Use Cases and AI Architecture
Identifying Your First Use Case
Before jumping into Navigator, take some time to think about the problems that you think could be solved by AI. The possibilities are limitless, so here are a few to get the creativity flowing.
-
You might have years of documentation that takes you hours to sort through to find a single answer. What if you had an expert you could chat with that was trained on these documents and would know the answer to your question in a matter of seconds, saving you hours of time?
-
You might be a safety compliance officer and you are hearing reports of some team members not wearing proper protective equipment in designated areas. What if your security cameras could alert you when a person is not wearing the proper safety equipment in that area?
-
You might have a database of images that need to be tagged with various features in order filter the images by type, color, material, etc. What if those images could automatically be tagged, ultimately saving you hours or days worth of manual image processing?
Use Case Solution Planning
Once you have identified a use case, you will want to break your use case down into incremental steps and build up from there.
Using an example from above, let's say you have a dataset of art supply images. You want to be able to easily tag these images, so you decide to train an image classifier with Navigator.
-
Start by identifying all the labels/tags you want to filter the images by. This could be very general to very specific. For example:
- Item Color: Red, Green, Blue, Black, White …
- Item Use: Drawing Supply, Painting Supply, Sculpting Supply …
- Item Type: Paint Brush, Pencil, Paint, Canvas, Clay …
-
Once your class labels have been identified, start with a broad label to start training and testing your classifier before moving to more specific detailed tagging.
-
In this example, start with the overall Item Use. Train your model on whether the image of the item is Drawing Supply, Painting Supply, Sculpting Supply.
-
Test your trained classifier with a test set of images and measure the accuracy at which it correctly labels the image based on the Item Use.
-
Make adjustments to your model or retrain if needed until the accuracy is acceptable for your use case.
-
With your classifier performing as expected for tagging the Item Use, move to training an additional classifier model that would label images by Item Type.
-
Repeat the steps outlined for the Item Use - Train, test, measure, and iterate the model until it performs as expected.
-
Before training additional classifiers, build a flow connecting your classifiers in sequence and ensure both models are still providing the correct labels.
-
Continue adding additional trained classifier models, if needed, until your images are tagged.
AI Architecture - Which one do I use?
So you have your use case identified as well as a starting point to tackle your solution, but how do you know what AI architecture to use?
Based on your use case and the data you have collected, determine if your use case is a Language (LLM) based solution, or a Computer Vision based solution.
Large Language Model (LLM) solutions
You need text based data for this solution, such as PDF's, Word Documents (.docx), or Text Documents (.txt)
LLM solutions will create an AI subject matter expert you can chat with through text prompts. This AI Expert lives locally on your computer.
Computer Vision solutions
You need image based data for this. The images can be a variety of still images, or it can be pre-recorded videos.
Computer Vision solutions create a model to analyze a series of images or video feeds to find an object in the frame or classify the entire image itself. Computer Vision solutions come in two options depending on what you want to solve for your use case:
Classifiers
A Classifier, such as Resnet, will label an entire image based on the information it is trained on. For example, if you want to identify an image of a Hot Dog vs an image that is Not Hot Dog, then your trained classifier will process the data it is given and if an image has a Hot Dog in it, then the image is labeled as HOT DOG. If an image does not have a Hot Dog in it, then the image is labeled as NOT HOT DOG.
Object Detector
An Object Detector is like a Classifier, but has the additional ability to locate the object you trained the model on displaying a bounding box around it. Let's use the previous example of Hot Dog vs Not Hot Dog. The Object Detector will review the image of the hot dog, draw a bounding box around the hot dog and label the bounding box as HOT DOG. With that, the whole image is not being labeled as HOT DOG, but the actual hot dog in the image is being found and labeled as HOT DOG. If no hot dog is found, the image would be labeled as NOT HOT DOG.