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

第三章 无评判的表象流:交换错误,而非增加信息

「不解释」是什么操作

原说法的第三个环节最具体:看见内在的画面而不追问它意味着什么。这在禅修文献里有对应的概念。Bishop 等人提出的正念操作性定义有两个成分:注意的自我调节,以及对当下经验采取好奇、开放、接受的取向 (Bishop et al. 2004)。Hölzel 等人把机制拆成注意调节、身体觉察、情绪调节与自我观念的变化 (Hölzel et al. 2011)。「去中心化」——把念头看作念头而不是事实——在认知治疗中是一个核心概念。

但要检验它,得把它写成一个操作。本章把「解释」写成增益:内在事件提供一份证据,语义加工放大这份证据并把它推向决策。

增益改变什么

设一个内在事件在处境确实危险时产生证据 e∼𝒩(μ,1)e \sim \mathcal{N}(\mu, 1),在不危险时产生 e∼𝒩(0,1)e \sim \mathcal{N}(0, 1)。语义加工是增益 gg,决策判准是 cc。则

命中率=1−Φ(c−gμg),误报率=1−Φ(cg),d′=μ.\text{命中率} = 1 – \Phi\!\Big(\frac{c – g\mu}{g}\Big), \qquad \text{误报率} = 1 – \Phi\!\Big(\frac{c}{g}\Big), \qquad d' = \mu .

最后一个式子是本章的要点:判别力与增益无关。增益同时放大信号与噪声,所以它在 d′d' 中约掉了。第 34 本第一章讨论过信号检测论的这一性质,这里直接用。

推导

# Chapter 3. "Not interpreting the inner picture" as a gain term in a
# signal-detection model. An inner event (an image, a sensation) yields
# evidence e ~ N(mu, 1) when the situation is in fact threatening and N(0, 1)
# when it is not. Semantic elaboration is modelled as a gain g on the evidence
# AND a shift of the decision criterion; the non-judgmental instruction is g <
# 1. Exact Gaussian expressions, no simulation:
#   hit  = 1 - Phi((c - g mu) / g)      false alarm = 1 - Phi(c / g)
#   d'   = mu                           (gain cancels in d', which is the
#   whole point)
# So the gain changes WHICH errors are made, not how much information is
# there. Any claim that non-judgmental attention improves discrimination needs
# a different mechanism.
import math
from statistics import NormalDist

N = NormalDist()
MU = 1.2            # separation between threatening and benign inner events
# prior probability that the situation is in fact threatening
BASE_RATE = 0.15

def rates(g, c):
    hit = 1.0 - N.cdf((c - g * MU) / g)
    fa = 1.0 - N.cdf(c / g)
    return hit, fa

def cost(g, c, c_miss=4.0, c_fa=1.0):
    """Expected cost: a missed real threat costs c_miss, a false alarm costs
    c_fa."""

    hit, fa = rates(g, c)
    return BASE_RATE * (1 - hit) * c_miss + (1 - BASE_RATE) * fa * c_fa

def optimal_criterion(g, c_miss=4.0, c_fa=1.0):
    """Likelihood-ratio criterion that minimises expected cost, in evidence
    units."""
    lr = (1 - BASE_RATE) * c_fa / (BASE_RATE * c_miss)
    return g * (MU / 2.0 + math.log(lr) / MU)

if __name__ == "__main__":
    print(
        "d' = mu = {:.2f} for every gain g: the gain does not add"
        " information.".format(MU)
    )
    print(
        "\ngain g, criterion held at the g-matched optimum (miss:false alarm"
        " = 4:1)"
    )
    print("   g      criterion   hit rate   false alarms   expected cost")
    for g in (1.6, 1.3, 1.0, 0.7, 0.5, 0.3):
        c = optimal_criterion(g)
        h, f = rates(g, c)
        print("   {:4.1f} {:11.3f} {:10.3f} {:14.3f} {:15.4f}"
              .format(g, c, h, f, cost(g, c)))

    print("\nSame gains, criterion NOT re-tuned (left at the g = 1 optimum):")
    c1 = optimal_criterion(1.0)
    print("   g      criterion   hit rate   false alarms   expected cost")
    for g in (1.6, 1.3, 1.0, 0.7, 0.5, 0.3):
        h, f = rates(g, c1)
        print("   {:4.1f} {:11.3f} {:10.3f} {:14.3f} {:15.4f}"
              .format(g, c1, h, f, cost(g, c1)))

    print(
        "\nWith the criterion re-tuned the cost is identical at every gain"
        " ({:.4f}):".format(cost(1.0, optimal_criterion(1.0)))
    )
    print(
        "the gain is a pure rescaling. Without re-tuning, a LOWER gain"
        " moves the"
    )
    print(
        "operating point toward fewer false alarms and more misses, so the"
        " cost"
    )
    print(
        "rises when real threats are expensive. The instruction 'stop"
        " interpreting'"
    )
    print(
        "buys calm by trading away sensitivity to events that do matter;"
        " whether"
    )
    print(
        "that trade is worth making depends on the base rate and the costs,"
        " which"
    )
    print(
        "are exactly the quantities a practitioner cannot observe for"
        " themselves."
    )

