Changelog

Append-only log of wiki ingests.

2026-04-20

• Initialized vault structure. Created raw/ and wiki/ folders, index.md, and CHANGELOG.md.
• Ingested week 1. Created week page weeks/week-01.md and 5 concept pages: perceptron.md, dot-product.md, decision-boundary.md, loss-function.md, maximum-likelihood-estimation.md. Updated index.md.

2026-04-24

• Ingested week 2. Created week page weeks/week-02.md and 5 concept pages: gradient-descent.md, learning-rate.md, sigmoid-function.md, binary-cross-entropy.md, gradient-descent-variants.md. Updated index.md.
• Ingested week 3. Created week page weeks/week-03.md and 6 concept pages: multi-layer-perceptron.md, computation-graph.md, backpropagation.md, softmax.md, overfitting.md, regularization.md. Extended binary-cross-entropy.md with a multiclass section linking to softmax. Updated index.md.

2026-04-27 (week 9)

• Ingested week 9. Created week page weeks/week-09.md and 7 concept pages: language-model.md (the framing — $P(W)$, autoregressive vs masked, applications), n-gram-language-model.md (chain rule + Markov assumption + MLE counting; size scaling $|V{|}^N$; zeros problem; the two structural limitations — no long-distance dependencies, no synonymy generalisation — that motivate neural LMs), perplexity.md (per-word geometric-mean inverse probability; intrinsic vs extrinsic evaluation; intuition as average branching factor), decoding-strategies.md (PDF→CDF sampling primitive; greedy, beam, ancestral sampling, temperature, top-K, top-P), word-embedding.md (word2vec, GloVe, emergent linear analogy geometry, static vs contextualized; the recipe also generalises to images/multimodal — connects to representation learning), recurrent-neural-network.md (weight sharing across time, vanishing/exploding gradients, BPTT, sequential processing as the GPU bottleneck), lstm.md (cell state with additive forget/input/output gates as the residual-connection-of-recurrence; BiLSTM for ELMo). Wove the week-09 problem set into the week page (Q1 unigram drill confirming $P(\text{the})=15/60$ and $P(\text{the the the <EOS>})=0.0039$; Q2 bigram drill with greedy search demonstrating the cycle failure mode — <SOS> the cat sat on the cat sat on the — illustrating why greedy decoding gets stuck in local loops). Survey-level treatment of Transformer architectures (encoder-only BERT family, decoder-only GPT family, encoder-decoder T5/Whisper), instruction tuning, RLHF, and CLIP — these point forward to week 11’s deep dive into attention rather than getting their own concept atoms (week 11 will likely create transformer/attention/BERT/GPT atoms in depth). The week-09 l21 transcript is essentially empty (national holiday); slides + l19/l20 transcripts + Jurafsky/Martin lineage carried the lecture content. Updated index.md. Note: the cross-link from representation-learning.md to word-embedding.md would strengthen the cross-week story but was not added in this pass — could be done in a future cross-link pass.

2026-04-27 (week 8)

• Ingested week 8. Created week page weeks/week-08.md and 2 concept pages: diffusion-model.md (the comprehensive DDPM atom) and latent-diffusion-model.md (Stable Diffusion as a distinct architecture). The DDPM page covers forward/reverse processes, the closed-form forward shortcut $q(x_t\mid x_0)$ via ${\bar{\alpha }}_t$, the noise-prediction parameterisation (predict $ϵ$, recover ${\mu }_{\theta }$ analytically), training Algorithm 1 and sampling Algorithm 2 from the DDPM paper, U-Net + timestep embedding, the quality/diversity/speed trilemma, and extensions (cascaded, super-resolution, image-to-image, segmentation). The LDM page covers the encoder/decoder around the diffusion process, the speed-up math, cross-attention conditioning for text-to-image, and disambiguates “latent” between AE / GAN-prior / LDM-grid uses (cross-linked to latent-representation.md). Wove the week-08 problem set into the week page (Q1 6-statement T/F on DDPM determinism / U-Net role / latent-code confusion; Q2 mismatched-eye-colour receptive-field bug). Conditional diffusion is not a new concept atom — it’s “diffusion with $c$ added as U-Net input”; reuses the existing conditional-generative-model.md for the posterior-estimation framing. Updated index.md. Note: no Notion notes for week 8.

