An AI Model Using Another AI Model?

Thurs >> Nov 20

This week, I was training one of my main models for my science fair project, but after several hours, it seemed to plateau at ~0.83 accuracy (which is measured by the Dice Similarity Coefficient, or DSC).

So in order to improve my current model, I thought of implementing another model that could take in the “raw” data, and then pass the output of that data onto the original model that I was training. This concept, I later learned, was called a “pipeline”, and it can apply in many other AI fields too, not just in research.

Fri >> Nov 21

Today, I decided that my pipeline was going to consist of two models: one deals with “object classification” which locates and identifies objects in a 3d model using “bounding boxes”, and the second model is the original model that I had tested before.

My original model’s task is image segmentation, which means tracing the outline of a certain object or multiple objects that’s wanted from an image or 3d model.

The rationale behind this strategy was to get one model to recognize the rough “pin-drop” locations of the objects using object detection, and then the actual image segmentation model to outline the object’s shape and identify it.

This image segmentation approach was used in a few research papers that I looked into, and they’ve specifically use the YOLO architecture to accomplish object classification, so I tried using that.

Sat >> Nov 22

In “pipelines”, the output data from one model can either be automatically or manually transferred to another model. For example, when an AI email analyzer takes a set of incoming emails and then filters and sends relevant emails into a Slack channel, the pipeline does this automatically, and is called a sequential pipeline. Non-sequential pipelines, in contrast, may not automatically transfer data between models and processes automatically, and thus will require manual transferring (e.g. through copy pasting, sending files, etc). Non-sequential pipelines can also have multiple inputs and outputs, depending on the use case.

My model pipeline in this case, would be a sequential pipeline, because the output of one model is going straight into another model.

Also, during the training of the YOLO model, I “froze” the model so it could adjust to my data. When I copied over the YOLO model architecture, I also copied over its weights, which are like it’s “learnings” from a certain training session. So, to keep some of the weights intact, I froze pars of the YOLO model. “Freezing” a model some layers' weights from changing during training. It is done to keep the pre-trained knowledge while adjusting the rest of the model to new data.

This “copying” over of weights is called transfer learning.