5D Library › 内感受解离、突显网络切换与内脏自主神经修复 · English

第二章 三网络的切换:驻留时间与它的代价

三个网络

证据

静息态网络(综述与原始研究)。 Raichle 等人提出「默认模式」:一组在静息、无任务时活动较高、在外向任务中活动下降的脑区,包括后扣带/楔前叶、内侧前额与双侧顶下叶 (Raichle et al. 2001; Raichle 2015)。Seeley 等人用静息态功能连接分离出两个此前被混为一谈的网络:以背侧前扣带与前岛叶为核心的突显网络,和以背侧额中与顶叶为核心的中央执行网络 (Seeley et al. 2007)。Menon 与 Uddin 提出岛叶在其中起「切换」作用 (Menon and Uddin 2010),Menon 进一步把三者的关系写成「三网络模型」并用于精神病理 (Menon 2011)。Sridharan 等人用格兰杰因果分析认为右前岛叶在执行网络与默认模式之间的切换中起关键作用 (Sridharan, Levitin, and Menon 2008)。

注

「网络」这个词承载了很多。 以上研究用的是功能连接,即体素或脑区时间序列之间的统计依赖,不是解剖连接,也不是因果。第 25 本讨论过网络指标的解释限度;第 52 本第一章给出过一个更极端的例子:一个只反映处方习惯的共现网络,照样可以算出漂亮的聚类系数与「核心节点」。读到「网络」时要问它的边是什么。

默认模式网络与自发思维、走神有关:Mason 等人把走神与默认网络联系起来 (Mason et al. 2007),Christoff 等人提出把自发思维写成动态的约束变化而不是一个状态 (Christoff et al. 2016)。在禅修研究中,Brewer 等人报告有经验的禅修者默认网络主要节点活动较低、连接模式不同 (Brewer et al. 2011),Garrison 等人报告这种差异在超出主动任务之后仍然存在 (Garrison et al. 2015)。

注

这些禅修研究是小样本的横断面或准实验研究。 Brewer 等人与 Garrison 等人的被试数为十余到数十人,且比较的是「有经验者与新手」或「状态之间」,不是随机分配的。横断面差异不能分辨「练习造成了差异」与「本来不同的人更容易坚持练习」。Van Dam 等人系统地讨论了这类问题 (Van Dam et al. 2018)。

把切换写成马尔可夫链

原说法的第二个环节是:内部事件(一个念头、一个身体感觉、一个画面)不再招募突显网络,于是人停在默认模式里不被打断。这可以精确化:把每个采样时刻的「主导状态」记为 D(默认)、S(突显)、E(执行),用转移矩阵描述切换。这类状态序列模型在静息态文献中有先例:Vidaurre 等人用隐马尔可夫模型刻画静息态网络的时间组织,并发现状态访问本身是分层组织的 (Vidaurre, Smith, and Woolrich 2017)。

本章的转移矩阵是作者选定的示意值,不是任何数据的拟合。程序的作用不是估计真实的占位率,而是算出「从一个转移矩阵能推出什么」——这部分是精确的。

推导

# Chapter 2. The triple network as a three-state Markov chain over dominant
# states (D = default mode, S = salience, E = central executive), fitted to
# nothing: the transition matrix is a schematic the author chooses. What the
# program computes is what follows FROM a transition matrix, which is exact:
# the stationary distribution (power iteration), expected dwell time
# 1/(1-p_ii) in units of the sampling step, the entropy rate, and mean first
# passage time from D to E. It then varies one entry, the D->S probability,
# which is the author's reading of "the salience network stops being recruited
# by inner events", and shows what that does to the rest.
import math

STEP = 0.72  # s per TR, a common resting-state fMRI sampling interval
STATES = ("D", "S", "E")

def stationary(P, iters=20000):
    v = [1.0 / len(P)] * len(P)
    for _ in range(iters):
        v = [
            math.fsum(v[i] * P[i][j] for i in range(len(P)))
            for j in range(len(P))
        ]
        s = math.fsum(v)
        v = [x / s for x in v]
    return v

def entropy_rate(P, pi):
    h = 0.0
    for i, row in enumerate(P):
        for p in row:
            if p > 0:
                h -= pi[i] * p * math.log(p, 2)
    return h

