Fig 9.

(A) Mean E firing rate of the network as a function of mean input η and signal amplitude ε. Black curve shows level set at 13 Hz, and is parametrized by the normalized arclength x in direction of arrow. Parameter set used throughout the rest of paper (ε = 0.5, η = −0.5) corresponds to x ≃ 0.3. (B) Illustration of input I_i(t) presented to a neuron as a function of the parameter x. The red line shows the mean η(x) while the blue line shows fluctuations of amplitude ε(x). For all parameters, the mean excitatory firing rate is constant. Red dot shows the benchmark regime (η = −0.5, ε = 0.5). (C) Ratio of stimulus amplitude ε to that of network interaction amplitude σ_E,I as derived from the mean field equations (see Eq (6)). Red dot shows the benchmark regime (η = −0.5, ε = 0.5). (D) The fraction of positive Lyapunov exponents #{λ_i > 0}/N as a function of the parameter x. Red dot shows the benchmark regime (η = −0.5, ε = 0.5). (E) Level curves of classification performance P = 1 in the (x, N_r) parameter space for several values of ρ_same. For parameter pairs to the left of these curves, the network achieves perfect classification on average. Stars indicate the corresponding N_same = ρ_same N. Red dot shows the benchmark regime (η = −0.5, ε = 0.5). (F) Classification performance P as a function of x and readout dimension N_r for three input similarity fractions ρ_same = 0, 0.5, 0.9. Stars indicate the corresponding N_same = ρ_same N. For all P values shown, quantities were sampled over 20 repeated simulations where networks, inputs and Tempotron training were re-processed. Ordered readouts, ρ_corr = 0, N = 500, K = 20.