A single threshold knob sets how much a hysteretic network remembers

What the work claims

This is a nonlinear-dynamics theory paper with a modest applied demonstration. The object of study is the discrete-time hysteresis neural network, a recurrent network whose neurons are binary elements with hysteresis: their output state depends not only on the current input but on the direction from which the input arrived, characterized by a single threshold parameter.¹ Depending on that parameter, the network can hold many coexisting binary fixed points, the standing candidates for associative memory.

The paper's contribution is to characterize not the fixed points themselves but their basins of attraction: the sets of initial network states that flow to each fixed point. Basin size is what a memory actually gets in practice, because retrieval is flow into a basin, and the authors quantify the distribution of basin sizes with its normalized entropy. Their central claim: the threshold parameter controls that distribution, and in particular can maximize the entropy, pushing the capacity split toward uniform across all stored patterns.¹

To make the analysis tractable they pose a simple problem: classification of a binary dataset, where fixed points play the role of class centers and basins play the role of classes. The demonstration dataset is synthetic, 200 binary response vectors over 20 items generated from a Gaussian model, and four representative vectors are chosen as class centers; the classification is then evaluated with two standard quantities from item response theory, the discrimination and ability parameters of a two-parameter logistic model.¹

How it works

Hysteresis gives each neuron a memory of its own recent history: the input must cross one threshold to switch up and a different, lower one to switch back, so brief noise does not flip the state. Wire many such elements into a recurrent loop and the network's dynamics become a map on binary vectors with multiple attracting fixed points, each a stored pattern. The basin of attraction of a fixed point is the region of state space from which the network converges to it; in memory terms, it is the set of corrupted or partial patterns that pattern can still retrieve.

The paper evaluates basins by brute-force enumeration over initial states, feasible only because the model is tiny (the authors note the enumeration grows as two to the N, which makes the analysis impossible as networks grow, one reason they restrict to small systems and synthetic data). For each setting of the threshold parameter they classify all elements of the dataset by which fixed point they flow to, giving a distribution of basin sizes, and compute its normalized entropy: high when every memory captures an equal share of state space, low when one or two basins dominate.

The operative finding is a dose-response. Push the threshold too low and the network holds few fixed points, and those weakly. Push it too high and the network proliferates fixed points, but many are undesired spurious memories that steal basin volume from the patterns you actually stored. Between those failure modes sits a regime where the basin-size entropy is maximized and capacity is shared roughly evenly across the intended patterns.¹

Where a skeptic should push

The load-bearing assumption is that basin geometry measured on a 20-bit synthetic toy tells you anything about substrates you care about. The dataset is generated from a Gaussian random model rather than drawn from real examinees, so the classification demonstration shows the machinery runs, not that it wins at anything. The paper is candid about scale: exhaustive basin analysis is exponential in network size, so everything reported lives in a regime far below any physical or biological realization.

There is also a sloppiness worth noting for anyone citing this work: the paper swaps its N and M notation between the figure caption and the equations, captioning the data as 200 students answering 20 problems while defining the mathematics the other way around. It does not change the result, but it is a reminder that this is a conference-style theory sketch, not a benchmark.

Demonstrated: in discrete-time hysteresis networks of small size, the threshold parameter controls the number and stability of fixed points and the entropy of the basin-size distribution, with a regime that maximizes capacity evenness. Asserted: that this constitutes an effective classification method or a design law for real memory hardware. The gap between those is large and the paper does not cross it.

Hysteresis, basins, and organoid memory capacity

The reason this toy deserves attention from the organoid and wetware side is that hysteresis is not a curiosity of a Japanese theory model; it is the default physics of the substrates everyone is actually building. Memristive devices switch with history-dependent thresholds. Synapses depress and facilitate with history-dependent time constants. Voltage-gated ion channels, the literal computing elements of neural tissue, are hysteretic by construction, opening at one voltage and closing at another. A neural organoid on a multielectrode array is, at the device physics level, a large network of hysteretic elements. This paper supplies the vocabulary that substrate has been missing: its memories are fixed points, its retrieval is basin flow, and its capacity is a distribution with a geometry you can measure and tune.

The opportunity is the single-knob result. If one threshold-like parameter sets how evenly capacity is allocated across stored patterns, then the homeostatic mechanisms already present in living tissue, excitation-inhibition set points, adaptation currents, plasticity thresholds, are candidate implementations of exactly that knob. An organoid that is good at holding many distinct activity patterns without one pattern dominating may be one whose homeostatic set point sits near the maximum-entropy regime, and longitudinal recording is one way to watch that state drift. That turns a vague claim, living tissue self-regulates, into a testable one: capacity evenness should track homeostatic state, and disturbing homeostasis should distort the basin distribution before it destroys the patterns.

The threat is the other side of the dose-response. The paper shows that pushing the threshold too high floods the network with spurious fixed points that cannibalize the intended memories. Anyone who has watched a cortical organoid slide into seizure-like, self-sustaining bursting will recognize the phenomenology: the tissue has too many stable attractors and cannot leave them. The framework suggests those states are not random failures but the predictable high-threshold regime of a hysteretic network, which reframes state control in organoid computing as threshold management rather than noise suppression. And the deepest caveat runs cold: basin volumes in real high-dimensional tissue are precisely the object this analysis cannot compute. The entropy-maximization law is a compass for what to measure, not a map of what is there.

The bottom line

Established: in small discrete-time hysteresis networks, a threshold parameter controls fixed-point count, stability, and the entropy of basin-size distributions, with an identifiable capacity-maximizing regime, verified on synthetic binary data. Hypothesis: living and memristive substrates, being hysteretic at the element level, obey the same dose-response, so their memory capacity and their pathological stuck states are two ends of one knob. What would confirm it: direct basin or capacity measurements on a real hysteretic substrate, memristive array or organoid, showing evenness of stored-pattern capacity varying with a homeostatic or switching-threshold parameter. What would break it: capacity in real substrates turning out to be dominated by high-dimensional effects invisible to the small-network analysis, which the paper's own scaling limits already concede.