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.
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.
Data Science | Data Engineering | Data Driven Insights
6 年Dear Abir Sir , Analogy here is really creative. Love it