2026-04-26 (week 7)

• Wove 3Blue1Brown’s Bayes lesson into bayes-theorem.md. Added a leading “Steve the librarian” intuition pump (Kahneman & Tversky) before the formula — concrete-counts framing (210 people: 10 librarians, 200 farmers; 40% vs 10% match rates → $4/24\approx 16.7\%$ posterior) plus a > [!warning] COMMON MISCONCEPTION callout for base-rate neglect. Added a “geometry: restricted possibility space” section explaining Bayes as area-ratio in a $1\times 1$ square, with Grant’s verbatim “the actual math of probability is really just the math of proportions” quote in a tip box and his “rationality is not about knowing facts” line in a quote box. Added a recall question on base-rate neglect. Kept the penguin example in place (it’s the bridge to conditional generative models). The Steve example is now the intuition pump; penguin is the module application.

2026-04-26 (week 7)

• Ingested week 7. Created week page weeks/week-07.md and 4 concept pages: generative-model.md, generative-adversarial-network.md, bayes-theorem.md, conditional-generative-model.md. The GAN page consolidates discriminator/generator losses, theoretical-vs-practical generator loss with the saturating-gradient explanation, the min-max algorithm, the Jensen-Shannon proof sketch, and DCGAN/BigGAN/StyleGAN as variants. The Bayes page is foundational (will be reused for week 8 diffusion); covers prior/likelihood/posterior with the penguin walkthrough plus the MMSE-regression-converges-to-conditional-mean digression. The conditional-GM page builds on Bayes to motivate cGAN/pix2pix and explains the L1+GAN hybrid loss. Wove the week-07 problem set lessons (Q1 MSE-regression-on-digits → conditional mean is a blurry compromise, Q2 cGAN tensor sizes, Q3 GAN true/false statements) into the week page. Cross-linked $z$-disambiguation (GAN’s $z$ is a prior sample, not an encoded latent) to latent-representation.md. Updated index.md. Note: no Notion notes for week 7, so the wiki draws from slides + transcripts + problem set only.

2026-04-26 (continued)

• Ingested week 6. Created week page weeks/week-06.md and 7 concept pages: autoencoder.md, latent-representation.md, representation-learning.md, self-supervised-learning.md, contrastive-learning.md, pretext-task.md, clever-hans-effect.md. The latent-representation.md page explicitly defines “latent” and disambiguates the AE-vs-SimCLR naming clash (in AE: $z$ is kept; in SimCLR: $h$ is kept, $z$ is discarded). Cross-linked from autoencoder.md, contrastive-learning.md, representation-learning.md, and the week page.
• Wove in remaining Notion supplementary material into week 6. Added the translation/re-translation pedagogical framing to autoencoder.md (motivates why the decoder exists as a grader, not the goal). Added the detective-vs-artist intuition to contrastive-learning.md (why contrast beats reconstruction for representation quality). Added a t-SimCNE section to contrastive-learning.md covering 2-d unsupervised dataset visualization with contrastive learning (slides 72–74). Added a word-embeddings aside to representation-learning.md to anchor the cross-domain generality of the recipe. Copied 5 figures from the Notion export into images/week-06/ (autoencoder architecture+loss, identity trap, latent walk on digit-7, MNIST clustering, compression pipeline) and embedded them across the week page and autoencoder.md. Wove the week-06 problem-set Q1 (skip-connections defeat AEs / lower loss ≠ better representation) and Q2 (combining hospital datasets → Clever Hans) into autoencoder.md and clever-hans-effect.md. Updated index.md. Note: week-06 ships only slides + problem set (no lecture transcripts), so the wiki draws primarily from slides and Notion notes.

2026-04-26

• Ingested week 4. Created week page weeks/week-04.md and 6 concept pages: image-representation.md, activation-functions.md, convolution.md, pooling.md, convolutional-neural-network.md, shift-invariance-equivariance.md. Updated index.md.
• Wove in Notion supplementary material into week 4. Added geometric/decision-boundary intuition to activation-functions.md; added “Weight sharing” section, end-to-end framing, “Three hyperparameters” section, and “What is a kernel (and a filter)?” definition to convolution.md. Embedded 13 week-04 figures across convolution.md, pooling.md, and convolutional-neural-network.md.
• Ingested week 5. Created week page weeks/week-05.md and 8 concept pages: weight-initialization.md, normalization.md, data-augmentation.md, dropout.md, residual-connection.md, transfer-learning.md, upsampling.md, u-net.md. Extended learning-rate.md with a learning-rate schedule section. Cross-linked overfitting.md, regularization.md, and convolutional-neural-network.md to the new concepts; added a forward-pointing ResNet subsection to convolutional-neural-network.md. Copied 11 figures from the Notion export into images/week-05/ with descriptive names; embedded across the new concept pages. Updated index.md.
• Naming violations flagged (not auto-renamed): raw/week-05/w5-notion.zip should be w05-notion.zip (zero-padded). The Notion export folder name (raw/week-05/ExportBlock-…) is auto-generated; not in vault convention. Neither was renamed because the .zip is referenced as a source citation.