From: What can ecosystems learn? Expanding evolutionary ecology with learning theory
 | Unsupervised correlation learning | Coevolution |
---|---|---|
 | Activation dynamics | Population dynamics |
a) | Neural activation level | Species density, x i |
b) | Neural activation pattern | Ecological state, X={x i ,x 2,...x N } |
c) | Synaptic connection strength, ω ij | Inter-species fitness interaction, ω ij |
d) | Neural network (weight matrix, W). | Ecological network (community matrix, Ω). |
e) | Neural activation dynamics: a non-linear weighted sum of inputs from other neurons (and external inputs). | Ecological population dynamics (Eq. 1): species growth is a non-linear function of the sum of weighted fitness interactions from other species (and environmental changes to carrying capacities). |
f) | External input patterns | Environmental forcing |
 | (aka. ‘training set’). | (in multiple environmental conditions) |
 | Correlation learning (unsupervised) | Evolution of interactions (individual selection only) |
g) | Positive feedback between activation strengths and connection strengths – aka. neurons that fire together wire together. Unsupervised correlation learning mechanism, Hebb’s rule: Δ ω ij =r x i x j , where r>0 is a learning rate. | Positive feedback between ecological densities and connections – or species that occur together wire together. Direct effects of individual natural selection on interactions: v ij =r x i x j , where \(r = \frac {m_{i}}{k_{\textit {ie}}} g \mu \) describes the available mutation (Eq. 3). |
 | Collective behaviours in neural networks (arising from e.g., Hebbian learning, Fig. 1 ) | Collective behaviours in ecosystems (arising from individual selection acting upon interspecific correlations) |
h) | Memory formation (Fig. 1, top panel) Hebb’s rule organises synaptic connections to reinforce the state of the system, decreasing sensitivity to changes in input. | Ecological memory formation(Fig. 3): natural selection organises ecological relationships in a manner that reinforces the current ecological state, decreasing sensitivity to changes in environmental conditions. (Attractors due to environmental variables become attractors of community dynamics [17].) |
i) | Distributed associative memory facilitates a memory of multiple patterns (Fig. 1 a): the capacity to store multiple patterns of activation in the organisation of synaptic connections and recall patterns from any initial conditions via activation dynamics. | Formation of alternative stable states (Fig. 5 a): the creation of a distributed ecological memory in the network of species interactions results in a system with attractors that mimic past ecological states. |
j) | Pattern reconstruction (Fig. 1 b): the recall of a complete pattern from a partial stimulus. | Ecological assembly dynamics (Fig. 5 b): reconstruction of a particular community composition, from a subset of that community. |
k) | Error correction (Fig. 1 c): the ability to remove noise from a pattern, repair imperfections and restore a complete pattern. | Ecological resilience (Fig. 5 c): the ability to recover from perturbations in species densities and restore the complete community. |
l) | Recognition or classification of an input or stimulus (Fig. 1 d): return the nearest attractor from ambiguous initial conditions. | Ecological sensitivity to initial conditions (Fig. 5 d): the switch-like change in response to small variation in initial species densities. |
m) | Holding state in dynamics: Hopfield networks and other recurrent networks have an internal state that allows them to display temporal dynamics (independent of input). | Ecosystems hold state in population dynamics (Fig. 6): in systems with multiple attractors this results in a communities capable of hysteresis with tipping points between states. |