The Rise of Self-Driving Cars and the Power of Machine Learning
A Tragic Crash in Florida
In May 2016, a Tesla in Florida hit a turning truck.
The Tesla was behind the truck.
The car did not slow down.
Its roof was torn open in the impact.
The Tesla kept moving and finally stopped after hitting another object.
The truck driver was safe.
The Tesla driver did not survive.
What Investigators Found
The autopilot was ON.
The driver trusted the system.
When the truck made a wrong turn, the autopilot failed to react in time.
This incident showed both the power and the limits of self-driving tech.
How It All Started
ALV — The First Attempt
In the 1980s, DARPA funded a project called the Autonomous Land Vehicle (ALV).
A simple system was used:
-
Green pixels = grass
-
Blue pixels = sky
-
Brown pixels = road
This helped the vehicle understand the path.
It worked, but only at 3 km/h and only on simple dirt roads.
When colours changed, the system failed.
The world was too complicated for fixed rules.
The Breakthrough: Neural Networks
ALVINN — Learning Instead of Coding
Researchers at Carnegie Mellon built ALVINN.
They tried a new idea:
Don’t tell the computer what to do.
Show it what humans do.
They collected driving videos.
The computer learned by watching.
This is machine learning.
This worked much better.
A Major Milestone
In 1995, Professor Dean Pomerleau drove across the U.S.
98% of the trip was done by the autonomous system.
Still, the system needed lots of training data.
It struggled whenever it saw something new.
The Modern Race: Tesla, Waymo & Others
Waymo’s Approach
Google launched Waymo.
They used:
-
Cameras
-
Sensors
-
3D maps
They collected tons of driving data.
In 2018, they launched a driverless taxi service.
Tesla’s Approach
Tesla chose a different path.
They used real-world video data from customer cars.
Millions of kilometres of footage helped train their system.
By the mid-2010s, Tesla rolled out self-driving features.
The Challenge
Even with huge amounts of data, errors still occur.
Some are small.
Some are surprisingly basic.
Why?
Because the real world is full of unique situations the system has never seen.
Examples of unseen scenarios
-
Odd lighting
-
New road colours
-
Strange vehicle shapes
-
Unexpected manoeuvres
-
Weather distortions
When a situation is not in the training data, the system struggles.
The 2016 Tesla crash was likely one such case.
Machine Learning Everywhere — Even in Investing
How Quants Use ML
In investing, a strategy called quant investing works the same way.
Models are trained on:
-
Price history
-
Trading volume
-
News
-
Reports
-
Sentiment
-
Unique secret data
The models look for patterns.
Then they predict what to buy and sell.
The Problem: Overfitting
If the model sees only past situations, it may fail in new ones.
This is called overfitting.
It is similar to self-driving cars failing in new scenarios.
That’s why quants use both machines and human judgement.
Why This Matters to You
You invest in markets.
The markets now include human investors and machine-driven investors.
Knowing how these systems work helps you understand how markets react.
ML models are powerful.
But they are not perfect.
Their training data limits them.
Self-Driving Tech Today
Tesla and others have made huge progress.
Self-driving cars can handle most of the driving.
They avoid many accidents humans might cause.
But they still need supervision.
And their success still depends heavily on data.
Conclusion
Machine learning changed cars, investing, and technology.
But ML systems are only as good as the data they learn from.
More data means better performance.
Less data means dangerous mistakes.
Human oversight remains essential — in cars, in markets, and in all AI-powered systems.
Comments
Post a Comment