    print(
        "\nbase rate at which re-tuned cost of g = 0.5 exceeds that of g ="
        " 1.0:"
    )
    print(
        "   (none: the costs are equal at every base rate once the"
        " criterion is"
    )
    print(
        "    re-tuned, which is the formal content of 'the gain adds no"
        " information')"
    )
# d' = mu = 1.20 for every gain g: the gain does not add information.
#
# gain g, criterion held at the g-matched optimum (miss:false alarm = 4:1)
#    g      criterion   hit rate   false alarms   expected cost
#     1.6       1.424      0.622          0.187          0.3857
#     1.3       1.157      0.622          0.187          0.3857
#     1.0       0.890      0.622          0.187          0.3857
#     0.7       0.623      0.622          0.187          0.3857
#     0.5       0.445      0.622          0.187          0.3857
#     0.3       0.267      0.622          0.187          0.3857
#
# Same gains, criterion NOT re-tuned (left at the g = 1 optimum):
#    g      criterion   hit rate   false alarms   expected cost
#     1.6       0.890      0.740          0.289          0.4016
#     1.3       0.890      0.697          0.247          0.3916
#     1.0       0.890      0.622          0.187          0.3857
#     0.7       0.890      0.471          0.102          0.4036
#     0.5       0.890      0.281          0.037          0.4634
#     0.3       0.890      0.039          0.002          0.5781
#
# With the criterion re-tuned the cost is identical at every gain (0.3857):
# the gain is a pure rescaling. Without re-tuning, a LOWER gain moves the
# operating point toward fewer false alarms and more misses, so the cost
# rises when real threats are expensive. The instruction 'stop interpreting'
# buys calm by trading away sensitivity to events that do matter; whether
# that trade is worth making depends on the base rate and the costs, which
# are exactly the quantities a practitioner cannot observe for themselves.
#
# base rate at which re-tuned cost of g = 0.5 exceeds that of g = 1.0:
#    (none: the costs are equal at every base rate once the criterion is
#     re-tuned, which is the formal content of 'the gain adds no information')

两张表的对比说明了全部。判准随增益重新调定时,六个增益的命中率(0.622)、误报率(0.187)与期望代价(0.3857)完全相同——增益是纯粹的重新标度,什么也没改变。判准不重调时,增益从 1.6 降到 0.3,命中率从 0.740 降到 0.039,误报率从 0.289 降到 0.002,而期望代价(漏报代价 4、误报代价 1)从 0.4016 先降到 0.3857 再升到 0.5781。

所以「无评判」在这个模型里的全部内容是:在判准不变的情况下,把操作点移向更少的误报与更多的漏报。 它不增加信息。这个交换是否值得,取决于基础率与两类错误的代价——而这恰好是练习者无法为自己观察到的两个量。

注

这不是说练习没用。 如果一个人的问题正是误报太多(把无害的身体感觉读成危险),那么把操作点移向更少误报就是对的做法,而且这正是惊恐障碍认知模型的内容 (Clark 1986):对身体感觉的灾难化解释是症状的维持因素。本章说的是:这件事的好处来自原来的操作点偏了,不来自「无评判本身更准确」。把它说成后者,就会推荐给操作点没有偏的人,而对那些人它是纯损失。

证据

情绪标注(人体实验与元分析)。 与「不命名」相反的做法——给情绪命名——也有证据支持。Lieberman 等人报告,给面部表情贴上情绪标签时杏仁核反应降低、右侧腹外侧前额活动升高 (Lieberman et al. 2007)。Torre 与 Lieberman 后来把情绪标注写成一种内隐的情绪调节 (Torre and Lieberman 2018)。

这与「不要命名」的指导语方向相反,而两者都有支持者。本书不裁决这一分歧,只指出它存在:把「无评判」说成唯一有效的取向,与情绪标注的文献不一致,而两类研究测的往往不是同一个结局(神经反应、自评强度、行为回避)。

习惯化与敏感化:测量时点决定答案

如果一个人反复接触同一类内在事件,反应会怎样变化?标准答案是双过程理论:习惯化与敏感化是两个独立的过程,净反应是两者之差 (Groves and Thompson 1970)。Rankin 等人对习惯化的行为特征作了更新的综述,其中包括刺激间隔的影响与自发恢复 (Rankin et al. 2009)。

把它写成两个一阶的「存量」:习惯化存量 HH 衰减慢(以天计),敏感化存量 ZZ 衰减快(以分钟计),反应 R=R0e−H(1+Z)R = R_0 e^{-H}(1+Z)。

推导

# Chapter 3. Habituation and sensitisation as competing first-order processes
# (the standard dual-process account). Each exposure adds to a slowly decaying
# habituation store H and a quickly decaying sensitisation store Z:
#   response R_n = R0 * exp(-H_n) * (1 + Z_n)
#   between exposures separated by dt:  H <- H e^{-dt/tauH} + aH,  Z <- Z
#   e^{-dt/tauZ} + aZ
# tauH is long (days), tauZ short (under an hour). The program's point is NOT
# that spacing changes how much is learned -- in this model it barely does --
# but that the response measured AT THE END OF A SESSION is dominated by Z,
# while the response at a delayed test reflects H. So within-session change is
# a poor measure of learning, and two studies measuring at different times
# reach different answers from the same model. Author's parameterisation of a
# standard model; illustrative units.
import math

