Frankly, I disagree. I think the conclusions are domain-specific.

In my field (machine learning/informatics in a medical context), it is very often the case that simple models are better as they generalize to the next set of patients better, even if a more complex model may fit the test dataset better. In some contexts, it requires a lot of hubris to think that your model is really describing reality and free from the intricacies of your particular dataset.

To generalize, complex models tend to overfit on the small datasets we typically have. One example of this is the difficulty in transferring transcriptomic models for risk prediction or other biological characteristics between clinical trials; a model with a LOOCV AUC of 0.9 can get used on a different clinical trial and perform quite poorly. Deep learning certainly has its place, but there is nothing inherently superior about a more complex model.

Of all the top Kaggle competition solutions I've looked at, very few could not be done on a laptop. I can think of only one exception and he was using one computer with 100GB+ RAM.

Doubling your data will likely only get you maybe 10% more accurate than the last increment. This seems like a lot until you realize that you went from a RMSE of 0.87567 to 0.87566. This doesn't make any material difference to your bottom line other than it takes more compute resources and more complexity in terms of coding. In this sense it may be actually costing you more to maintain a larger model and also the results may be less clear.

I would encourage ML practitioners to try the following exercise: half a problem's data and re-train your models. I think you'll find that your training will typically be much faster and the difference in the performance will be minimal.

Could not agree with your statements more. Doing statistical modelling with a small subset or too many independent variables leads to problems, which seems to be similar to what you are describing. Although I know nothing about machine learning, it would make sense that would be the case. This is off topic, but I'm an undergrad in bioinformatics, and was wondering if you have any tips into where new research may head? I will have to partner with a PI later this fall, and was wondering what kind of project I would pursue

If you put your email in your profile, I will email you.

Awesome, I have updated my profile.

Agree with the main thrust, but think that it misses a couple additional key costs: lack of reproducibility and lack of collaboration. Both of these are actually covered by the product that Domino sells, too!

Essentially, work done on laptops often doesn't make it past the laptop. Perhaps the modelling code makes it into Git, but the test data gets lost, the process that lead to it gets lost, etc. etc. Then reproducing and potentially modifying the work that has already been done becomes a labor and the work gets left behind or completely redone as a result. Let alone all the differences in environment and setup that can manifest between different laptops.

Reproducibility and provenance are essential for the data science workflows. And it's important to maintain these at scale. Take a look at Pachyderm some time as well if you get a chance (I work on the project for full disclosure). We version data in addition to code for full reproducibility. Even for distributed multi-stage and multi-language pipelines.

I agree with the article's main premise. I have a huge VPS on GCP that I can quickly spin up, make a few runs, then halt it. Cost is about $0.60/hour

The post states that the accuracy of the GLM is reported for a test set (out of sample data), but I cannot tell if the other models' accuracies are reported for a test set or training set.

Also the statement "Decreased model complexity leads to less accurate models" isn't necessarily true.

Their interface has nice fonts. I could also see their work being useful in statistics classes.

Nice but unreadable semitransparent & thin fonts.

The contrast is too low.

100% agree with this article. The cloud is where heavy computational work belongs but it's not trivial to leverage these resources today given the complexity of existing offerings. We built https://paperspace.com to make this easy and affordable. Ping me if you want to test out the platform or if you have any questions!