Data and AI Learning Group

  • 1.  AutoAI's Impact on the Data Science Workflow

    Posted Thu March 12, 2020 07:13 PM
    An emerging trend in AI is the availability of technologies in which automation is used to select a best-fit model, perform feature engineering and improve model performance via hyperparameter optimization. Automated solutions are becoming available to address each step in the Data Science workflow. Market research firms have predicted automation will help address the skills gap in Data Science and improve the productivity of Data Scientists. What are your thoughts about the capabilities of AutoAI? Where within the Data Science workflow do you expect it to have the most impact? 

    Data Science Workflow

    Dakuo Wang, Justin D. Weisz, Michael Muller, Parikshit Ram, Werner Geyer, Casey Dugan, YlaTausczik, Horst Samulowitz, and Alexander Gray. 2019. Human-AI Collaboration in Data Science: Exploring Data Scientists' Perceptions of Automated AI. Proc. ACM Hum.-Comput. Interact. 3, CSCW, Article 211 (November 2019), 24 pages. https://doi.org/10.1145/3359313


    ------------------------------
    Meredith Mante
    ------------------------------


  • 2.  RE: AutoAI's Impact on the Data Science Workflow

    Posted Fri March 13, 2020 11:26 AM
    Thank you Meredith for posting this!  With respect to the "modeling" stage, what we currently have with AutoAI Experiment on Cloud Pak for Data (and IBM Cloud)  is truly impressive. It selects the right model, applies hyperparameter optimization, ensembles, and pick the best model in terms of a metric that you can choose. That brings the question, can it be done 100% automated? I think the answer is no. There are a few things that you should know before you go into AutoAI (well, having AutoAI Experiment in mind). Basically, I think there are three basically things you should know before you can use AutoAi (Experiment) succesfully. And the nice thing, it has nothing to do with statistics or math. Unless, of course, you want to dive into hyperparameter optimization, which will take you years of study to understand what they are doing exactly. But, as said, no stats or math required to apply Auto AI successfully. But there a few other things to know before you start (still, with AutoAI Experiment in mind as a representative instance of AutoAI). Also, with the respect to the "Preparation" stage, I think it cannot be 100% automated.  Well, interesting questions, and the cliffhanger, of course, is the question: "what  three things are required to apply AutoAI in the modeling stage successfully?" :-). Once again, thank you for posting this Meredith!



    ------------------------------
    JOS den RONDEN
    ------------------------------



  • 3.  RE: AutoAI's Impact on the Data Science Workflow

    Posted Fri March 13, 2020 12:03 PM
    Thanks for your input Jos. I agree that hyperparameter optimization is certainly a step in the Data Science workflow where AutoAI can be helpful for a Data Scientist. That is a good point that you make with regards to the preparation stage too.

    ------------------------------
    Meredith Mante
    ------------------------------