5D Library › Neurobiology of Ritual, Spatial Boundary Transitions and Computational Biomedicine of Self-Transcendence · 中文

Chapter 4 From Rumination to the Present: the Triple-Network Model and Salience-Network Switching

Three large-scale networks

The resting brain is not quiet. Book 8 discussed the energy budget of the default mode network (DMN), its hub structure centred on the posterior cingulate and medial prefrontal cortex, and evidence of DMN deactivation during meditation in experienced practitioners; that is only recalled here. Menon (2011) placed three large-scale networks into one framework, the triple-network model (Menon 2011):

  • the default mode network, centred on the posterior cingulate cortex (PCC) and medial prefrontal cortex (mPFC), active in self-referential thought, remembering the past and imagining the future;
  • the central executive network (CEN), centred on the dorsolateral prefrontal and posterior parietal cortex, active when information about an external task must be held and manipulated;
  • the salience network (SN), centred on the anterior insula and dorsal anterior cingulate cortex, which detects “important” events inside and outside the body and switches between the DMN and the CEN.

Menon argued that many psychiatric disorders can be understood as dysregulated relations among these three networks: salience detection goes wrong, or switching fails, leaving a person stuck in a DMN-dominated, self-referential state. The three networks of Posner discussed in Book 24 (alerting, orienting, executive control) are a different division and must not be mixed up with these.

Evidence

The “switching” role of the right fronto-insular cortex (human fMRI). In an auditory event-segmentation task, a visual oddball task and rest, Sridharan, Levitin and Menon (2008) used Granger-causal and other effective-connectivity methods to estimate the direction of influence between networks, found that the right fronto-insular cortex influenced CEN and DMN nodes more than the reverse, and proposed that it plays a causal role in switching between the two networks (Sridharan et al. 2008). Note that this “causality” is inferred from the temporal order of blood-oxygen signals, whose temporal resolution is limited and whose regional haemodynamic delays may differ; it is not the result of a direct intervention such as stimulation or a lesion.

Evidence

Rumination and the DMN (synthesis of human studies). Hamilton, Farmer, Fogelman and Gotlib (2015) reviewed and pooled functional-connectivity studies of depression and argued that connectivity between the DMN and the subgenual anterior cingulate is increased in depression and tracks the degree of rumination, calling rumination the “dark matter” of clinical neuroscience (Hamilton et al. 2015). This is correlational evidence: increased connectivity and rumination go together, which does not show that lowering the connectivity would reduce rumination.

The framework says that multisensory ritual input, gated by the anterior insula, switches PCC–mPFC narrative self-rumination to “minimal embodied presence”. The first half of this claim is written as a model below; the second half is discussed at the end of the chapter.

The author’s model: Wilson–Cowan-type rate equations

Wilson and Cowan (1972) described the dynamics of local neural populations with equations for the mean firing rates of excitatory and inhibitory populations (Wilson and Cowan 1972). This book treats each of the three networks as one population and writes simplified Wilson–Cowan-type rate equations (the author’s model, not fitted to any data):

τdxdt=−x+F(ux),F(u)=11+e−(u−0.5)/0.08,\tau\frac{dx}{dt} = -x + F(u_x), \qquad F(u) = \frac{1}{1 + e^{-(u-0.5)/0.08}},

uD=0.6D−0.8C−0.8S+ID,uC=0.6C−0.8D+0.6S+IC,uS=0.4S+IS.u_D = 0.6D – 0.8C – 0.8S + I_D,\quad u_C = 0.6C – 0.8D + 0.6S + I_C,\quad u_S = 0.4S + I_S.

The DMN (DD) and CEN (CC) each excite themselves and inhibit each other; the salience network (SS) inhibits DD and excites CC; ID=0.4I_D = 0.4 is the self-referential drive, ICI_C the drive from an external task, and ISI_S the external salient input (in the framework, the multisensory ritual input). The program does three things: A, a salient input lasting 20 time constants, starting from the DD-dominated state; B, a slow rise and fall of ISI_S to find the switching points; C, added noise, measuring the mean time spent before the DD state is left.

