---
title: 'Loss landscapes and terrain'
date: 2026-10-12
summary: 'Seed. The slope-reading intuition from backcountry skiing keeps wanting to be a statement about optimisation geometry, and I have not made it rigorous yet.'
status: seed
topics: [neural-networks, skiing]
canonical: 'https://yhzhan.me/notes/loss-landscapes-and-terrain/'
source:
  type: own
linked_from: [attention-is-all-you-need, sequence-and-sight-reading, terrain-as-a-first-filter]
links_to: [attention-is-all-you-need, terrain-as-a-first-filter]
---
An honest seed: the analogy is suspicious and I want to either make it precise
or discard it.

The skiing version of reading terrain is mostly about *second-order*
information — not the slope angle in front of you but how it is changing, and
where that puts you relative to a feature you cannot see. Convexity, basically,
inferred from very sparse samples.

What I do not know is whether the resemblance to curvature in optimisation is
load-bearing or merely a pun. Related: [[attention-is-all-you-need]],
[[terrain-as-a-first-filter]].