5

Source document

[2026-03-20 00:39] xenonfun: Scalability claim is thin — you flag it in limitations but ResNet-18 on CIFAR is not a demanding test of scalability. The 5-task constraint is also mild. Future work section says the right things but a reviewer may note the gap between "scales to CNNs" in the abstract and the actual scope of experiments. Wire LM connection The structural isolation principle here is essentially the same intuition driving your wire encoding work — physical separation of representations as a substitute for learned interference management. The junction layer is doing what Kuramoto coupling does in a different domain: routing between isolated subsystems. Worth keeping that conceptual thread in mind if you're thinking about a unified theory across the two research directions. Minor Reference [3] is misattributed — that's the Synaptic Intelligence paper (Zenke et al.), not A-GEM (Chaudhry et al. 2019). A-GEM is a different Chaudhry et al. paper. Fix before submission. The reproducibility section references a W&B entity that looks like a personal account. You may want to migrate to a Monumental Systems org before making this public. Overall: the empirical story is solid, the framing is clean, and the honest LoRA comparison is a good differentiator. The task-ID oracle issue is the one thing that could seriously wound you in review if not addressed head-on. [2026-03-20 01:23] lisamegawatts: we just reran 5 seed dual net [2026-03-20 01:23] lisamegawatts: it still sucks lol [2026-03-20 01:28] lisamegawatts: (files: structural_isolation_zero_replay_catastrophic_forgetting_prevention_neurips_submission.pdf) [2026-03-20 01:28] lisamegawatts: it explicitly does mean we know the routes, it is self organizing [2026-03-20 01:28] lisamegawatts: we have labeled and unlabeled [2026-03-20 01:29] lisamegawatts: follow up paper forthcoming [2026-03-20 01:30] lisamegawatts: also we did do cnns, agem, gem, packnet, and cifar 5, 10, 50, and 100 and fashion mnist permuted and unpermuted [2026-03-20 01:32] lisamegawatts: and parameter matched and ablations all ci 95+ [2026-03-26 22:36] rolandnsharp7643: new white paper out hehe (files: Resonance_A_Neural_Network_Built_from_One_Equation.pdf) [2026-03-27 00:01] lisamegawatts: hahawe only got to 4 params [2026-03-27 00:02] foxhop.: https://unworkbench.com/docs/symmetric-embedding-workbench.pdf [2026-03-27 00:02] lisamegawatts: impressive [2026-03-27 00:02] foxhop.: new white paper & huge main contributor [2026-03-27 00:03] foxhop.: reading now! [2026-03-27 00:15] foxhop.: <@360636009433661440> do you mind if i try to reproduce? [2026-03-27 00:16] foxhop.: a single equation as a kernel is it's own operating system among new class of operating systems. [2026-03-27 00:22] rolandnsharp7643: sure. i just pushed an update. check the readme: https://github.com/rolandnsharp/resonance [2026-03-27 00:30] lisamegawatts: also you know the posers 3d stuff? i had it model the boes as graph edges and the oscillators as joints