def mfpt(P, src, dst):
    """Mean first passage time src -> dst, in steps, by solving (I-Q) t =
    1."""
    idx = [i for i in range(len(P)) if i != dst]
    Q = [[P[i][j] for j in idx] for i in idx]
    n = len(idx)
    A = [
        [(1.0 if i == j else 0.0) - Q[i][j] for j in range(n)] + [1.0]
        for i in range(n)
    ]
    for c in range(n):                              # Gauss-Jordan
        p = max(range(c, n), key=lambda r: abs(A[r][c]))
        A[c], A[p] = A[p], A[c]
        d = A[c][c]
        A[c] = [v / d for v in A[c]]
        for r in range(n):
            if r != c and A[r][c] != 0.0:
                f = A[r][c]
                A[r] = [a - f * b for a, b in zip(A[r], A[c])]
    return A[idx.index(src)][n]

def build(p_ds):
    """Schematic chain; the D row is renormalised so only D->S is varied
    against D->D."""
    rest = 0.10                                    # D -> E held fixed
    return [[1.0 - p_ds - rest, p_ds, rest],
            [0.22, 0.55, 0.23],
            [0.18, 0.12, 0.70]]

if __name__ == "__main__":
    print(
        "schematic transition matrix (rows sum to 1), sampling step {:.2f} s"
        .format(STEP)
    )
    P0 = build(0.25)
    print("        " + "".join("{:>8s}".format(s) for s in STATES))
    for s, row in zip(STATES, P0):
        print(
            "   {:5s}".format(s) + "".join("{:8.2f}".format(v) for v in row)
        )

    pi = stationary(P0)
    print("\nstationary occupancy: " +
          "  ".join("{} {:.3f}".format(s, v) for s, v in zip(STATES, pi)))
    print("expected dwell (s):   " +
          "  ".join("{} {:.2f}".format(s, STEP / (1 - P0[i][i]))
                    for i, s in enumerate(STATES)))
    print("entropy rate: {:.4f} bits/step = {:.4f} bits/s"
          .format(entropy_rate(P0, pi), entropy_rate(P0, pi) / STEP))
    print("mean first passage D -> E: {:.2f} steps = {:.2f} s"
          .format(mfpt(P0, 0, 2), mfpt(P0, 0, 2) * STEP))

    print("\nvarying only D -> S (D -> E fixed at 0.10):")
    print(
        "   p(D->S)   occ D   occ S   occ E  dwell D (s)  H (bits/s)  MFPT"
        " D->E (s)"
    )
    for p_ds in (0.40, 0.30, 0.25, 0.15, 0.08, 0.03):
        P = build(p_ds)
        q = stationary(P)
        print("   {:7.2f} {:7.3f} {:7.3f} {:7.3f} {:12.2f} {:11.4f} {:14.2f}"
              .format(p_ds, q[0], q[1], q[2], STEP / (1 - P[0][0]),
                      entropy_rate(P, q) / STEP, mfpt(P, 0, 2) * STEP))

    print(
        "\nLowering D -> S raises default-mode occupancy and lengthens its"
        " dwell"
    )
    print(
        "time; it also LOWERS the entropy rate and LENGTHENS the path to the"
    )
    print(
        "executive state. A model in which inner events stop recruiting"
        " salience"
    )
    print(
        "is therefore also a model of becoming slower to engage a task."
        " That is"
    )
    print("a cost, not a side effect.")
# schematic transition matrix (rows sum to 1), sampling step 0.72 s
#                D       S       E
#    D        0.65    0.25    0.10
#    S        0.22    0.55    0.23
#    E        0.18    0.12    0.70
#
# stationary occupancy: D 0.362  S 0.293  E 0.345
# expected dwell (s):   D 2.06  S 1.60  E 2.40
# entropy rate: 1.2747 bits/step = 1.7704 bits/s
# mean first passage D -> E: 6.83 steps = 4.92 s
#
# varying only D -> S (D -> E fixed at 0.10):
#    p(D->S)   occ D   occ S   occ E  dwell D (s)  H (bits/s)  MFPT D->E (s)
#       0.40   0.285   0.351   0.364         1.44      1.8348           4.47
#       0.30   0.332   0.315   0.353         1.80      1.8035           4.74
#       0.25   0.362   0.293   0.345         2.06      1.7704           4.92
#       0.15   0.440   0.234   0.326         2.88      1.6437           5.43
#       0.08   0.519   0.174   0.307         4.00      1.4673           6.02
#       0.03   0.596   0.116   0.288         5.54      1.2472           6.66
#
# Lowering D -> S raises default-mode occupancy and lengthens its dwell
# time; it also LOWERS the entropy rate and LENGTHENS the path to the
# executive state. A model in which inner events stop recruiting salience
# is therefore also a model of becoming slower to engage a task. That is
# a cost, not a side effect.

