we hope.

Here is the summary of the research. Ill post it soon, but for now we are reviewing and validating the results and additional research on this topic.

Model within a Model: a computer vision training approach using hierarchical learning seeks to determine the optimal methodology for training-related data sets in a computer vision model. While this paper focused on training content for a computer vision model to read and understand architectural construction documentation, this approach could be used on many other applications. This paper compares two training approaches: 1) all-at-once and 2) the Model-within-a-Model (Mw/M), incorporating Bayesian statistical thinking. In the end, when we trained data all-at-once, immediate Precision (P) and Recall (R) scores were high, but they did not improve when more data was added. Compared to the Mw/M approach, which saw continuous improvements as more data was added to the training pipeline. This paper will further explain this process and how others could adopt this approach.

Keywords: Construction, 2D, Computer Vision, Vision Language Model, Data Training