Posts
A small story:
A junior data scientist walks up to a senior. He asks, "my model accuracy is not good; can you take a look?". The senior takes a look into the model, and sees that the model is a self attention model without pre-training; and it is trained on a 400 row datasets. This can happen if you don't have a bit of experience in machine learning and delve into deep learning too fast. But then how much knowledge in machine learning is needed to start working on deep learning?
Why are people rushing to deep learning?
Deep learning, the machine learning part established on neural networks; has started back in 1950s. After two winters in deep learning, just now, when processor speeds are at maximum and computation cost is at the lowest of all times, deep learning is booming in both academics as well as industries. From Tesla to openai, from stanford to MIT, everywhere academics are working on exciting new things, the different new architectures are now too many to even count and therefore all new students and professionals are running towards deep learning like the insects fly towards the fire.
So yes, indeed the world of deep learning is ever expanding. Each year, or at least once in two years, there is normally a break-through in deep learning field in some field. And along with that break through, comes a barrage of research papers, software, state of art architectures and many other things. Now the important thing is that if you want to stay up to date in the deep learning then you will need to read and get practiced with these.
The balance and how to find it:
Now the problem therefore in front of us is to balance between the huge task of staying updated with the deep learning and understanding good machine learning to become a valid enough data scientist.
The way to resolve this issue is to try and solve problems with appropriate technology and architecture. i.e. if you have a small dataset (<1000 rows) and small number of features; among which there are no text or something sort of a data which is a series type or sequence type data, then it is enough to use normal machine learning techniques to solve this problem instead of using a higher technology like deep learning or self-attention models.
In doing so, you will learn two things:
(1) applying proper model in proper scenario. When you will not be in a modeling only platform, and will be supposed to apply models in real life data; then you will apply the proper model in the proper settings and therefore save both computational as well as your reputation. Also, in business, unlike academic scenario, models are supposed to be explainable glass box models instead of leader board supported accuracy oriented models; whose inner workings are often not discussed.
(2) what are the basics and necessary things in a normal modeling scenario, tuning, pruning, cleaning, imputing, reporting, visualizing, feature engineering by hand and through business understanding and thousands other things which are not taught when you are trying to fit a black box model into your data to get an awesome end result.
In conclusion:
So the end advice will be it is okay to start deep learning with a basic knowledge of machine learning; but you should learn the basic nuisances of model training before getting into deep learning too hard, and also you should treat a model always like food, i.e. " you should know what you put into your mouth" because just like your body, your data and the model results are also sensitive.
Thanks for reading! I write about more technical stuffs generally and visit some of my other posts if you are interested in machine learning. Have a great day!
Comments
Post a Comment