Network Change and Submissions

Wow! It’s been a long time since I’ve last written about my work here, I’ve been overwhelmed with high school starting and upcoming deadlines for papers and lab demos.

Fortunately after numerous all-nighters, I was able to submit my paper for peer review in the IEEE International Conference on Robotics and Automation (ICRA) last night. It was an arduous process leading up to that moment, but I’m proud of the work I’ve submitted and look forward to writing more in the future.

While I’m waiting for my paper to be upload to arXiv, I’ll start by describing the biggest change to my work, the network architecture.

Network Change

With SqueezeNet I was struggling to find results with modal insertion, the network was actually performing worse with this additional information that should’ve been extremely useful. Although the modal insertion could’ve been applied in a proper manner in SqueezeNet, because of the size of the network it became difficult to try each possilbe insertion point to optimize the results of the insertion, within the timeframe of submission for ICRA.

For these reasons, I completely pivoted away from SqueezeNet into a custom network that our lab has been using for awhile, Z2Color. This network reflected my results perfectly. However, I wasn’t satisfied with the overall performance of the network, so I decided to add batch normalization to the network which helped it learn at a much faster rate. This was added throughout the network after each convolutional layer, which dramatically improved performance of our network.

Picture

Figure 1: Z2Color w/ Batch Normalization and Modal Insertion

This network switch allowed me to try out each insertion point for the network. As seen in Fig. 1, the point after the first batch normalization layer served as a perfect point to add this information. This point was found not only through repeated training of networks on this dataset and observation of validation loss, but also through qualitative observation of the network’s performance on-the-road.

Future Posts

Now that my work has calmed down a bit, I hope to post a bit more often, maybe once every 1-2 weeks. I’ll also be updating this site with my arXiv page once it’s submitted, so all of you can read through my work. I’ll also keep this blog updated with the progress of my ICRA submission and my future research work. Wish me luck! :pray:

Written on September 15, 2017