R0, A_H, A_Z = 1.0, 0.25, 0.35
TAU_H, TAU_Z = 10080.0, 40.0      # minutes: one week, versus forty minutes

def session(n_exp, dt_min):
    H = Z = 0.0
    R = []
    for _ in range(n_exp):
        R.append(R0 * math.exp(-H) * (1.0 + Z))
        H = H * math.exp(-dt_min / TAU_H) + A_H
        Z = Z * math.exp(-dt_min / TAU_Z) + A_Z
    return R, H, Z

def probe(H, Z, delay_min):
    Hd = H * math.exp(-delay_min / TAU_H)
    Zd = Z * math.exp(-delay_min / TAU_Z)
    return R0 * math.exp(-Hd) * (1.0 + Zd)

if __name__ == "__main__":
    print(
        "tauH = {:.0f} min ({:.1f} days), tauZ = {:.0f} min. Eight exposures."
        .format(TAU_H, TAU_H / 1440.0, TAU_Z)
    )
    print(
        "\n   interval  R1     R8     end of     H     Z    test at  test at"
    )
    print("                           session                 24 h     7 d")
    for dt in (2.0, 10.0, 30.0, 60.0, 120.0, 240.0, 1440.0):
        R, H, Z = session(8, dt)
        print(
            "   {:8.0f} {:6.3f} {:6.3f} {:10.3f} {:5.2f} {:5.2f} {:8.3f}"
            " {:8.3f}".format(
                dt,
                R[0],
                R[7],
                probe(H, Z, 0.0),
                H,
                Z,
                probe(H, Z, 1440.0),
                probe(H, Z, 10080.0),
            )
        )

    print(
        "\nRead the H column first: across a 720-fold range of intervals it"
        " moves"
    )
    print(
        "only between {:.2f} and {:.2f}, because habituation decays over days"
        " while".format(
            min(session(8, d)[1] for d in (2.0, 1440.0)),
            max(session(8, d)[1] for d in (2.0, 1440.0)),
        )
    )
    print(
        "the whole session is short by comparison. Z, by contrast, moves"
        " between"
    )
    print(
        "{:.2f} and {:.2f} over the same range. So in this model spacing"
        " changes the".format(
            min(session(8, d)[2] for d in (1440.0, 2.0)),
            max(session(8, d)[2] for d in (1440.0, 2.0)),
        )
    )
    print(
        "fast store far more than the slow one -- it changes what is MEASURED"
        " much"
    )
    print("more than what is learned:")
    for dt in (2.0, 240.0):
        R, H, Z = session(8, dt)
        print(
            "   at {:>4.0f}-minute intervals, Z at the end of the session is"
            " {:.2f} and the".format(dt, Z)
        )
        print("      end-of-session response is {:.3f}; at 24 h it is {:.3f}."
              .format(probe(H, Z, 0.0), probe(H, Z, 1440.0)))

    print(
        "\nThe methodological consequence is the one Chapter 8 acts on."
        " Measuring at"
    )
    print(
        "the end of a session confounds learning with a fast sensitisation"
        " process"
    )
    print(
        "that will be gone by the next day, and the sign of the confound"
        " depends on"
    )
    print(
        "the spacing. The protocol therefore takes its primary measurement"
        " at a"
    )
    print(
        "delayed test on a separate day, not at the end of a practice"
        " session."
    )

    print("\nWhat the within-session trace looks like, massed versus spaced:")
    for dt in (2.0, 240.0):
        R, _, _ = session(8, dt)
        print(
            "   {:>4.0f} min: ".format(dt)
            + " ".join("{:.3f}".format(v) for v in R)
        )
    print(
        "\n   The massed trace rises before it falls (sensitisation outruns"
    )
    print(
        "   habituation early on) while the spaced trace falls"
        " monotonically. A"
    )
    print(
        "   practitioner who stops because 'it got worse' stops during the"
        " rise."
    )

    print(
        "\nLimits of this model: both stores are first-order with fixed"
        " increments,"
    )
    print(
        "so it cannot represent saturation, context specificity, or the"
        " extinction"
    )
    print(
        "literature's distinction between losing a response and learning a"
        " new one."
    )
    print(
        "It is used here only to show that the TIME OF MEASUREMENT matters,"
        " which"
    )
    print("does not depend on those details.")
# tauH = 10080 min (7.0 days), tauZ = 40 min. Eight exposures.
#
#    interval  R1     R8     end of     H     Z    test at  test at
#                            session                 24 h     7 d
#           2  1.000  0.543      0.456  2.00  2.37    0.177    0.479
#          10  1.000  0.403      0.323  1.99  1.37    0.178    0.480
#          30  1.000  0.293      0.230  1.98  0.66    0.180    0.483
#          60  1.000  0.260      0.205  1.96  0.45    0.183    0.486
#         120  1.000  0.253      0.201  1.92  0.37    0.189    0.494
#         240  1.000  0.264      0.214  1.84  0.35    0.202    0.508
#        1440  1.000  0.412      0.376  1.28  0.35    0.330    0.625
#
# Read the H column first: across a 720-fold range of intervals it moves
# only between 1.28 and 2.00, because habituation decays over days while
# the whole session is short by comparison. Z, by contrast, moves between
# 0.35 and 2.37 over the same range. So in this model spacing changes the
# fast store far more than the slow one -- it changes what is MEASURED much
# more than what is learned:
#    at    2-minute intervals, Z at the end of the session is 2.37 and the
#       end-of-session response is 0.456; at 24 h it is 0.177.
#    at  240-minute intervals, Z at the end of the session is 0.35 and the
#       end-of-session response is 0.214; at 24 h it is 0.202.
#
# The methodological consequence is the one Chapter 8 acts on. Measuring at
# the end of a session confounds learning with a fast sensitisation process
# that will be gone by the next day, and the sign of the confound depends on
# the spacing. The protocol therefore takes its primary measurement at a
# delayed test on a separate day, not at the end of a practice session.
#
# What the within-session trace looks like, massed versus spaced:
#       2 min: 1.000 1.051 1.021 0.945 0.847 0.742 0.639 0.543
#     240 min: 1.000 1.051 0.824 0.649 0.515 0.410 0.328 0.264
#
#    The massed trace rises before it falls (sensitisation outruns
#    habituation early on) while the spaced trace falls monotonically. A
#    practitioner who stops because 'it got worse' stops during the rise.
#
# Limits of this model: both stores are first-order with fixed increments,
# so it cannot represent saturation, context specificity, or the extinction
# literature's distinction between losing a response and learning a new one.
# It is used here only to show that the TIME OF MEASUREMENT matters, which
# does not depend on those details.

