Not really sure, but I can't stop thinking that, perhaps, simpler models can perform similarly, and that this is just an attempt to find a problem that fits to

Not really sure, but I can't stop thinking that, perhaps, simpler models can perform similarly, and that this is just an attempt to find a problem that fits to the existing solution. In this specific case, the problem space is so well defined that we can just spend more time on training than building more complicated models. You see, GPUs are much cheaper to roll than researchers. p.s. Meh, any wisdom here?