Model centric approaches are costly
Most modern neural network (NN) architectures are the product of ongoing research by specialized teams, collaborating across labs and companies. Implementing cutting-edge NN architecture could be very time consuming.
Often, the improved accuracy afforded by a complex NN model may only be warranted in a pure research environment. For commercial use cases, there are simply too many trade-offs involved.
The cost/benefit for improving models is often poor
For many use cases, the performance delta between modern models and older ones will be small after a large enough training set has been curated, and the data has been processed, cleaned, and normalized. In these cases, it’s better to use the less modern architecture, because the slightly improved accuracy offered by modern architecture won’t outweigh the drawbacks of increased cost, increased time to market, increased compute use, and decreased stability.12