What made the difference.
BEFORE
Machine learning required experts to decide which features mattered.
For image recognition, someone specified what to measure. That was slow and limited.
WHAT CHANGED
Deep networks learn the features themselves from raw data.
Nobody tells the system what an edge is. It discovers that edges are useful.
WHAT MADE IT POSSIBLE
Substantially more data, from the internet Substantially more computation, particularly graphics processors Better training techniques
Those three arriving together produced the change.
WHAT IT ENABLED
Image recognition at and beyond human level Usable speech recognition Machine translation that reads naturally Language models that write coherently
WHAT IT DID NOT SOLVE
Reasoning reliably Knowing what it does not know Working from few examples Explaining itself
THE PATTERN
Deep learning is exceptional at perception and pattern-matching, and weaker at reasoning.
That shapes where it helps and where it fails.
