Exponential Capacity in Multilayer Hetero-Associative Neural Networks

Exponential Hopfield networks store a number of patterns that grows exponentially with the number of neurons, and in their classical formulation they are auto-associative: they complete a corrupted copy of a memory into the memory itself. Many of the tasks one wants such a network to perform are instead hetero-associative, mapping a cue to a different target. We introduce and analyse an exponential neural network of $L$ layers of $N$ binary neurons, each layer carrying its own dataset, whose energy is an exponential of the product of the per-layer Mattis overlaps, so that it is minimised precisely when every layer retrieves the pattern of the same index; the stored association must be a surjective function of the cue, and we show why nothing else can be stored at all. A cavity/signal-to-noise analysis, made exact at leading order by a large-deviation evaluation of the noise, shows that the aligned hetero-associative state is a fixed point of the zero-temperature dynamics up to a number of stored patterns $P_c\sim e^{Nρ_L}$, exponential in the layer size, with an explicit rate $ρ_L$ that grows like $L\log 2$; enlarging the basins of attraction lowers the rate but never destroys its exponential character. Comparing the theory with structured data we find that the exponential capacity and the predicted basins survive correlated, many-to-one patterns: the network is a near-perfect content-addressable memory. The same closed forms describe, without refitting, a synthetic manifold, real T-cell-receptor/epitope triples and natural-language intent data, so the mechanism is domain-universal. Generalisation to unseen cues, though significantly above chance, stays below memorisation, and it is the geometry of the encoding, rather than the data domain, that sets how far above chance it reaches. In this family, exponential storage and strong generalisation are distinct capabilities.

How Many Bits Can an Adapter Write? Measuring the Capacity and Memorization of Parameter-Efficient Fine-Tuning

A LoRA adapter is a few megabytes that almost everyone treats as a skill rather than a record of the data behind it. We put that assumption on a scale. Extending compression-based memorization analysis to the frozen-base setting, we measure directly, in bits, how much a low-rank adapter writes into a model it never changes. The answer is both smaller than full fine-tuning and less lawful than parameter counting would predict. Adapters store a couple of bits per trainable parameter, well short of a full model's budget, but that figure turns less on how many parameters an adapter carries than on where they sit. Move the same parameter budget from attention into the MLP and it holds nearly twice as much; strip the frozen base of its structure and the capacity all but disappears. Applied to realistic fine-tunes of Qwen2.5, the same instrument shows privacy leakage rising with the bits an adapter writes rather than the parameters it nominally has, and it draws a clean line between supervised and reinforcement learning: the secrets that supervised fine-tuning copies down verbatim, an adapter trained on verifiable rewards never records. Measuring what fine-tuning writes, rather than attacking it after the fact, turns a piece of folklore into a quantity one can design against.