Industry Machine Learning Project Success?
Hi All, I’ve worked in ML for some time now. ML shows great promise in mass market products (Google Search, Photos, Face Lock in phones, okish self-driving etc.) But looking at the success of ML in these products, businesses everywhere are trying to use ML for their problem, and think of it like a magic wand! I’ve worked with a few ml projects in amzn and other tech companies, that I felt did not provide as much value, the models are overfitted to align with business performance goals, but then they do not perform good in practice. I just want to understand how many of the ML projects really pan out, or am I being too optimistic when measuring my skills. Austin, 2yoe,140k ~~[Optional]~~ For Eg. take the alexa org, they started wit pretty ambitious goals to make an end-to-end system, but many portions of that had to be converted to rules. And this had to be done because Alexa is a product that is consumer facing, so there you cannot be just happy with performance in the training set and satisfying made-up business performance goals, things need to work in practice. But a large amount of internal projects are just ongoing based on some training set related goals and don't perform good in practice and don't do anyone any good :(, this makes me feel pretty sad that I'm not able to make a tangible contribution. #machine-learning #data-science
Worked on ML and provided value to the business. But the question is was ML necessary? No. Could it have been done without ML? Yes. Did ML gave more accurate results? Can't say because that's the only method we did. Why did we use ML then? Because it simply gave more credibility to the business and leadership. Also, they want to show we are using ML to solve our problems so we are cutting edge lol
Yes everything doesn’t have to use ML. But some companies do abuse the words AI and ML. There is no wrong in doing rule based systems.
Find teams solving problems that ML works for. Recommendation engineers, speech recognition, NLP, image recognition, etc.
This is not directly related to production but we use random forests to improve the interpretability and accuracy of our A/B test analysis and our counterfactual analysis as well.
I have noticed this too. Would like to hear some experienced folks' opinions on this.
So ML is a very powerful tool in modeling complex problems but just Machine Learning per se may not just work. At the End of the day the we need to solve problems and solve them at scale to improve user experiences. Business values is a huge part in determining how well a solution may fare. If something like linear regression gives you the performance and flexibility to scale and it helps your business so be it. With so much experimentation, we should be ready to let our babies go and move on.