Teach Your MODELs Not To Memorise NOISE.

Design, train and tune a Machine Learning Model is very much like raising a kid through infancy, childhood and then adolescence.

Their success depends extensively on the kind of education & exposure they might be getting through their growing up period. Think about,

  • Multiple subjects, extracurricular activities & hobbies - creates Higher Variance in thinking
  • Where as, exposure to different surroundings, problems, or even travel - Lowers Bias in the child's psychology

Definitely the above is intuitively a much desirable state. High Variance & Low Bias models are much more accurate and predictable on average, they are complex mechanisms/algorithms and may take more time and investment from the guardians.

However, the critical phase is always the tuning, finding the optimal balance in the final stage. You need to carefully deal with the flexible underlying structure of the model. Making sure it do not memorise the noise instead of signals. The model doesn't overfit. The tradeoff 'should only' tilts to the positive side, otherwise there is a high cost. The Bias-Variance Trade-off is an extremely important concepts to understand for supervised machine learning and predictive modelling. It has some practical implications but due to dense mathematics it is often considered very difficult to grasp.

Happy learning!

Abir Mukherjee.

Sultan Amed

Director of AI @Contango by ADQ || xMcKinsey (QB) | xCitiBank

6 年

Congratulations and thank you for finding out time to write this excellent piece. This is quite an abstract. Conveys a lot of information but in the form of arts which makes really interesting to read and at the same time easy to understand. However, it leaves us with lots of curiosity and questions. I believe those are on our way. One, that came immediately in my mind : given the importance of the topic how often do you feel we discuss this among our analytical fraternity and from your standpoint do you believe analyst should invest time to understand the concept and from leadership standpoint what do you believe can a optimum way to increase the awareness of these things which are more physical phenomena and let us visualise the art of our work.

Sushil Pandey

Data Science | Data Engineering | Data Driven Insights

6 年

Dear Abir Sir , Analogy here is really creative. Love it

要查看或添加评论,请登录

Abir Mukherjee的更多文章

  • Arch supporter

    Arch supporter

    Today I woke up to find myself transformed into my own arch supporter. By 7:30 I had selected my beverage for the…

    2 条评论
  • Daylight to me

    Daylight to me

    When I give you the task I take it you have a clear understanding of the task, and I won't interfere with the way you…

    7 条评论
  • Predictive & Causal Relationships

    Predictive & Causal Relationships

    With the rise of Big Data predictive modeling has grown explosively in the last decade, while causal models have taken…

    1 条评论
  • Seduction of the Underdog

    Seduction of the Underdog

    Tactics, Innovation & Evolution The phenomenon to switch our allegiances for a singular contest just based on who's up…

    2 条评论
  • p-value = Indian Head Nod ?

    p-value = Indian Head Nod ?

    I like to think of p-value in statistics similar to an Indian head nod. I might expect to answer yes or no in relation…

    1 条评论
  • Empiricism vs. Rationalism : What sort of Analytical thinker are you ?

    Empiricism vs. Rationalism : What sort of Analytical thinker are you ?

    There are two broad camps of philosophies behind analytics that people fall into. And that defines how do you like to…

    3 条评论
  • Why Second Price Bidding works for Google?

    Why Second Price Bidding works for Google?

    Let me begin by clarifying my intention to write this article is solely to depict some peculiarity of this subject and…

  • Opposite Proverbs problem in analytics

    Opposite Proverbs problem in analytics

    Just like the proverbs "pen is mightier than sword" and "action speaks louder than words" we have many similar analogy…

    3 条评论

社区洞察

其他会员也浏览了