读表的顺序是这样的。给定矩阵,平稳占位为 D 0.362、S 0.293、E 0.345,期望驻留时间(step/(1−pii)\text{step}/(1-p_{ii}),采样步长 0.72 秒)为 2.06、1.60、2.40 秒,熵率 1.2747 比特/步、1.7704 比特/秒,从 D 到 E 的平均首达时间 6.83 步即 4.92 秒。

然后只改一个数:p(D→S)p(\text{D}\to\text{S}),也就是「内部事件招募突显网络」的概率,而 p(D→E)p(\text{D}\to\text{E}) 固定在 0.10。结果是单调的,而且是三件事一起变:

  • 默认模式占位从 0.285 升到 0.596;
  • 默认模式驻留时间从 1.44 秒升到 5.54 秒;
  • 熵率从 1.8348 降到 1.2472 比特/秒,到执行状态的首达时间从 4.47 秒升到 6.66 秒。
三网络马尔可夫链中,只改变 p(\mathrm{D}\to\mathrm{S}) 的后果。左:默认模式占位率;中:默认模式期望驻留时间;右:熵率(金色实线,左轴)与从 D 到 E 的平均首达时间(灰色虚线,右轴)。向左移动即「内部事件较少招募突显网络」。

前两项是原说法想要的,后两项是它没有提的代价。这个代价不是经验发现,而是矩阵的算术后果:让一个状态更黏,就同时让整个链的行为更可预测、让离开它更慢。

假说 · 尚未验证

作者的假说(尚未确立): 以「不招募突显」为训练目标的练习,在提高默认模式驻留时间的同时,会在练习后的一段时间内延长对外部任务的切换延迟,表现为任务开始时的反应时变长或遗漏增多。什么样的结果能证明它是错的: 在事先登记的试验中,以练习后固定时点的任务切换代价(例如线索任务切换范式中的切换代价)为预先指定结局,若练习组的切换代价与结构匹配的对照组相比没有变长,且置信区间排除了预先设定的最小关注差别,则本假说错。

注

熵率下降不是「坏」,首达时间变长也不是「坏」。 它们只是代价。一个在工作中需要频繁切换的人,和一个被反复侵入的念头打断的人,付得起的代价不同。本书不替读者权衡,只要求把代价写出来;第八章把这件事落实为一个预先指定的阴性对照结局与一个切换代价的次要结局。

「安抚默认模式网络」这句话的问题

原说法里有一句「高阶觉知安抚 DMN」。这句话在两个层面上需要修正。

第一,方向。默认模式网络的活动在走神时较高,但它不是「紧张」的网络;外向的、需要努力的任务会降低它的活动。如果练习提高了默认模式的占位与驻留(上节的结果),那么从网络指标上说,练习是让默认模式更活跃,不是「安抚」它。禅修研究报告的默认网络活动降低 (Brewer et al. 2011; Garrison et al. 2015),与「停在默认模式里」是两个不同的描述;二者能否同时成立,取决于「默认模式活动」与「主导状态为 D」是不是同一件事,而它们不是。

第二,层次。「高阶觉知」在第一章的模型里是参数 πh\pi_h,一个缩放第二层通路的标量。把它指认为某个脑区或某个网络需要独立证据。前岛叶的「切换」角色 (Sridharan, Levitin, and Menon 2008; Menon and Uddin 2010) 是一个候选,但那是关于切换的证据,不是关于「某个网络下调另一个网络的精度」的证据。

证据

可以确定的是什么。 三个网络可以在静息态数据中分离 (Seeley et al. 2007);它们的活动在时间上呈现可用状态序列刻画的结构 (Vidaurre, Smith, and Woolrich 2017);前岛叶/背侧前扣带与注意的重定向有关 (Menon and Uddin 2010; Uddin 2015);默认模式活动与自发思维相关 (Mason et al. 2007; Christoff et al. 2016)。不能确定的是:哪一个网络「调节」哪一个网络的精度,以及这种调节是否是任何练习的作用机制。

从状态序列能测到什么

假设真的记录了一段状态序列(例如用隐马尔可夫模型从功能磁共振或脑电中估计),能回答什么问题?上节程序给出的三个量——占位率、驻留时间、熵率——都是从转移矩阵推出来的,而转移矩阵要从有限长的序列估计。这里有一个与第三章相同的统计问题:短序列上的熵估计有系统性的向下偏差。第三章程序 p13 给出量级,并说明校正方法;那里的对象是自评的表象类别序列,这里是网络状态序列,数学是同一个。

因此,本书不建议把熵率作为任何主要结局。可用的是占位率与驻留时间:它们是比例与倒数,偏差小得多,而且有直接的解释。第八章把前额的注意相关指标列为次要结局,并明确它们不改变结论。