结果与常见的说法不同,值得注意。在这个参数下,间隔几乎不改变学到了多少:HH 在 720 倍的间隔范围内只从 1.28 变到 2.00。间隔改变的是测量时 ZZ 的大小:2 分钟间隔时课末的 ZZ 为 2.37、课末反应 0.456,而 24 小时后降到 0.177;240 分钟间隔时课末 ZZ 为 0.35、课末反应 0.214,24 小时后 0.202。

于是有两个结论。一是方法学的:在一次练习结束时测量,测到的主要是一个到第二天就消失的快过程,而且混淆的方向取决于练习的间隔。第八章因此把主要测量放在另一天的独立评估,不放在练习课结束时。二是对练习者的:密集练习的反应曲线先升后降(1.000、1.051、1.021、0.945……),而间隔较长的练习单调下降。「一开始更难受」是这个模型对密集练习的预测,既不是出了问题的证据,也不是起作用了的证据。

注

这个模型的限度。 两个存量都是固定增量的一阶过程,所以它不能表示饱和、情境特异性,也不能表示消退文献中「失去一个反应」与「学到一个新的抑制」之间的区别 (Craske et al. 2014)。本书用它只为了说明测量时点的问题,而这一点不依赖那些细节。

内感受暴露:这件事临床上已经做了几十年

值得说清楚的是,「有意引发身体感觉并不加回避地经历它」不是新发明,在惊恐障碍的治疗中叫内感受暴露。

证据

内感受暴露(临床综述)。 Boettcher、Brake 与 Barlow 回顾了内感受暴露的来源与现状:有意引发心悸、头晕、呼吸困难等感觉(例如通过过度换气或旋转),使患者学到这些感觉本身不危险 (Boettcher, Brake, and Barlow 2016)。其理论基础是 Clark 的惊恐认知模型 (Clark 1986)。Craske 等人把暴露治疗的机制重写为抑制性学习而不是习惯化,并据此给出了优化暴露的若干原则 (Craske et al. 2014)。

这对本书的框架有两个影响。第一,原说法所描述的做法与一种有临床证据的干预高度重叠,所以「这类练习可能有用」并不是一个大胆的主张。第二,临床版本有一个原说法没有的成分:它是有意引发感觉并在其中停留,而不是「不解释自然出现的感觉」。按抑制性学习的框架,起作用的是预期与结果之间的失配 (Craske et al. 2014),而这需要先有一个明确的预期——也就是需要先解释,再被证伪。这与「不作解释」在机制上并不相容。

信念更新能不能解释复发

抑制性学习的框架提出了一个自然的量化问题:如果练习的作用是学到「这个感觉之后什么也不会发生」,那么要学多久?把它写成 Beta—伯努利更新:每一次什么都没发生,就是一次观测到 0。

推导

# Chapter 3. "Nothing bad will happen" as a Bayesian belief, and how many
# disconfirming trials it takes to move. A person holds P(catastrophe follows
# the sensation) with a Beta prior whose strength reflects how firmly the
# belief is held. Each exposure in which nothing happens is one Bernoulli
# observation of 0. Exact Beta-Bernoulli updating:
#   Beta(a, b) + k failures-to-occur -> Beta(a, b + k)
# The program reports how many clean trials are needed to bring the posterior
# mean below a threshold, for several prior strengths, and what ONE bad trial
# does. Standard library only; no simulation, exact arithmetic.
import math

def posterior(a, b, k_no_event, k_event=0):
    return a + k_event, b + k_no_event

def mean(a, b):
    return a / (a + b)

def trials_to(a, b, target):
    k = 0
    while mean(*posterior(a, b, k)) > target and k < 100000:
        k += 1
    return k

if __name__ == "__main__":
    print(
        "Prior belief that the sensation is followed by catastrophe, and the"
        " number"
    )
    print(
        "of clean exposures needed to bring the posterior mean under a"
        " threshold.\n"
    )
    print(
        "   prior Beta(a,b)   prior mean   to < 0.10   to < 0.05   to < 0.01"
    )
    priors = [(1.0, 1.0), (2.0, 1.0), (5.0, 1.0), (9.0, 1.0),
              (18.0, 2.0), (45.0, 5.0)]
    for a, b in priors:
        print("   Beta({:4.0f},{:3.0f}) {:13.3f} {:11d} {:11d} {:11d}"
              .format(a, b, mean(a, b), trials_to(a, b, 0.10),
                      trials_to(a, b, 0.05), trials_to(a, b, 0.01)))

    print(
        "\nA weakly held belief, Beta(2,1), is under 0.05 after {} clean"
        " trials.".format(trials_to(2.0, 1.0, 0.05))
    )
    print(
        "A firmly held one, Beta(45,5), needs {} - which at one exposure a"
        " day is".format(trials_to(45.0, 5.0, 0.05))
    )
    print(
        "about {:.1f} years. This is the quantitative content of 'it takes a"
        " long".format(trials_to(45.0, 5.0, 0.05) / 365.0)
    )
    print(
        "time': not a platitude, but a consequence of how much evidence a"
        " confident"
    )
    print("prior is worth.")

