Team Sprague visited Dr. Rolf Mueller’s lab at Virginia Tech last Friday to learn more about bats and collect data! To the left is a picture of the wall of plastic leaves that resides in the Mueller lab as a part of their research on how bats deal with foliage as they hunt. We used our device to collect data from various angles and distances away from this wall. We got to see real bats, observe the lab’s data collecting equipment, and are currently awaiting a dataset from the lab that we can use in our neural networks.
Speaking of neural nets…
To the left is a subset of test outputs, predicted depth maps based on digitalized audio inputs, and squared error maps (darker means less error). To the right is the same output subset with the corresponding depth map predictions using spectrogram inputs, as well as the respective errors maps. The difference between the two neural networks’ mean squared errors was not significant enough to declare one type of audio input as The One we should move forward with in future network designs. So we will continue fine-tuning both of these neural nets. We don’t yet have criteria for qualitatively analyzing the predictions, but hope to develop a list soon as we continue observing these depth maps.
Moving forward this next week, we are aiming to collect a much larger dataset (at least 100,000 samples) using our device, which has new and improved bat ears. The dataset will contain samples taken from various locations in the ISAT building. We also want a second dataset containing samples revolving around one object in a 3D space. Discussions on improving neural network architecture to follow.