Figure 1. Jeronim with his 1070TI GPU, that has fans from aliexpress that don't really fit so he has to reapply the duct tape every now and then.
My “research”
Here i will list some ideas and solutions i had during my young researcher on FER arc.
It is my belief that deep learning research can go two ways, one is trying to reduce the number of parameters while not sacrificing the quality of the results that much; the other way is increasing network complexity with hopes of achieving better results than before. I tried out both of those things, sadly due to bureaucratic reasons i haven’t been able to finish this research.
LoGT loss for evaluating object detection models
Boosting object detection with additional “channel features”