    print("\nAsymmetry: what ONE occurrence does, after 200 clean trials.")
    print(
        "   prior         after 200 clean   then 1 event   change   trials to"
        " undo"
    )
    for a, b in [(5.0, 1.0), (18.0, 2.0), (45.0, 5.0)]:
        a1, b1 = posterior(a, b, 200)
        m1 = mean(a1, b1)
        a2, b2 = posterior(a1, b1, 0, 1)
        m2 = mean(a2, b2)
        back = 0
        while mean(a2, b2 + back) > m1 and back < 100000:
            back += 1
        print("   Beta({:4.0f},{:3.0f}) {:16.4f} {:14.4f} {:+8.4f} {:15d}"
              .format(a, b, m1, m2, m2 - m1, back))

    print(
        "\nThis is where the model FAILS to reproduce the phenomenon it is"
        " offered"
    )
    print(
        "to explain. One occurrence after 200 clean trials moves the"
        " posterior mean"
    )
    print(
        "by between +0.003 and +0.005, and between 5 and 41 further clean"
        " trials"
    )
    print(
        "undo it. A Bayesian learner of this kind cannot relapse: by"
        " construction,"
    )
    print(
        "200 observations outweigh 1. Yet sudden return of fear after a"
        " single bad"
    )
    print(
        "episode is among the best-attested findings in the extinction"
        " literature."
    )
    print(
        "So the arithmetic above is evidence AGAINST simple evidential"
        " updating as"
    )
    print(
        "the mechanism, not evidence for it. Something else - context,"
        " attribution,"
    )
    print(
        "or a separate inhibitory memory rather than a revised probability -"
        " has to"
    )
    print("carry the asymmetry, and this model has no term for it.")

    print(
        "\nWhat the model does NOT say: it assumes the trials are"
        " independent and"
    )
    print(
        "that the person updates on them. Both fail in the cases that matter"
        " most -"
    )
    print(
        "avoidance makes trials non-independent (the person attributes safety"
        " to the"
    )
    print(
        "avoidance, not to the absence of danger), and a belief maintained"
        " for other"
    )
    print(
        "reasons will not update at all. Those are the standard reasons a"
        " purely"
    )
    print(
        "evidential account of exposure is incomplete; the arithmetic above"
        " is an"
    )
    print(
        "upper bound on how fast belief could move, not a prediction that it"
        " does."
    )
# Prior belief that the sensation is followed by catastrophe, and the number
# of clean exposures needed to bring the posterior mean under a threshold.
#
#    prior Beta(a,b)   prior mean   to < 0.10   to < 0.05   to < 0.01
#    Beta(   1,  1)         0.500           8          18          98
#    Beta(   2,  1)         0.667          17          37         197
#    Beta(   5,  1)         0.833          44          94         494
#    Beta(   9,  1)         0.900          80         170         890
#    Beta(  18,  2)         0.900         160         340        1780
#    Beta(  45,  5)         0.900         400         850        4450
#
# A weakly held belief, Beta(2,1), is under 0.05 after 37 clean trials.
# A firmly held one, Beta(45,5), needs 850 - which at one exposure a day is
# about 2.3 years. This is the quantitative content of 'it takes a long
# time': not a platitude, but a consequence of how much evidence a confident
# prior is worth.
#
# Asymmetry: what ONE occurrence does, after 200 clean trials.
#    prior         after 200 clean   then 1 event   change   trials to undo
#    Beta(   5,  1)           0.0243         0.0290  +0.0047              41
#    Beta(  18,  2)           0.0818         0.0860  +0.0042              12
#    Beta(  45,  5)           0.1800         0.1833  +0.0033               5
#
# This is where the model FAILS to reproduce the phenomenon it is offered
# to explain. One occurrence after 200 clean trials moves the posterior mean
# by between +0.003 and +0.005, and between 5 and 41 further clean trials
# undo it. A Bayesian learner of this kind cannot relapse: by construction,
# 200 observations outweigh 1. Yet sudden return of fear after a single bad
# episode is among the best-attested findings in the extinction literature.
# So the arithmetic above is evidence AGAINST simple evidential updating as
# the mechanism, not evidence for it. Something else - context, attribution,
# or a separate inhibitory memory rather than a revised probability - has to
# carry the asymmetry, and this model has no term for it.
#
# What the model does NOT say: it assumes the trials are independent and
# that the person updates on them. Both fail in the cases that matter most -
# avoidance makes trials non-independent (the person attributes safety to the
# avoidance, not to the absence of danger), and a belief maintained for other
# reasons will not update at all. Those are the standard reasons a purely
# evidential account of exposure is incomplete; the arithmetic above is an
# upper bound on how fast belief could move, not a prediction that it does.

前半部分给出的是「要多久」的量化答案:弱持有的信念 Beta(2,1) 在 37 次干净试验后降到 0.05 以下,而牢固持有的 Beta(45,5) 需要 850 次,按每天一次约 2.3 年。这把「需要很长时间」从一句套话变成了一个数。

后半部分是本章的第二个否证结果。200 次干净试验之后的 1 次事件,只把后验均值推高 0.0033 到 0.0047,而 5 到 41 次干净试验即可抵消。也就是说,这样的贝叶斯学习者不可能复发——按构造,200 次观测压倒 1 次。但单次不良经历后恐惧迅速回复,是消退文献中证据最充分的现象之一 (Bouton 2002)。所以这段算术是反对单纯证据性更新作为机制的证据。承载这种不对称的必须是别的东西:情境、归因,或者一段独立的抑制性记忆而不是一个被修订的概率 (Bouton 2002; Craske et al. 2014)。

