Why Convolution Fits Grids
Convolution fits grids because grids have local neighborhoods.
In an image, nearby pixels often form edges, corners, textures, and parts. In an audio waveform, nearby samples form local patterns. In a sequence, nearby tokens can form local phrases.
Convolution uses three assumptions:
- local patterns matter
- the same local pattern can matter in many positions
- deeper layers can combine local patterns into larger ones
These assumptions are not universal. A table of unrelated features may not have meaningful neighboring columns. In that case, convolution may be the wrong tool.
The architecture should match the structure of the data.
Enter 1 if convolution fits grids partly because nearby values can form local patterns, or 2 if it assumes all input positions are unrelated.
Compute it first, then check your number.
Enter 1 if convolution is always ideal for any table, or 2 if it may be a poor fit when neighboring columns have no meaningful relationship.
Compute it first, then check your number.