Contribution
It shows a correspondence between individual-level selection on interspecific interactions and unsupervised Hebbian learning, and uses a Hopfield-style simulation to show that a community with no community-level selection can store and recall more than one past composition. It offers this as a third option beside “no system-level organisation” and group selection.
Key insight
Selection makes each species less sensitive to the competitors it meets most, and meeting is proportional to joint density, so “species that occur together wire together”. If a species cannot reduce all competition at once, because of a fixed budget, the relative pattern of who competes with whom comes to encode the compositions the community has been forced into, and those compositions become attractors.
Assumptions
- Ecology (eq. 1 as corrected by the erratum): dx_i/dt = (m_i x_i / k_ie)(k_ie + Σ_j ω_ij x_j), ω_ii = −1, all ω_ij ≤ 0 (competition only).
- Selection (eqs. 2–3): a mutant in i with ω_ij′ raised by g has selection coefficient s = (m_i/k_ie) g x_j′; with fixation probability ∝ s, the rate of change is v_ij = (m_i/k_ie) g µ x_i x_j. Only changes to ω_ij (how j affects i) are selected; side-effects are ignored (Appendix A).
- Trade-off (Appendix C): after each update every row and column sum of the off-diagonal ω is renormalised to its initial value Q_i.
- Simulation: symmetric ω, N = 400, ecology run to equilibrium (τ = 5000 steps) between evolutionary updates; environments E1 and E2 shift carrying capacities k0 ± α (k0 = 10, α = 0.1) in two 20 × 20 pictogram patterns, switching every update (T = 1).
Key results
- Hebbian rule. v_ij ∝ x_i x_j, Table 1(g); robust in sign pattern to two models with linkage or rare outcrossing (Appendices B, D, Figs. 7–8).
- One memory (Experiment 1): after 800 cycles in E1 the community returns to the E1 pattern under E0 or E2 forcing, “some distortion” in E2 (Fig. 3).
- Two memories (Experiment 2): with E1 and E2 alternating, the evolved community without forcing has two attractors, E1 and E2, and not their union; it completes patterns from small parts, repairs corrupted ones and goes to the nearer of the two from mixtures (Fig. 5).
- Hysteresis. Before evolution the response to forcing moving from E1 to E2 is proportional; after, it is switch-like with hysteresis, and only for forcing along remembered directions; arbitrary forcing stays linear (Fig. 6, Appendix E, Fig. 9).
- Breakdown. The two-attractor state appears around generation 525, then one attractor outcompetes the other and, after about 575 generations, the remaining one degrades (Figs. 10–11, Appendix F).
- The trade-off is necessary: without it, selection “make[s] high density species fitter in all conditions” and no memory forms (Appendix H).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Individual selection changes ω_ij at a rate proportional to x_i x_j | moderate | eq. 3 under strong-selection, weak-mutation fixation; robust in sign (App. B, D) |
| C2 | With normalising trade-offs these changes have the sign pattern of Hebb’s rule | moderate | Figs. 4, 8, for the simulated cases |
| C3 | A community can hold and recall several past compositions without community-level selection | moderate | simulation, symmetric ω, one parameter set |
| C4 | The memory persists | weak | contradicted in the long run by the paper’s own Figs. 10–11 |
| C5 | Critical transitions may be the recall of remembered states | weak | interpretive, from Fig. 6 |
Concepts
- Ecological memory (after Thompson et al.): past conditions and selection encoded in the current structure of a community.
- Unsupervised correlation learning: changing connections to reinforce frequent co-activations, with no reward signal.
- Normalising constraint: a fixed total of competitive interaction per species, so lowering one raises others.
Connections
The neural side is Hopfield’s associative memory (LIT-918), whose symmetric weights give a Lyapunov function, as symmetric ω does here. Watson and Szathmáry’s “evolution can learn” programme, of which this is the ecological instance, is not registered. The symmetric Lotka–Volterra community is the setting of CLAIM-369.
Bearing on the record
- QUESTION-083 asks whether processes sharing one rule can come to hold differentiated, mutually regulated organization with no governance role coded in. This paper is a positive instance in its own domain, but with a coded-in constraint: the normalising trade-off. The authors say the result depends on it (Appendix H), and it is imposed as a renormalisation step, not derived. The question’s caveat, that a positive answer shows organization from fixed local objectives and not self-constitution (CLAIM-288), applies here too.
- CLAIM-369: the simulations keep ω symmetric, which gives the community a potential, and the paper relies on that (“the dynamics have only fixed point equilibria”). It is CLAIM-369’s class with the interaction matrix evolving slowly. Evolution here reshapes the objective the community descends, which the claim’s fixed-U setting does not cover.
- QUESTION-132: the long-run breakdown, one attractor outcompeting the other, is an instance of the question’s “entrenched asymmetry” outcome emerging from initially balanced memory.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Equation (1) as published has a stray ”+”: “dx_i/dt = (m_i x_i / k_ie) + (k_ie + Σ ω_ij x_j)”. The erratum corrects it to the product. The arXiv v1 has the same error.
- The renormalisation equations (10)–(11) divide each entry by its row (or column) sum, which sets the sums to 1 and flips the sign of negative entries, while the text says the sums are held at Q_i < 0. The factor Q_i is missing from the printed formulas.
- Parameter inconsistencies. Appendix C gives g = 0.01 and α = 0.1; the Figure 7 caption uses g = 0.1 and the Figure 8 caption g = 0.1 and α = 3.5, without saying these runs differ.
- Text slips. In result iv, the forcing moves “from E1 to E1” and the system “switch[es] to E1”; E2 is meant in two of those places. The Figure 5 caption labels its fourth panel “e”.
- Symmetry is imposed in simulation though individual selection acts on ω_ij only; the authors argue the selective pressures are symmetric “except for the influence of individually-varying carrying capacities” (Appendix H), which are exactly what the forcing varies.
- Fixation probability is taken equal to s; the usual approximation is 2s. A constant, so the direction of the results is unaffected.
- Transience. The two-memory state lasts a bounded window of evolutionary time before collapse.
Open questions
- Does a stoichiometric or resource-explicit model, which the authors suggest, produce the normalising trade-off endogenously?
- Do the results survive asymmetric ω, where cycles are possible?