怎样测量一条不被解释的表象流

最后一个方法学问题。如果练习的要求是不解释内在画面,那么研究者怎样测量这条画面流而不要求受试去解释它?一个不要求语义判断的办法是:只记录类别的变化,然后算这条类别序列的转移熵——高熵对应流动的、不重复的流,低熵对应反复回到同一类的流。

这里的困难是统计的。短序列上的熵估计有系统性的向下偏差,Miller 很早就给出了偏差的一阶校正 (Miller 1955),Paninski 系统地分析了熵与互信息估计的性质 (Paninski 2003)。

推导

# Chapter 3. Measuring an imagery stream without interpreting it: transition
# entropy of a categorical sequence, and why the estimate is biased when
# sequences are short. A practitioner reports the category of each inner event
# at fixed intervals (K = 6 coarse categories). Two generators: a FREE stream
# (near-uniform transitions) and a STUCK stream (high probability of repeating
# the same category, the formal version of "going round the same thought").
# The plug-in entropy estimate is biased DOWNWARD, and the bias is of the same
# size as the effect being looked for, so the program applies the Miller-Madow
# correction and reports both. Synthetic data; fixed seed; standard library.
import math
import random

K = 6

def chain(stickiness):
    """Transition matrix with P(stay) = stickiness, rest spread uniformly."""
    off = (1.0 - stickiness) / (K - 1)
    return [
        [stickiness if i == j else off for j in range(K)] for i in range(K)
    ]

def sample(P, n, rng):
    s = rng.randrange(K)
    out = [s]
    for _ in range(n - 1):
        r, acc = rng.random(), 0.0
        for j in range(K):
            acc += P[s][j]
            if r <= acc:
                s = j
                break
        out.append(s)
    return out

def true_entropy_rate(P):
    """Conditional entropy H(X_{t+1} | X_t). The chain is symmetric, so its
    stationary
    distribution is uniform and every row has the same entropy: one row
    suffices.
    """
    return -math.fsum(p * math.log(p, 2) for p in P[0] if p > 0)

def plugin_rate(seq):
    """Plug-in conditional entropy H(X_{t+1} | X_t) in bits."""
    counts = [[0] * K for _ in range(K)]
    for a, b in zip(seq, seq[1:]):
        counts[a][b] += 1
    total = len(seq) - 1
    h = 0.0
    for i in range(K):
        ri = sum(counts[i])
        if ri == 0:
            continue
        for j in range(K):
            if counts[i][j]:
                p = counts[i][j] / ri
                h -= (ri / total) * p * math.log(p, 2)
    return h

def miller_madow(seq):
    """Plug-in plus (m - 1) / (2 N ln 2), m = number of observed transition
    cells."""
    counts = [[0] * K for _ in range(K)]
    for a, b in zip(seq, seq[1:]):
        counts[a][b] += 1
    total = len(seq) - 1
    nonzero_rows = sum(1 for i in range(K) if sum(counts[i]) > 0)
    m = sum(1 for i in range(K) for j in range(K) if counts[i][j] > 0)
    return plugin_rate(seq) + (m - nonzero_rows) / (2.0 * total * math.log(2))

if __name__ == "__main__":
    rng = random.Random(606)
    print(
        "K = {} categories. True conditional entropy from the generator,"
        " against".format(K)
    )
    print("estimates from sequences of the length a real session provides.\n")
    print("                             n=40          n=200         n=2000")
    print(
        "   stream  P(stay)  true H  plug-in  MM    plug-in  MM   "
        " plug-in  MM"
    )
    for lab, stay in (("free ", 1.0 / K), ("mild ", 0.35), ("stuck", 0.70)):
        P = chain(stay)
        th = true_entropy_rate(P)
        cells = []
        for n in (40, 200, 2000):
            pl = [plugin_rate(sample(P, n, rng)) for _ in range(300)]
            mm = [miller_madow(sample(P, n, rng)) for _ in range(300)]
            cells.append((math.fsum(pl) / len(pl), math.fsum(mm) / len(mm)))
        print(
            "   {}  {:7.3f} {:7.3f} {:8.3f} {:6.3f} {:8.3f} {:6.3f}"
            " {:8.3f} {:6.3f}".format(
                lab,
                stay,
                th,
                cells[0][0],
                cells[0][1],
                cells[1][0],
                cells[1][1],
                cells[2][0],
                cells[2][1],
            )
        )

    Pf, Pm = chain(1.0 / K), chain(0.35)
    hf, hm = true_entropy_rate(Pf), true_entropy_rate(Pm)
    pl40 = (
        math.fsum(plugin_rate(sample(Pf, 40, rng)) for _ in range(300)) / 300
    )
    mm40 = (
        math.fsum(miller_madow(sample(Pf, 40, rng)) for _ in range(300)) / 300
    )
    print(
        "\nAt n = 40, the length of a 20-minute session sampled every 30"
        " s, the"
    )
    print(
        "plug-in estimate understates the free stream by {:.2f} bits, and the"
        .format(hf - pl40)
    )
    print(
        "Miller-Madow correction cuts that to {:.2f} bits. For comparison,"
        " the whole".format(hf - mm40)
    )
    print(
        "true difference between the free and the mild stream is {:.2f}"
        " bits: the".format(hf - hm)
    )
    print(
        "uncorrected bias is about {:.0f} times the effect being looked"
        " for, and".format((hf - pl40) / (hf - hm))
    )
    print(
        "even the corrected bias is about {:.0f} times it. An uncorrected"
        " estimator".format((hf - mm40) / (hf - hm))
    )
    print(
        "would report differences in sequence LENGTH as differences in the"
        " stream."
    )

