Characteristics of ML
'yo bro, if ML is rly THAT good, why don't we just spam ML everywhere huh???' - Nobel Peace Prize Award Recipient (certificate of participation)
“Out of every one hundred men, ten shouldn't even be there, eighty are just targets, nine are the real fighters, and we are lucky to have them, for they make the battle. Ah, but the one, one is a warrior, and he will bring the others back.” - Heraclitus
From the way I see, industries slap their products with the labels of ML for it being this trendy esoteric concept that few understand. Meanwhile, others believe plucking a few intuitive variables off their head, or pasting a best-fit line amounts to ML.
Simply no. Like with all things in the world, where dolphins belong in the sea (not zoo) and humans on land, there are areas where ML excel in and doesn't.
Characteristics of ML
This should probably be labelled as 'guidelines for a successful machine learning' scenario.They are
A pattern or trend is likely to exist
There is no mathematical proof
Lots of input data
1. Existing pattern or trend
In its simplest, if you cannot logically convince yourself that the subject matter can be predictable, it would be likely that the machine would not be able to as well.
Scenarios where a pattern or trend may not exist:
RNG (random number generator) program
Scenarios where a pattern or trend may exist:
Price of a stock
Probability of cancer
Score of a test
Despite having gazillion rows of data for a RNG program, we will unlikely be able to predict the very next number to be generated (given it is not a bad program that defeats its very purpose, but do research on 'true random' vs 'pseudo random' if you're interested).
2. No mathematical proof
The scenario that we want to predict must not be able to be derived mathematically.
For example, given the radius, we will be able to precisely and mathematically tell both the
Area of the circle, and
Circumference of the circle
Hence, there will never be a case where one attempts to predict the area of a circle given its radius.
3. Large data set
The underlying logic can be explained using statistics, and is further elaborated in the next topic. It goes by:
The larger the dataset we have (sample size),
The better the dataset is able to represent the population (Law of Large Numbers),
The better the machine is able to learn about the population with the help of this large sample
With a small data set, it is likely that we can't conclusively say anything about the population, despite having a groundbreaking hypothesis.
This is also the reason for the preference of test set error over train set error, as the model would have only truly learnt when applied to unseen data sets, whereas swimming with its own merely amounts to 'memorising'.
Last updated