本章小结

三个网络在静息态数据中可以分离,它们的时间组织可以用状态序列模型刻画,这些都有证据。把「内部事件不再招募突显网络」写成转移概率的下降之后,结果是三件事同时发生:默认模式占位从 0.285 升到 0.596、驻留时间从 1.44 升到 5.54 秒,而熵率从 1.8348 降到 1.2472 比特/秒、到执行状态的首达时间从 4.47 升到 6.66 秒。前两项是原说法想要的,后两项是它的代价,且这个代价是算术必然,不是经验发现。「安抚默认模式网络」这句话在方向与层次上都需要修正:提高默认模式驻留与降低默认模式活动是两个不同的描述,而把「高阶觉知」指认为某个网络的精度调节还缺证据。最后,状态序列上的熵率估计在短序列上有系统偏差,不宜作主要结局。

参考文献

Brewer, Judson A., Patrick D. Worhunsky, Jeremy R. Gray, Yi-Yuan Tang, Jochen Weber, and Hedy Kober. 2011. “Meditation Experience Is Associated with Differences in Default Mode Network Activity and Connectivity.” Proceedings of the National Academy of Sciences 108 (50): 20254–59. https://doi.org/10.1073/pnas.1112029108.
Christoff, Kalina, Zachary C. Irving, Kieran C. R. Fox, R. Nathan Spreng, and Jessica R. Andrews-Hanna. 2016. “Mind-Wandering as Spontaneous Thought: A Dynamic Framework.” Nature Reviews Neuroscience 17: 718–31. https://doi.org/10.1038/nrn.2016.113.
Garrison, Kathleen A., Thomas A. Zeffiro, Dustin Scheinost, R. Todd Constable, and Judson A. Brewer. 2015. “Meditation Leads to Reduced Default Mode Network Activity Beyond an Active Task.” Cognitive, Affective and Behavioral Neuroscience 15: 712–20. https://doi.org/10.3758/s13415-015-0358-3.
Mason, Malia F., Michael I. Norton, John D. Van Horn, Daniel M. Wegner, Scott T. Grafton, and C. Neil Macrae. 2007. “Wandering Minds: The Default Network and Stimulus-Independent Thought.” Science 315 (5810): 393–95. https://doi.org/10.1126/science.1131295.
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.
Menon, Vinod, and Lucina Q. Uddin. 2010. “Saliency, Switching, Attention and Control: A Network Model of Insula Function.” Brain Structure and Function 214: 655–67. https://doi.org/10.1007/s00429-010-0262-0.
Raichle, Marcus E. 2015. “The Brain’s Default Mode Network.” Annual Review of Neuroscience 38: 433–47. https://doi.org/10.1146/annurev-neuro-071013-014030.
Raichle, Marcus E., Ann Mary MacLeod, Abraham Z. Snyder, William J. Powers, Debra A. Gusnard, and Gordon L. Shulman. 2001. “A Default Mode of Brain Function.” Proceedings of the National Academy of Sciences 98 (2): 676–82. https://doi.org/10.1073/pnas.98.2.676.
Seeley, William W., Vinod Menon, Alan F. Schatzberg, Jennifer Keller, Gary H. Glover, Heather Kenna, Allan L. Reiss, and Michael D. Greicius. 2007. “Dissociable Intrinsic Connectivity Networks for Salience Processing and Executive Control.” Journal of Neuroscience 27 (9): 2349–56. https://doi.org/10.1523/JNEUROSCI.5587-06.2007.
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.
Uddin, Lucina Q. 2015. “Salience Processing and Insular Cortical Function and Dysfunction.” Nature Reviews Neuroscience 16: 55–61. https://doi.org/10.1038/nrn3857.
Van Dam, Nicholas T., Marieke K. van Vugt, David R. Vago, Laura Schmalzl, Clifford D. Saron, Andrew Olendzki, Ted Meissner, et al. 2018. “Mind the Hype: A Critical Evaluation and Prescriptive Agenda for Research on Mindfulness and Meditation.” Perspectives on Psychological Science 13 (1): 36–61. https://doi.org/10.1177/1745691617709589.
Vidaurre, Diego, Stephen M. Smith, and Mark W. Woolrich. 2017. “Brain Network Dynamics Are Hierarchically Organized in Time.” Proceedings of the National Academy of Sciences 114 (48): 12827–32. https://doi.org/10.1073/pnas.1705120114.

Fan, Y. (2026). Interoceptive Decoupling, Salience-Network Switching and Autonomic Visceral Restoration: From Precision-Weighting in the Default Mode Network and the Redistribution of Vagal Tone to the Neurophenomenology of a Non-Judgmental Imagery Stream, with a Textual Examination of neiguan in the Huangdi Neijing [内感受解离、突显网络切换与内脏自主神经修复] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24020418 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).