You have to be really careful with every single level of the stack. - metrics code - batching code - data parsing/formatting/quality - data sanitation(avoid

You have to be really careful with every single level of the stack. - metrics code - batching code - data parsing/formatting/quality - data sanitation(avoiding leakage) - devising a proper dummy baseline early in the design process - sufficient patience/grit, at a personal level - data sufficiency - checking proper gradient flows - knowledge of the hardware - knowledge of each framework’s quirks - knowledge of the probability theory - knowledge of numerical stability - ability to setup a complicated stack - ability to debug both math problems and extremely complicated program stacks - data normalization - knowledge of nn architectures - knowledge of a constantly accelerating milieu of the state of the art - knowledge of which versions of which libraries/dependencies are buggy - ability to compare research papers and research code very carefully and critically - ability to make results reproducible - ability to share and communicate results - ability to prioritize beforehand which few of the many possible options you could try, are actually worth your limited time at the moment - ability to find prior art that has been overlooked by the current popular methods - ability to optimize code to run fast - ability to scale up the process to a larger distributed system when required - ability to unit test stochastic code - ability to modularize extremely bespoke spaghetti research code once its gotten ugly