@peeleraja Funny how life flips the script: IIT topper stuck on UPSC, Manipal guy thriving in NYC, JNU poet in Hiranandani, fashion club girl teaching math, class clown designing Blinkit ads, doctor turned comic. Success isn't linear, it's chaotic
One friend never changes their opinion.
No matter what facts you show.
Another friend changes opinions
based on the last Instagram reel they watched.
Both are exhausting.
In machine learning, this is the bias-variance tradeoff.
High bias → model is too rigid.
High variance → model reacts to every tiny fluctuation.
One ignores reality.
The other overreacts to noise.
The goal is balance.
Not too stubborn.
Not too unstable.
Good models and good decisions
live somewhere in between.
Which one are you closer to bias or variance?
your plant isn't growing because you keep adding more water
sometimes you need to cut the dead branches first
in ML this is called Pruning
after training a neural network you check each weight and ask is this actually contributing anything?
the ones close to zero? gone.
structured pruning removes entire neurons or filters unstructured pruning removes individual weights both are trying to answer the same question
does this complexity earn its place?
a VGG-16 model has 138 million parameters pruning can cut that by 90% with almost no drop in accuracy
because most of those weights were just noise
same in life
the habits, commitments, relationships with near zero contribution they're not neutral they're dead weight slowing your forward pass
cut them
a leaner model generalizes better so does a leaner life
less parameters. sharper output. faster inference.
that's pruning
linear regression draws a line through your data.
but here's the question nobody asks at first, how does it know where the best line is?
that's exactly where the cost function comes in.
think of it like this. you're trying to park a car perfectly in a spot. every time you're off, someone measures exactly how far off you are. too left, too right, too far forward. they keep score.
that score is your cost.
technically it works like this...
the model makes a prediction. compares it to the actual value. squares the difference to penalize bigger mistakes harder. averages it across all data points.
that's Mean Squared Error. the most common cost function in linear regression.
now the model has a number. one honest number that says this is how wrong you are right now.
and the goal becomes simple.
minimize that number.
every iteration the model adjusts the line slightly. checks the cost. adjusts again. keeps going until the cost stops dropping.
that process is called gradient descent.
lower cost = better fit.
better fit = predictions you can actually trust.
the cost function isn't just a formula.
it's the only honest feedback the model ever gets.
when you first learned about cost functions did it click immediately or did gradient descent make it finally make sense?