Exercise 2: The Genuine Parallel Between Pretrained GloVe and nn1-7's AlexNet — Possible Solution
====================================================================

WHAT nn1-7's AlexNet ACTUALLY LEARNED
------------------------------
Per nn1-7, AlexNet was trained on ImageNet's millions of labeled images
across many categories, and in the process learned internal features —
edge detectors, texture patterns, shape components — that turned out to
be generally useful, not narrowly specific to the exact categories
ImageNet happened to label. Those learned features could be reused as a
starting point for entirely different image tasks the network was never
originally trained on.

WHAT PRETRAINED GLOVE VECTORS LEARNED, ANALOGOUSLY
------------------------------
Per this chapter, GloVe was trained on massive general-purpose text
corpora (Wikipedia, Common Crawl) to produce word vectors that capture
broad semantic relationships — again, not narrowly specific to any one
downstream task, but generally useful representations of what words
mean and how they relate to each other, learned once from a large,
general training process.

WHY BOTH COUNT AS THE SAME KIND OF TRANSFER LEARNING
------------------------------
In both cases, a model was trained on a large, general task (classify
millions of diverse images; predict co-occurrence patterns across
billions of words) and produced internal representations useful well
beyond that original training task. Reusing AlexNet's learned features
for a new image task, and reusing GloVe's learned word vectors for a
new text task, are structurally the same move: take a representation
learned once, at scale, on a general problem, and reuse it as a head
start on a smaller, more specific problem, rather than learning
everything again from nothing.

WHY THE DIFFERENT DATA TYPES DON'T MATTER TO THE UNDERLYING PATTERN
------------------------------
Transfer learning as a general principle isn't about images or text
specifically — it's about the relationship between a large general
training process and a smaller specific one. Per this chapter, "a
representation learned once, on a huge general corpus, reused as a head
start rather than relearned from nothing" describes GloVe's case
exactly, and the identical description applies to AlexNet's own
ImageNet-trained features, just with images substituted for text. The
modality (pixels vs. words) is incidental to the pattern; what matters
is that both are instances of "don't relearn from scratch what's
already been learned well elsewhere."

WHY THIS WORKS AS AN ANSWER
------------------------------
It identifies what each model actually learned during its own large-
scale general training, shows that both produced representations
useful beyond their original training task, and explains that transfer
learning is defined by this large-general-to-small-specific reuse
pattern rather than by any particular data modality, which is why both
examples genuinely qualify as the same underlying technique.