In Deep Learning, we rely on ML Models inference that generates a probable prediction for our input. Deep Learning has been a game changer specially when it comes to solving unstructured data related tasks such as Computer Vision, Natural Language Understanding etc.
However it is also very important for us to realize the limitations of Deep Learning as the ML Models are only as good as the Training Data Set that they were built upon, the Hyperparameters and Algorithms that were used for optimizing and tuned at the time. As most Machine Learning practitioners know that building a deploying model is never a one time job. they always need updates and tuning so as to keep up with real world data.
This boils down to the fundamental idea of Algorithm based Determinism, that the idea that real world is not deterministic (my opinion) therefore Deterministic or Probabilistic Models (Map != Territory) would never* be perfectly predict for all the Real World Problem.
Hopefully this serves as a helpful reminder specifically today when Deep Learning is regarded as a Silver Bullet in the AI community.
However it is also very important for us to realize the limitations of Deep Learning as the ML Models are only as good as the Training Data Set that they were built upon, the Hyperparameters and Algorithms that were used for optimizing and tuned at the time. As most Machine Learning practitioners know that building a deploying model is never a one time job. they always need updates and tuning so as to keep up with real world data.
This boils down to the fundamental idea of Algorithm based Determinism, that the idea that real world is not deterministic (my opinion) therefore Deterministic or Probabilistic Models (Map != Territory) would never* be perfectly predict for all the Real World Problem.
Hopefully this serves as a helpful reminder specifically today when Deep Learning is regarded as a Silver Bullet in the AI community.
Comments
Post a Comment