Calculation

# Wilson-Cowan-type rate model of the triple network (Chapter 4). The author's model.
# Three populations: D (default mode), S (salience), C (central executive). Each obeys
#   tau dx/dt = -x + F(u_x),  F(u) = 1 / (1 + exp(-(u - 0.5) / 0.08))
#   u_D = 0.6 D - 0.8 C - 0.8 S + I_D      (self-referential drive I_D = 0.4)
#   u_C = 0.6 C - 0.8 D + 0.6 S + I_C      (task drive I_C)
#   u_S = 0.4 S + I_S                       (salient, multisensory input I_S)
# S inhibits D and excites C, so it can switch the D-C competition (Menon's proposal).
# A: a 20-tau pulse of I_S, starting in the D-dominant state, for two task drives.
# B: a slow ramp of I_S up and down (hysteresis). C: with noise, the mean time before the
# D state is left, against I_S.
import math, random

ID, TAU, DT = 0.4, 1.0, 0.1

def F(u):
    return 1 / (1 + math.exp(-(u - 0.5) / 0.08))

def step(x, i_s, ic, rng=None, sig=0.0):
    d, s, c = x
    ud = 0.6 * d - 0.8 * c - 0.8 * s + ID
    uc = 0.6 * c - 0.8 * d + 0.6 * s + ic
    us = 0.4 * s + i_s
    n = [rng.gauss(0, sig * math.sqrt(DT)) for _ in range(3)] if rng else (0, 0, 0)
    return (d + DT / TAU * (F(ud) - d) + n[0], s + DT / TAU * (F(us) - s) + n[1],
            c + DT / TAU * (F(uc) - c) + n[2])

def run(amp, ic, t_end=80.0):
    x, k, out = (0.9, 0.05, 0.05), 0, {}
    while k * DT < t_end:
        t = k * DT
        x = step(x, amp if 20 <= t < 40 else 0.0, ic); k += 1
        out[k] = x
    return out

print("A  pulse of salient input I_S between t = 20 and t = 40 (units of tau)")
for ic in (0.1, 0.2):
    for amp in (0.2, 0.4):
        tr = run(amp, ic)
        cells = "  ".join(f"t={t:2d} D {tr[t * 10][0]:.2f} C {tr[t * 10][2]:.2f}"
                          for t in (19, 39, 79))
        print(f"   I_C {ic:.1f}, I_S {amp:.1f}:  {cells}")

def ramp(ic):
    x, up, down = (0.9, 0.05, 0.05), None, None
    for k in range(401):
        i_s = 0.8 * k / 400
        for _ in range(30):
            x = step(x, i_s, ic)
        if up is None and x[2] > x[0]:
            up = i_s
    for k in range(401):
        i_s = 0.8 * (1 - k / 400)
        for _ in range(30):
            x = step(x, i_s, ic)
        if down is None and x[0] > x[2]:
            down = i_s
    return up, down

def survives(x, i_s, ic, which, n=3000):        # does the state stay D (0) or C (2)?
    for _ in range(n):
        x = step(x, i_s, ic)
    return x[which] > x[2 - which]

def edge(ic, which, lo=0.0, hi=0.8):            # bisection for where a state ends
    start = (0.9, 0.05, 0.05) if which == 0 else (0.05, 0.95, 0.95)
    if which == 2 and survives(start, lo, ic, 2):
        return None                             # C exists even at I_S = 0
    for _ in range(18):
        mid = (lo + hi) / 2
        ok = survives(start, mid, ic, which)
        if (which == 0) == ok:
            lo = mid
        else:
            hi = mid
    return (lo + hi) / 2

