How Smart Software Works: Well, No One Is Sure It Seems
March 21, 2024
This essay is the work of a dumb dinobaby. No smart software required.
The title of this Science Daily article strikes me a slightly misleading. I thought of my asking my son when he was 14, “Where did you go this afternoon?” He would reply, “Nowhere.” I then asked, “What did you do?” He would reply, “Nothing.” Helpful, right? Now consider this essay title:
How Do Neural Networks Learn? A Mathematical Formula Explains How They Detect Relevant Patterns
AI experts are unable to explain how smart software works. Thanks, MSFT Copilot Bing. You have smart software figured out, right? What about security? Oh, I am sorry I asked.
Ah, a single formula explains pattern detection. That’s what the Science Daily title says I think.
But what does the write up about a research project at the University of San Diego say? Something slightly different I would suggest.
Consider this statements from the cited article:
“Technology has outpaced theory by a huge amount.” — Mikhail Belkin, the paper’s corresponding author and a professor at the UC San Diego Halicioglu Data Science Institute
What’s the consequence? Consider this statement:
“If you don’t understand how neural networks learn, it’s very hard to establish whether neural networks produce reliable, accurate, and appropriate responses.
How do these black box systems work? Is this the mathematical formula? Average Gradient Outer Product or AGOP. But here’s the kicker. The write up says:
The team also showed that the statistical formula they used to understand how neural networks learn, known as Average Gradient Outer Product (AGOP), could be applied to improve performance and efficiency in other types of machine learning architectures that do not include neural networks.
Net net: Coulda, woulda, shoulda does not equal understanding. Pattern detection does not answer the question of what’s happening in black box smart software. Try again, please.
Stephen E Arnold, March 21, 2024
Comments
-
- Subscribe to Beyond Search
Feature archive
News archive
-
