At this point in time, based solely on test loss, the best performing model was modification 1, trained on 27 crops (referred to as “modification 1.27” from now on). Though this model has the lowest test loss at 0.148, I think it is possible that some of the other models may actually perform better and may just be picking up bees that I didn’t identify in my labels, leading to a higher loss. Therefore, I’m not going to fully commit to any network yet; however, when looking at various interesting possibilities for the network, this modification 1.27 network is going to be where I start by default.

One such thing I looked at was the possibility of rotating the test cube to see if the network can identify bees in various orientations. I want to be sure that, if looking at a swarm of bees that may be in a variety of orientations, the network will be able to identify them all well, even if they are not oriented in a way that was seen in training. In order to do this, I used the same test cube (the original cube, cropped in a way that was unseen in training), but I rotated it first 90 degrees and then 180 degrees along dimensions (0,2). Though it would have been more rigorous to test this cube rotated in a variety of degrees, I started with these simple ones so as not to have to worry about feeding a crooked cube to the network/having to re-crop it. The first thing I noticed was that the model’s performance had decreased and given the original test cube, the loss had increased to 0.656; I reran this multiple times, and this loss stayed consistent (though in a different notebook, the loss remains around 0.148). Though this is something I intend to look into more at a later time, currently I think this model is still a fine way to compare how the model performs given rotated cubes.

1) 0 degrees rotated:

• Test loss = 0.656
• 
• 

2) 90 degrees rotated:

• Test loss = 0.948
• 
• 

3) 180 degrees rotated:

• Test loss = 0.709
• 
• 

Given these three different results, it appears that the network performs best given the original orientation of the cube, but the different orientations don’t completely take the model off guard. Though the loss increased, it is still less than 1 in all cases, and therefore significantly better than the original architecture, at the very least.

Additionally, I looked more into assessing the neural networks I had already trained in order to try to assess which one is the overall best. Though it’s hard to tell at this stage, I think modification 7 might be the best contender, so I sent this pre-trained network to Danielle to be applied to a larger swarm of bees to see how well it does. Then, I spent the rest of the week labeling more bees. Danielle and I decided that for the time being, trying to continue to improve the network without more training data seems less helpful than producing more training data, so until I’m done labeling the small swarm, that will be my main priority and what I spent most of my time on this week. By the end of this week, I was able to finish labeling about 240 bees.