Learning speed matches when inputs are uncorrelated, structure matters with correlations
A study compares convolutional and locally connected networks to see how locality and weight sharing affect learning, especially when inputs are not independent.
- Publication
- arXiv
- Stage
- Preprint
- What we read
- Summary of the paper
- Authors
- Francesco Mori, Melanie Weber
- Universities and research institutions
- Not yet supplied in verified metadata; the Brief does not guess.
What they did and found
Researchers compared two space-aware designs. When input samples are not related, both learn at about the same overall pace, but the design with shared weights gains speed proportional to the number of positions. With correlated inputs, results vary and timing shifts.
Why it matters
Choosing the right layout and scale can cut needed samples, but real data correlations can narrow the edge of weight sharing depending on data and training time.
This work compares two space-aware designs and how input layout affects learning speed and error, using a simple, clear setup.
When data are patchy with little cross-talk, shared weights can help early learning; with strong spatial links, adaptation is faster at first, but outcomes depend on training length.
What remains uncertain
Findings come from a mathematical model and specific assumptions; real-world results may differ with different data and multiple training passes.
Original sources · 1
Check the original paper for its authors, methods, version and access terms.
What is your take?
Ask a question, add useful context or share a different perspective. Keep the conversation respectful and grounded.