    print(
        "\nCan a session-length sequence tell the two streams apart?"
        " Distribution of"
    )
    print("the Miller-Madow estimate, 2000 sessions of each, n = 40:")
    for lab, stay in (("free ", 1.0 / K), ("mild ", 0.35)):
        P = chain(stay)
        v = sorted(miller_madow(sample(P, 40, rng)) for _ in range(2000))
        print("   {}: median {:.3f}, 5th-95th percentile {:.3f} to {:.3f}"
              .format(lab, v[1000], v[100], v[1900]))
    vf = sorted(miller_madow(sample(Pf, 40, rng)) for _ in range(2000))
    vm = sorted(miller_madow(sample(Pm, 40, rng)) for _ in range(2000))
    import bisect
    # P(a random mild session scores below a random free session) = the AUC
    auc = math.fsum(bisect.bisect_left(vm, x) for x in vf) / (
        len(vf) * len(vm)
    )
    print("   P(a mild session scores below a free one) = {:.3f}".format(auc))
    print("   (0.5 would be no discrimination at all, 1.0 perfect)")
    n_needed = None
    for n_sessions in range(1, 61):
        # SE of the per-person mean over n sessions, pooled SD from the two
        # arms
        sd = math.sqrt(
            0.5
            * (
                sum((x - vf[1000]) ** 2 for x in vf) / len(vf)
                + sum((x - vm[1000]) ** 2 for x in vm) / len(vm)
            )
        )
        if abs(vf[1000] - vm[1000]) / (sd / math.sqrt(n_sessions)) >= 2.8:
            n_needed = n_sessions
            break
    print(
        "\n   A single session discriminates at {:.2f}, which is far too"
        " weak to".format(auc)
    )
    print(
        "   classify one person's stream. Averaging over sessions per person,"
        " about"
    )
    print(
        "   {} sessions are needed before the person-level means of the two"
        " streams".format(n_needed)
    )
    print(
        "   are separated by 2.8 standard errors. That is why the Chapter 8"
        " protocol"
    )
    print(
        "   pre-specifies a per-person average over sessions and the"
        " Miller-Madow"
    )
    print(
        "   estimator, and records the sampling interval, which fixes n and"
        " so fixes"
    )
    print(
        "   the bias: an arm that reports more events per session would"
        " otherwise"
    )
    print("   score as having a higher-entropy stream for that reason alone.")
# K = 6 categories. True conditional entropy from the generator, against
# estimates from sequences of the length a real session provides.
#
#                              n=40          n=200         n=2000
#    stream  P(stay)  true H  plug-in  MM    plug-in  MM    plug-in  MM
#    free     0.167   2.585    1.928  2.265    2.472  2.579    2.574  2.585
#    mild     0.350   2.443    1.793  2.093    2.329  2.438    2.433  2.444
#    stuck    0.700   1.578    1.062  1.217    1.456  1.555    1.571  1.578
#
# At n = 40, the length of a 20-minute session sampled every 30 s, the
# plug-in estimate understates the free stream by 0.65 bits, and the
# Miller-Madow correction cuts that to 0.34 bits. For comparison, the whole
# true difference between the free and the mild stream is 0.14 bits: the
# uncorrected bias is about 5 times the effect being looked for, and
# even the corrected bias is about 2 times it. An uncorrected estimator
# would report differences in sequence LENGTH as differences in the stream.
#
# Can a session-length sequence tell the two streams apart? Distribution of
# the Miller-Madow estimate, 2000 sessions of each, n = 40:
#    free : median 2.267, 5th-95th percentile 2.005 to 2.500
#    mild : median 2.108, 5th-95th percentile 1.785 to 2.396
#    P(a mild session scores below a free one) = 0.741
#    (0.5 would be no discrimination at all, 1.0 perfect)
#
#    A single session discriminates at 0.74, which is far too weak to
#    classify one person's stream. Averaging over sessions per person, about
#    9 sessions are needed before the person-level means of the two streams
#    are separated by 2.8 standard errors. That is why the Chapter 8 protocol
#    pre-specifies a per-person average over sessions and the Miller-Madow
#    estimator, and records the sampling interval, which fixes n and so fixes
#    the bias: an arm that reports more events per session would otherwise
#    score as having a higher-entropy stream for that reason alone.

量级值得记下:在 n=40n = 40(20 分钟、每 30 秒取样一次)时,朴素的插入式估计把自由流的熵低估 0.65 比特,Miller—Madow 校正把偏差降到 0.34 比特;而自由流与「轻度黏着」流之间的真实差别只有 0.14 比特。也就是说,未校正的偏差约为所求效应的 5 倍,校正后仍约 2 倍。更要紧的是:偏差随序列长度变化,所以一个报告了更多事件的组会仅因此被估为熵更高。

单次练习的判别力只有 0.74(随机为 0.5),远不足以给一个人的流分类;按个人均值平均,约需 9 次练习才能使两种流的个人均值相差 2.8 个标准误。第八章据此规定:以个人在多次练习上的平均值为单位、使用 Miller—Madow 估计量、并记录取样间隔。

假说 · 尚未验证

作者的假说(尚未确立): 以「不解释画面」为训练目标的练习,会提高自评表象流的转移熵(流动性),而这一变化先于身体警觉分数的下降。什么样的结果能证明它是错的: 在事先登记的试验中,若练习组与结构匹配对照组在第 2 周的表象流转移熵(按 Miller—Madow 估计、个人均值、取样间隔固定)上的差别的 95% 置信区间包含零,或其变化不先于主要结局的变化,则本假说错。