print("B  slow ramp of I_S (0 -> 0.8 -> 0): switching points seen by the scan (they")
print("   include sweep lag); bisection with long settling gives the state edges")
for ic in (0.1, 0.2):
    up, down = ramp(ic)
    back = f"{down:.3f}" if down is not None else "never (C persists at I_S = 0)"
    print(f"   I_C {ic:.1f}: ramp D -> C at I_S = {up:.3f}; C -> D at I_S = {back}")
    d_end, c_start = edge(ic, 0), edge(ic, 2)
    cs = f"from I_S = {c_start:.3f}" if c_start is not None else "already at I_S = 0"
    print(f"      edges: D state lost above I_S = {d_end:.3f}; C state exists {cs}")
print("C  noise sd 0.22, I_C 0.1: mean time (tau) before the D state is left, 40 runs;")
print("   runs stopped at the 300-tau cap make the mean a lower bound")
for i_s in (0.0, 0.1, 0.2, 0.25, 0.3):
    rng = random.Random(24); dwell = []
    for _ in range(40):
        x, k = (0.9, 0.05, 0.05), 0
        while x[0] > x[2] and k < 3000:
            x = step(x, i_s, 0.1, rng, 0.22); k += 1
        dwell.append(k * DT)
    cap = sum(1 for d in dwell if d >= 300)
    print(f"   I_S = {i_s:.2f}: {math.fsum(dwell) / len(dwell):6.1f}"
          + (f"  lower bound: {cap} of 40 runs stopped at the cap" if cap else ""))
# A  pulse of salient input I_S between t = 20 and t = 40 (units of tau)
#    I_C 0.1, I_S 0.2:  t=19 D 1.00 C 0.00  t=39 D 1.00 C 0.00  t=79 D 1.00 C 0.00
#    I_C 0.1, I_S 0.4:  t=19 D 1.00 C 0.00  t=39 D 0.00 C 1.00  t=79 D 1.00 C 0.00
#    I_C 0.2, I_S 0.2:  t=19 D 1.00 C 0.00  t=39 D 1.00 C 0.00  t=79 D 1.00 C 0.00
#    I_C 0.2, I_S 0.4:  t=19 D 1.00 C 0.00  t=39 D 0.00 C 1.00  t=79 D 0.00 C 0.97
# B  slow ramp of I_S (0 -> 0.8 -> 0): switching points seen by the scan (they
#    include sweep lag); bisection with long settling gives the state edges
#    I_C 0.1: ramp D -> C at I_S = 0.324; C -> D at I_S = 0.226
#       edges: D state lost above I_S = 0.312; C state exists from I_S = 0.241
#    I_C 0.2: ramp D -> C at I_S = 0.324; C -> D at I_S = never (C persists at I_S = 0)
#       edges: D state lost above I_S = 0.312; C state exists already at I_S = 0
# C  noise sd 0.22, I_C 0.1: mean time (tau) before the D state is left, 40 runs;
#    runs stopped at the 300-tau cap make the mean a lower bound
#    I_S = 0.00:  104.0  lower bound: 4 of 40 runs stopped at the cap
#    I_S = 0.10:   86.1  lower bound: 3 of 40 runs stopped at the cap
#    I_S = 0.20:   55.8
#    I_S = 0.25:   29.6
#    I_S = 0.30:   11.3

Part A: a salient input of IS=0.2I_S = 0.2 is not enough and the state does not change; with IS=0.4I_S = 0.4, during the input DD falls to 0 and CC rises to 1. What happens after the input ends depends on the task drive ICI_C: with IC=0.1I_C = 0.1 the state returns to DD dominance by t=79t = 79; with IC=0.2I_C = 0.2 the CC-dominated state persists (0.97). Part B shows hysteresis: with IC=0.1I_C = 0.1, in the slow ramp scan ISI_S must rise past 0.324 for the switch from DD to CC and fall below 0.226 for the switch back; these values include the lag of the scan, and bisection with full settling at each ISI_S gives the state edges: the DD-dominated state disappears above IS=0.312I_S = 0.312, and the CC-dominated state exists from IS=0.241I_S = 0.241. With IC=0.2I_C = 0.2 the CC-dominated state already exists at IS=0I_S = 0, so the system does not switch back even when ISI_S falls to 0. Part C: noise lets the state switch spontaneously without enough input, and the mean time spent shortens as ISI_S grows, from 104.0 time constants at IS=0I_S = 0 to 11.3 at IS=0.3I_S = 0.3; at IS=0I_S = 0 and 0.1, 4 and 3 runs stopped at the 300 cap, so 104.0 and 86.1 are only lower bounds. Figure 4 shows parts A and B.

Left: in the author’s triple-network model, activity of the DMN (D) and central executive network (C) during and after salient input (grey band); solid and dashed lines are two task drives. Right: hysteresis loop of D activity as the salient input slowly rises and falls.

