This project a work of fiction written from the perspective of a 2020 researcher traveling back in time to mid 2013 to share some 2020 xNNbased application ideas - Object Detection using Faster RCNN. The references to credit the actual inventors of the various ideas is provided at the end
- Now it's 2013 and we'd like to jump start the community with more complex vision, language, speech or games related xNN techniques from 2020; as before, we don't want anyone to know that we're from the future and got here via time machine
-
So we're going to create a presentation for conference where we've been invited to give a guest talk as if we're the inventor of a new xNN based technique
-
We choose a vision related application - object detection using Faster CNN and motivate why it's important.
-
We Show how the application is transformed into a classification and or regression problem and efficiently addressed using xNN based methods known to 2020
-
We Describe why we architected the network in the way we did (e.g., to mix strong features with well localized features, to allow mixing across all the words / features in a sentence, to handle unknown alignments between network outputs and speech labels, to handle astronomically large state spaces in games etc.
-
A demo is useful for capturing the attention and imagination of an audience, so we're going to create 1 as part of the presentation.
-
In an ideal world, we'd code from scratch everything needed by the application
-
Realistically, in a ~ 2 week time frame it's unlikely that this is going to happen, especially if training, hyper parameter optimization and limited computational resources are taken into account
-
Fortunately for us, we also brought the code and trained model from the original application designer back in time with us
-
So while we don't have to code an application from scratch, we do have to be able to run an application created by another person in TensorFlow or PyTorch to generate results