Comparing ML Infrastructure at a Startup Versus Big Tech

Having worked several years in both Big Tech and startups, I wanted to examine some of the differences between how Machine Learning Engineers and Data Scientists approach ML problems. At a high level, most AI practitioners are familiar with some variant of the ML lifecycle.

This usually involves building and curating your datasets, training and testing your model, and deploying and managing your model somewhere. However, while the high level process is generally agreed upon, there are massive differences in what is the best course of action is to actually get your model deployed and predicting stuff out in the wild.

To label this trove of data there are often teams of data labelers coupled with internally built applications to place guardrails on the labeling quality of these labelers. Beyond just labels, there are sophisticated models built to extract key features to enrich the data quality. There is a lot of emphasis on quality, as the business cost of a wrong prediction is much higher when you are a company like Microsoft or Google. To maintain datasets, there are often very well thought out internal tools which can track data and model versions, hyper-parameters, and metrics.

The beauty of working in big tech is that the questions you can ask are limitless. If you want to build your own question-answering model, you have the computing power to create embeddings on top of all of Quora, do the same with Reddit, and compare your results. The challenges of how you will get the data, label it, and what tools you will use to validate it are all taken care of by the processes in place and tools that have been built. As a last thought, without large companies publishing their models to the open source world, startups wouldn’t have a chance and the pace of innovation across the industry would dramatically reduce.

Activities like building an active learning framework to improve your feedback loop, using industry favorite tools like Docker and Kubernetes, and writing most of your code in python are common stories. Usually, it is good to start by looking for an open source solution for a particular problem.

Related posts:

In Spring beans are basically components which are managed by Spring IoC Container. They are objects which create a backbone of Spring application. To create bean Spring container needs Java POJO and…

I started consulting within the banking industry years ago and I’ve interacted with wealthy clients from a variety of different backgrounds. I’ve worked with founders from all industries, including…

Quiet quitting means nothing more than living. It means having a life outside of work. It means establishing healthy boundaries in your workspace & having balance. What it doesn’t mean is sacrificing…