Exercise 3: Why Forcing a CNN or RNN Onto This Capstone Would Be "Architecturally Dishonest" — Possible Solution ==================================================================== WHAT CNNs AND RNNs WERE SPECIFICALLY BUILT FOR ------------------------------ Per nn1-7, CNNs use convolution and parameter sharing across spatial position, specifically designed for data with genuine 2D spatial structure — nearby pixels are meaningfully related, and a kernel's own translation invariance is only useful when the same visual pattern can plausibly appear at different positions within a grid. Per nn1-8, RNNs maintain a hidden state carried through time, specifically designed for data where order and sequential position carry real meaning — what came before genuinely influences how a later element should be interpreted. WHAT THIS CAPSTONE'S OWN DATASET ACTUALLY LOOKS LIKE ------------------------------ Per this chapter's own scope note, "this capstone's own dataset is tabular — a genuinely different data shape from the images (nn1-7) and sequences (nn1-8/nn1-9) those architectures were specifically built for." The employee-attrition table's own columns (department, age, years_at_company, salary) have no 2D spatial relationship to each other at all — "age" isn't spatially "near" "salary" the way one pixel is near an adjacent pixel — and they have no meaningful sequential order either; reordering the columns (salary before age, for instance) changes nothing about what the data represents, unlike reordering words in a sentence. WHY APPLYING EITHER ARCHITECTURE HERE WOULD BE DISHONEST, NOT JUST UNUSUAL ------------------------------ Using a CNN would mean applying convolutional kernels designed to detect spatially-local patterns to a set of columns with no genuine spatial relationship — the "translation invariance" convolution provides would be meaningless here, since there's no 2D grid for a pattern to be "translated" across in the first place. Using an RNN would mean processing these four columns as if they were a genuine time sequence, when their actual order is arbitrary and carries no real information — the RNN's own hidden-state mechanism would be modeling a sequential relationship that doesn't actually exist in the data. In both cases, the architecture's own core assumption about the data's structure would be false for this specific dataset — not merely a non-ideal choice, but a genuine mismatch between what the model assumes and what the data actually is. WHY "DISHONEST" IS THE RIGHT WORD, NOT JUST "SUBOPTIMAL" ------------------------------ Per this chapter, forcing one of these architectures onto this dataset "here would be architecturally dishonest, not a real demonstration." The word "dishonest" specifically targets the fact that doing so would misrepresent what the resulting model is actually demonstrating — it would present a result as if it meaningfully exercised the architecture's own real strengths (spatial pattern detection, or genuine temporal dependency modeling), when in fact neither strength has anything real to act on in this specific dataset. This is a stronger, more precise critique than simply calling it a poor performance choice. WHY THIS WORKS AS AN ANSWER ------------------------------ It identifies precisely what structural assumption each architecture (CNN, RNN) depends on, explains why this capstone's own tabular dataset genuinely lacks both kinds of structure, and explains why applying either architecture anyway would misrepresent what the resulting demonstration actually shows, rather than merely being a less-than- ideal modeling choice.