I have heard numerous claims recently that the underparameterisation of neural networks can be implied due to the polysemanticity of its neurons, which is prevalent in LLMs.
Whilst I have no doubt that polysemanticity is the only solution to an underparameterised model, I urge on the side of caution when using polysemanticity as proof of underparametarisation.
In this note I claim that: even when sufficient capacity is available, superposition may be the default due to its overwhelming prevalence in the solution space. Disentangled, monosemantic solutions occupy a tiny fraction of the total low-loss solutions.
This suggests that superposition arises not just as a necessity in underparametarised models, but also is an inevitability of the search space of neural networks.
In this note I show a comprehensible toy example where this is the case and hypothesise that this is also the case in larger networks.
These were very rough Sunday musings so I am very interested about what other people think about this claim :).
You seem to be equating superposition and polysemanticity here, but they’re not the same thing.
Thank you for pointing this out to me! I will read further and update the blog.