注

表象不是所有人都有的。 视觉表象的生动程度在人群中差异极大,有人完全没有自发的视觉表象,即先天性心盲 (Zeman, Dewar, and Della Sala 2015);Pearson 对表象的神经科学作了综述,并讨论了表象强度的个体差异与其测量 (Pearson 2019)。Dijkstra 等人综述了表象与知觉共享的神经机制 (Dijkstra, Bosch, and Gerven 2019)。以「观看内在画面」为核心的练习对无表象者不适用;任何推广这类练习的材料都应写明这一点,第八章把表象生动度作为预先指定的分层变量与效应修饰检验。

本章小结

把「不解释内在画面」写成信号检测模型中的增益项之后,判别力 d′d' 与增益无关:判准重新调定时,各增益的期望代价完全相同(0.3857);判准不重调时,增益 0.5 把命中率从 0.622 压到 0.281、把代价从 0.3857 抬到 0.4634。所以「无评判」交换的是错误类型,不是信息量;它的好处来自原来的操作点偏了,这正是惊恐认知模型所描述的情形。情绪标注的文献给出方向相反而同样有支持的做法。双过程模型显示,间隔几乎不改变学到了多少,而改变课末测量到的快过程,所以主要测量应放在另一天。Beta—伯努利更新给出了「要多久」的数(牢固信念约 850 次干净试验),同时无法解释复发,因此是反对单纯证据性更新的证据。最后,表象流的转移熵在练习长度的序列上偏差约为效应的 2 到 5 倍,必须校正、固定取样间隔,并按个人均值分析。

参考文献

Bishop, Scott R., Mark Lau, Shauna Shapiro, Linda Carlson, Nicole D. Anderson, James Carmody, Zindel V. Segal, et al. 2004. “Mindfulness: A Proposed Operational Definition.” Clinical Psychology: Science and Practice 11 (3): 230–41. https://doi.org/10.1093/clipsy.bph077.
Boettcher, Hannah, C. Alex Brake, and David H. Barlow. 2016. “Origins and Outlook of Interoceptive Exposure.” Journal of Behavior Therapy and Experimental Psychiatry 53: 41–51. https://doi.org/10.1016/j.jbtep.2015.10.009.
Bouton, Mark E. 2002. “Context, Ambiguity, and Unlearning: Sources of Relapse After Behavioral Extinction.” Biological Psychiatry 52 (10): 976–86. https://doi.org/10.1016/S0006-3223(02)01546-9.
Clark, David M. 1986. “A Cognitive Approach to Panic.” Behaviour Research and Therapy 24 (4): 461–70. https://doi.org/10.1016/0005-7967(86)90011-2.
Craske, Michelle G., Michael Treanor, Christopher C. Conway, Tomislav Zbozinek, and Bram Vervliet. 2014. “Maximizing Exposure Therapy: An Inhibitory Learning Approach.” Behaviour Research and Therapy 58: 10–23. https://doi.org/10.1016/j.brat.2014.04.006.
Dijkstra, Nadine, Sander E. Bosch, and Marcel A. J. van Gerven. 2019. “Shared Neural Mechanisms of Visual Perception and Imagery.” Trends in Cognitive Sciences 23 (5): 423–34. https://doi.org/10.1016/j.tics.2019.02.004.
Groves, Philip M., and Richard F. Thompson. 1970. “Habituation: A Dual-Process Theory.” Psychological Review 77 (5): 419–50. https://doi.org/10.1037/h0029810.
Hölzel, Britta K., Sara W. Lazar, Tim Gard, Zev Schuman-Olivier, David R. Vago, and Ulrich Ott. 2011. “How Does Mindfulness Meditation Work? Proposing Mechanisms of Action from a Conceptual and Neural Perspective.” Perspectives on Psychological Science 6 (6): 537–59. https://doi.org/10.1177/1745691611419671.
Lieberman, Matthew D., Naomi I. Eisenberger, Molly J. Crockett, Sabrina M. Tom, Jennifer H. Pfeifer, and Baldwin M. Way. 2007. “Putting Feelings into Words: Affect Labeling Disrupts Amygdala Activity in Response to Affective Stimuli.” Psychological Science 18 (5): 421–28. https://doi.org/10.1111/j.1467-9280.2007.01916.x.
Miller, George A. 1955. “Note on the Bias of Information Estimates.” In Information Theory in Psychology: Problems and Methods, edited by Henry Quastler, 95–100. Glencoe, Illinois: Free Press.
Paninski, Liam. 2003. “Estimation of Entropy and Mutual Information.” Neural Computation 15 (6): 1191–1253. https://doi.org/10.1162/089976603321780272.
Pearson, Joel. 2019. “The Human Imagination: The Cognitive Neuroscience of Visual Mental Imagery.” Nature Reviews Neuroscience 20: 624–34. https://doi.org/10.1038/s41583-019-0202-9.
Rankin, Catharine H., Thomas Abrams, Robert J. Barry, Seema Bhatnagar, David F. Clayton, John Colombo, Gianluca Coppola, et al. 2009. “Habituation Revisited: An Updated and Revised Description of the Behavioral Characteristics of Habituation.” Neurobiology of Learning and Memory 92 (2): 135–38. https://doi.org/10.1016/j.nlm.2008.09.012.
Torre, Jared B., and Matthew D. Lieberman. 2018. “Putting Feelings into Words: Affect Labeling as Implicit Emotion Regulation.” Emotion Review 10 (2): 116–24. https://doi.org/10.1177/1754073917742706.
Zeman, Adam, Michaela Dewar, and Sergio Della Sala. 2015. “Lives Without Imagery — Congenital Aphantasia.” Cortex 73: 378–80. https://doi.org/10.1016/j.cortex.2015.05.019.

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).