New Research isn't Good Research (sometimes)

Thursday, November 13th

Today I looked through the recently-published Duke research paper and dataset that I mentioned in the last paper, at first it looked they had a lot of valuable data that I could use, but I was mistaken. As I looked closer, the data that they had didn’t fit my needs: their data didn’t have data that was not labelled with a high level of specificity.

All datasets have:

1. The original (aka “raw”) data (the duke dataset had this)

2. The labelled data (basically what the AI model should result in after processing the original data, kind of like an answer key)

My project is about creating and training an AI model that can “divide up” a picture into “regions” representing the different object that are present in the image (including the background, with could be the sky or just blank space): this task is called Image Segmentation.

The thing with the labelled data of the Duke dataset wasn’t necessarily “wrong”, but that it didn’t identify all the separate object in the image as their own “label”, but instead put every region in the image into two labels: background and not background.

My goal for the model was to not just classify background and not background, but also to classify what each object was what.

Saturday, November 15th

When looking for mechanisms that I could add onto a publicly-available model architecture, I came across a mechanism that dealt with “attention”, which I decided was worth trying. Attention is concept in AI models that refers to the “amplification” of certain parts of an image that we value more than other than the other parts of the image.

The funny part? The research paper in which this certain attention mechanism was introduced was published in 2018, 7 years before the present. Now, I’m using the findings in this research paper over the findings of a more recent paper…

This shows that really, it’s the scope and quality of the paper or research that matters to one’s project, and not necessarily the recency of a research paper. If a recent paper is either out of scope or low quality for someone’s research, then in that case, that new research is not good research.