Deep Learning: Making It Learnable v1.2.1 — reciprocal companion interface

Point release establishing the audited public-anchor contract and reciprocal routing with Deep Learning: Making It Trainable. The mathematical and experimental claims remain those of v1.2; all 133 frozen stdout blocks are byte-identical.

v1.2 — comprehensive audit and convention parity

Deep Learning: Making It Learnable v1.2

Version 1.2 is the comprehensive audit and convention-parity release.

• The HTML edition remains canonical; the attached PDF is its fixed 552-page print
  conversion.
• All three interludes now have independent figure namespaces, unnumbered equations,
  closing retrieval checks, and the same public authoring conventions as numbered
  chapters.
• The attention-as-test-time-regression interlude makes three memory solvers
  inspectable without overstating their relation to softmax attention or state-space
  models.
• Book-wide Plan → Code, exercise-tag, source, book-voice, frozen-output, and PDF
  text-layer contracts are enforced in CI.
• The exercise bank and evidence apparatus expand across optimization,
  generalization, sequence modeling, attention, pretraining, calibration, and
  generative modeling.
• The temperature thread now runs from kernel bandwidth, through learned similarity
  scale, to CLIP's training-time logit scale and a frozen model's post-hoc calibration.
• The epilogue owns the E. figure namespace, cites its test-time-control and
  mixture-of-experts sources directly, and leaves its control equations unnumbered.

At release, all 133 retained stdout blocks are byte-identical across HTML and TeX,
all manuscript and Python audits pass, no missing glyphs or corrupted PDF text are
detected, and every internal cross-reference resolves.

PDF SHA-256: 264263182a9d601460ed5f061aad62ee6d6933e1566fea8c26d9685dc8e4d10b

v1.1 — readable code, estimator discipline

Pedagogy: the learned-feature-space figure (ch. 3); the forget-gate diagnostic
(ch. 10); seven predict-before-run prompts; two provenance footnotes (RMSProp's
Lecture-6e lineage, weights-as-images); digit-embedding (Audit.) exercise
(ch. 13); mixture-of-experts in the epilogue's roads-not-taken.

Code: the equation/kernel/harness contract with the five-part visibility test
(style guide); canonical listings as tested source — dlbook module with Listing
4.1 (fit_supervised) and Listings 10.1/10.2, printed once via include, imported
by chs. 6/8/14 which show only their deltas; pure-harness cells folded.

Estimator discipline: "What a batch may estimate — and what it may not"
(ch. 4, three cases); case-named reminders at chs. 11/15/18/19/20; the ch. 19
"Why this batches" bridge and two verified (Audit.) traps (the β·D_x reduction
identity; the aggregate-posterior Jensen bias, β-TC-VAE cited).

Infrastructure: weekly Execution Audit (from-scratch re-execution of every
cell); include-code-files vendored; docs/compatibility.md is the living home of
version-fragile engineering (Appendix B points there).

Acceptance: every refactored chapter verified content-bit-identical against its
pre-change build. Page count 502 vs v1.0's 498: Plan-v2 changes net ≈ −1 page
(folds and dedup paid for the new sections); the +4 traces to the six
author-commissioned conceptual figures that preceded the plan's ceiling.

Deep Learning: Making It Learnable v1.0

Deep Learning: Making It Learnable — v1.0

Version 1.0 is the first stable, citable edition of Heman Shakeri's UVA DS 6050
course companion.

What this edition contains

• The complete twenty-chapter arc, from linear models through multimodal learning,
  with two bridge interludes, an epilogue, and four just-in-time appendices.
• The derivation-first test-time memory extension: attention as local-constant kernel
  regression, the KV cache as its retained dataset, and linear/delta state as distinct
  solver contracts. The epilogue closes with a carefully scoped memory-to-planning
  frontier.
• A pinned Chapter 1–9 corpus for the Chapter 10/14 language-model rematch, seeded
  CPU experiments, paired HTML/PDF freezes, and explicit evaluation disclosures.
• A consistent presentation covenant: experiments, derivations, assertions, and
  numerical audits expose their code; pure concept diagrams keep executable drawing
  source in the repository without printing pages of coordinates.
• Stable citation metadata in CITATION.cff and HTML revision notes for migrating
  pre-v1.0 page references.

Release artifact

The attached Deep-Learning--Making-It-Learnable.pdf is the verified 498-page edition.

Suggested citation: Shakeri, Heman. 2026. Deep Learning: Making It Learnable.
Version 1.0. https://shakeri-lab.github.io/dl-book/.

Text and figures are licensed CC BY-NC-SA 4.0. Code is MIT licensed; third-party
notices and asset provenance remain in the tagged source.