Read on Twitter
It's funny how modern day AI/ML has almost nothing to do with "intelligence".
ML is one of the most popular buzzwords in tech these days. There's no magic in ML. It's all a data play.
We've been building smart data programs for decades, ML has just made it possible to make it more contextual and specific.
We've been building smart data programs for decades, ML has just made it possible to make it more contextual and specific.
A lot of the buzz is largely because people who're speaking about ML are not the ones building it.
ML infrastructure is too hard for regular users to build. We're all just plugging-n-playing. That's why it seems like magic to most of us.
ML infrastructure is too hard for regular users to build. We're all just plugging-n-playing. That's why it seems like magic to most of us.
At the fundamental level, ML models are just hyper-efficient math programs. They take data, understand the math behind it and do calculations.
Let's decompose Classification and Regression, the two most popular ML models people talk about.
Regression /1
Suppose you have a excel sheet with 3 years of previous weather data. It contains factors like sunlight, wind level, humidity.etc which influence weather.
Suppose you have a excel sheet with 3 years of previous weather data. It contains factors like sunlight, wind level, humidity.etc which influence weather.
A ML model will basically try to understand the co-relation between all these factors. It'll form a statistical formula based on its understanding.
Maybe something like :
(Sun + wind + dust / humidity x temperature) = Weather
Maybe something like :
(Sun + wind + dust / humidity x temperature) = Weather
In technical terms, it'll identify the dependant and independent variables in your data.
Their co-relation is what the model actually learns. (Machine "learning")
Their co-relation is what the model actually learns. (Machine "learning")
Once the model has learnt the statistics forming the data, it can perform calculations for us.
You can give it random values for factors like sunlight, humidity.etc and it'll give a prediction for the weather. That's simple regression.
You can give it random values for factors like sunlight, humidity.etc and it'll give a prediction for the weather. That's simple regression.
Classification /2
Suppose you have an excel sheet with 3 years of sales data for an ecommerce site. It has data vectors like time-spent, items purchased, amount.etc
Suppose you have an excel sheet with 3 years of sales data for an ecommerce site. It has data vectors like time-spent, items purchased, amount.etc
The excel sheet is divided into high-value and low-value customers (2 classes) Such data is called "labelled" data.
The model will basically go through both these labelled classes and try to find the statistics behind each customer in these classes.
By the end of it, the model will have math logic like :
if x + y + z = p, then high-value
if x + y + z = q, then low-value
By the end of it, the model will have math logic like :
if x + y + z = p, then high-value
if x + y + z = q, then low-value
Once done, you can give some random values for items purchased, amount spent, time spent.etc and the model will tell you whether the user is high-value or low-value.
This could've been done by humans, but it'd be very tedious. Besides, humans would've just calculated the value of a customer by adding their total purchases so far. They ignore other factors which subtly contribute to the overall value.
A ML model looks at all possible values and combinations before arriving to an answer. It doesn't have limitations and biases like us.
That's the only tangible ML has over humans as of now. Depth and speed.
That's the only tangible ML has over humans as of now. Depth and speed.
Important thing to note is that all of ML is highly dependant on data. Your model will be as good as the data you feed it.
PS: GPT-3 and the likes train on billions of data parameters.
PS: GPT-3 and the likes train on billions of data parameters.
Looking back, you'll see how all of this ML buzz is just math and data. There's no 'intelligence'. Sure, we have deep learning and neutral networks which take ML to the next level. But they're very heavy and intensive for most usecases as of now.
Feels weird tweeting this thread because I've never written too much about ML on my feed. Lemme know if you find it useful.
