Nerea Luis

Artificial Intelligence seen from the software development lifecycle perspective by Nerea Luis

Intelligent models often work very well in laboratory or "controlled" datasets but when it comes to testing them with big, real world, datasets we often suffer from some lack of maturity in the technical side of things: we need infrastructure, QA processes, automation pipelines, constant data ingestion, deploy evolutions of the model... to name a few

In this talk we will discuss MLOps (what it is, why we need multiple software development profiles...), where are the most common bottlenecks and some open-source initiatives to ease the "go to production" phase such as Gradio's Hugging Face, Streamlit or Kubeflow.

Talk Questions

  • Question 78
    Have you tried out AWS SageMaker yet? It pretty much covers all of the MLOps features you talked about. There's also options for local development so you don't have to deploy to the cloud constantly.
  • Question 72
    Do you think is mandatory to have access of real data when training a model of machine learning in beta or pre stages?
  • Question 73
    About the data , for machine Learning, is It best using a relational Database or maybe neuronal o graph databases are much better for machine learning or that shouldn't be important for machine learning?
  • Question 80
    Once a model is released into production how do you stop bad actors from polluting the model with faulty inputs?
  • Question 76
    Do you use DataDog’s AI features and how do you find them?
  • Question 77
    How do you validate the model using CT and CD to avoid side effects like over fitting?
  • Question 74
    Why do we call it Artificial Intelligence, when it can’t pass the Turing test?
  • Question 81
    CI, CD, CT; CM - continuous modeling also - make sense in practice?
  • Question 82
    How are teams organized? Is there just a single MLOps team? It seems collaboration is more important than tools in what you have explained. What do you think about this?
  • Question 84
    She talked about how ops can help ML but how can AI help Ops
  • Question 86
    What does "going to production" mean in ML/AI? What do ML/AI systems need in terms of availability, resilience compared to other non-AI software platforms and services?
  • Question 83
    What do you think about nocode/lowcode?