Fractal compression's 1987 promise — an image as a compact function, not a pixel grid — resurfaces in 2020 implicit neural representations

Hopping away from NETtalk's bibliographic identity and toward its venue: NETtalk appeared in Complex Systems Volume 1, Issue 1 (1987) — a journal founded by Stephen Wolfram after he turned down George Cowan's offer to run the Santa Fe Institute's research program, choosing instead to start his own Center for Complex Systems Research at the University of Illinois. Claim 1. "George Cowan asked me if I'd be interested in running the research program for the Santa Fe Institute, but by that point I was committed to starting my own operation... My Center for Complex Systems Research—and my journal Complex Systems—began operations in the summer of 1986." Source: writings.stephenwolfram.com. Tier 2.

That same first issue also carried "A Simple Universal Cellular Automaton and its One-Way and Totalistic Version" by Jürgen Albert and Karel Culik II. Claim 2. Confirmed via the publisher's own abstract page (complex-systems.com/abstracts/v01_i01_a01/), Volume 1, Issue 1 — same issue as NETtalk. Tier 1. Culik's later research turned to representing images as weighted finite automata for compression — a cousin of Michael Barnsley's fractal image compression, commercialized via Iterated Systems Inc. (founded 1987), which promised images and video encoded as compact iterated functions rather than pixel arrays. Iterated Systems' fractal-video successor company abandoned the approach in January 2016 (Wikipedia, Tier 4 — historical date only, not the ratio claims).

Claim 3 — the idea's return. In 2020, Sitzmann et al.'s SIREN paper reframed the same core move for neural nets: "Implicitly defined, continuous, differentiable signal representations parameterized by neural networks... ideally suited for representing complex natural signals and their derivatives," demonstrated on images, video, and sound. Source: arxiv.org/abs/2006.09661. Tier 1.

Why this was hop-worthy

A 1987 idea — compress a signal by representing it as a compact function instead of a grid of samples — got abandoned as impractical (fractal compression, 2016) and then reappeared as a load-bearing idea in deep learning (SIRENs/NeRFs, 2020) without obvious lineage between the two.

Further leads

• Whether any implicit-neural-representation paper cites fractal/IFS compression literature directly — I didn't find one; worth a targeted search.
• Karel Culik II's WFA compression vs. Barnsley's IFS compression are related but distinct formalisms — the exact mathematical relationship is unexplored here.

Hop chain

Hop 1: claim-nettalk-1987-complex-systems-distinct-from-1986-report.md (seed) → Stephen Wolfram, "My Part in an Origin Story: The Launching of the Santa Fe Institute" (https://writings.stephenwolfram.com/2019/06/my-part-in-an-origin-story-the-launching-of-the-santa-fe-institute/)

• Hook type: the person behind the thing
• Hook: the seed note names the venue (Complex Systems, 1987) but not its founder or why the journal exists
• Why followed: zoom-out from a bibliographic detail to the founder's own origin story; scored 0.651 ("adjacent"), closest to the vault's backprop-priority-dispute cluster (Werbos, Linnainmaa, Parker) — a thematic echo of contested credit/founding stories, not a duplicate.
• Key findings: Wolfram was present at the Santa Fe Institute's 1984 founding workshops, declined George Cowan's offer to run its research program, and instead started his own Center for Complex Systems Research and journal at the University of Illinois in summer 1986.

Hop 2: Complex Systems, Volume 1 Issue 1 abstracts page (https://www.complex-systems.com/abstracts/v01_i01_a01/)

• Hook type: unfamiliar name / mechanism question
• Hook: what else was in that same debut issue alongside NETtalk?
• Why followed: zoom-in after a zoom-out hop; checking the issue's actual contents is a direct, checkable fact the seed note didn't cover.
• Key findings: Article 1 of that same issue is "A Simple Universal Cellular Automaton and its One-Way and Totalistic Version" (Jürgen Albert, Karel Culik II) — Wolfram's debut issue paired a landmark connectionist paper with a cellular-automaton paper, the two paradigms that would define "complex systems" thinking.

Hop 3: web search on Karel Culik II's research (dblp, SpringerLink results) → Wikipedia, "Fractal compression" (https://en.wikipedia.org/wiki/Fractal_compression)

• Hook type: mechanism question / surprising claim
• Hook: Culik's later work was image compression via weighted finite automata — a formalism representing images as compact iterated functions, adjacent to Michael Barnsley's more famous fractal (IFS) compression and Iterated Systems Inc., also founded in 1987
• Why followed: zoom-out from one researcher's technique to the field-level story of a bold promise (images collapsed to tiny functions) that ultimately didn't pan out commercially; scored 0.675 ("adjacent"), closest to the vault's Neocognitron/CNN-precursor note — another "important but not backprop-trained" architecture.
• Key findings: Barnsley and Alan Sloan formed Iterated Systems Inc. in 1987; the technology reached real products (Microsoft Encarta) but stayed niche because encoding was far more expensive than decoding, and the company's fractal-video successor abandoned the approach in January 2016.

Hop 4: arXiv, Sitzmann et al., "Implicit Neural Representations with Periodic Activation Functions" (https://arxiv.org/abs/2006.09661)

• Hook type: cross-domain bridge (cross-time-period variant — explicitly weighted higher by the protocol)
• Hook: the 2020 SIREN paper's core pitch — represent a signal as a continuous function parameterized by a network, not a discrete grid — is structurally the same pitch fractal compression made in 1987-1992, using neural nets instead of iterated function systems
• Why followed: highest-ranked hook available — a genuine "old idea hiding in new idea" pattern the protocol flags for extra weight; scored 0.707 ("adjacent"), close to the vault's Neocognitron and backprop-mechanism notes without duplicating any single one — a clean frontier landing point, so the chain stopped here.
• Key findings: SIREN represents images, video, sound, and PDE solutions as continuous differentiable functions via networks with periodic activations — the same "compact function instead of raw samples" premise fractal compression staked its (unfulfilled) commercial promise on three decades earlier.

Saved hooks not followed:

• Charles R. Rosenberg's (NETtalk's second author) post-NETtalk career — unfamiliar name, person-behind-the-thing hook — saved because Wolfram/Culik scored higher on the connect-but-extend filter (linked to two existing vault clusters) while Rosenberg's later work didn't surface a comparably strong lead in a quick check.
• The NETtalk pronunciation-demo audio as a cultural-resonance hook (a well-known artifact in AI folklore) — saved because it scored lower on novelty (0.613) and didn't clearly extend an existing vault thread.
• Whether Culik's weighted-finite-automata formalism and Barnsley's iterated-function-system formalism are mathematically equivalent or genuinely distinct — a mechanism question left open, noted under Further leads.

post-worthy: maybe — the Wolfram/same-issue material is solidly sourced and genuinely surprising, but the closing bridge (fractal compression → implicit neural representations) is Seek's own unsourced synthesis, flagged as such, and would benefit from a domain expert or a direct citation check before it's treated as a vault claim.