In this model, three things can be said. First, salient input must exceed a threshold to move the system out of DD dominance; weak input does nothing. Second, whether the switch outlasts the input depends not on the input itself but on whether there is enough task drive afterwards; with nothing “to do” afterwards, the system returns to DD dominance. Third, in a noisy system, salient input acts more by shortening dwell times than by changing the state once and for all. All three are properties of the model; the parameters are arbitrary, and another set would change the threshold and the dwell times.

What “minimal embodied presence” is

The “minimal embodied presence” of the framework roughly means, in the philosophical and meditation literature, a non-narrative awareness of the body and the present moment. There is no such state in this book’s model: the CC-dominated state is an “external task” state, which is not the same as “presence”. To put “presence” into the model one would need at least another variable for interoceptive attention, with measurable counterparts (say insular activity, or performance on a heartbeat-detection task). So “switching to minimal embodied presence” is the author’s interpretation, not a conclusion of the model or of the evidence.

A question of direction also arises. Even if DMN activity falls during a ritual, this may simply be because the person is doing something that needs attention; any moderately demanding task lowers DMN activity. Showing that ritual is special requires comparison with a control task matched for attentional demand.

Hypothesis · not established · Multisensory ritual and rumination

The author’s hypothesis (not established): in adults prone to rumination, a 10-minute multisensory ritual with a fixed action sequence, scent and sound leads to fewer spontaneous self-referential thoughts in the 10 minutes after it ends (measured by randomly prompted experience sampling) than a single-sense task matched for attentional demand and duration (for example listening to the same sound and pressing a key). What would show it wrong: under a pre-registered protocol, a confidence interval for the difference in the share of self-referential thoughts that lies within the smallest effect of interest; or a difference that appears only during the ritual and vanishes as soon as it ends (which matches this chapter’s model prediction of a return to DD without later task drive, but does not support the claim of switching to presence).

Chapter summary

Menon’s triple-network model is an integrative framework; Sridharan and colleagues inferred from effective-connectivity analysis that the right fronto-insular cortex plays a part in network switching, and Hamilton and colleagues reviewed correlational evidence linking rumination to increased DMN connectivity. The author’s Wilson–Cowan-type model shows that, in this model, salient input must exceed a threshold (the DD-dominated state disappears above about 0.312; the slow ramp scan gives 0.324) to switch states, that whether the switch lasts depends on later task drive, and that with noise the input acts by shortening dwell times. “Minimal embodied presence” is not in the model; it is the author’s interpretation. The DMN and attention networks discussed in Books 8 and 24 are only recalled.

References

Hamilton, J. Paul, Madison Farmer, Phoebe Fogelman, and Ian H. Gotlib. 2015. “Depressive Rumination, the Default-Mode Network, and the Dark Matter of Clinical Neuroscience.” Biological Psychiatry 78 (4): 224–30. https://doi.org/10.1016/j.biopsych.2015.02.020.
Menon, Vinod. 2011. “Large-Scale Brain Networks and Psychopathology: A Unifying Triple Network Model.” Trends in Cognitive Sciences 15 (10): 483–506. https://doi.org/10.1016/j.tics.2011.08.003.
Sridharan, Devarajan, Daniel J. Levitin, and Vinod Menon. 2008. “A Critical Role for the Right Fronto-Insular Cortex in Switching between Central-Executive and Default-Mode Networks.” Proceedings of the National Academy of Sciences 105 (34): 12569–74. https://doi.org/10.1073/pnas.0800005105.
Wilson, Hugh R., and Jack D. Cowan. 1972. “Excitatory and Inhibitory Interactions in Localized Populations of Model Neurons.” Biophysical Journal 12 (1): 1–24. https://doi.org/10.1016/S0006-3495(72)86068-5.

Fan, Y. (2026). Neurobiology of Ritual, Spatial Boundary Transitions and Computational Biomedicine of Self-Transcendence: From Predictive-Coding Free-Energy Minimisation, Default-Mode Network Decoupling and Vagal–Endocrine Restoration to Classical Chinese Medicine on Shifting the Spirit (Yijing Bianqi) [仪式神经学、空间边界切换与自我转移的计算生物医学] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23839910 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).