Toughest part of managing ML datasets?
Hi everyone,
I'm hoping there are some ML developers here. We are building an end-to-end AI data management platform that helps ML teams annotate, manage and debug datasets at scale.
I'm interested in a discussion about what people are struggling with most when it comes to data annotation/labeling, data set management, experiment tracking, etc. What challenges are people having?
Happy to share our knowledge... let us know what you're facing and we'll try to help!
I have struggled with experiment tracking for my projects. Particularly keeping track of what dataset was used to train which model, what hyper-parameters were used, and how well each model performed. I have looked at Neptune and AWS sagemaker, but not sure how good they are. Did you have any suggestions?
Using data sets that are not exact data targets of the models design is unpredictable. Without human feedback loops the results of the model are questionable.
Show me the Wizard of Oz prototype? https://www.youtube.com/watch?v=ZgVulzUl_c4 Min 27
ML products lack real world use cases. To build a solution for the makers is a weak approach. Look for more marketable solutions. It seems ML solutions work best in a hybrid solution.
Thanks for your comments. Not sure I agree that ML products lack real-world use cases... there are a TON of good real-world use cases being built for right now. Maybe I didn't fully understand your comment.
All vitamins, never a pain killer