5D Library › 具身认知、伯恩斯坦自由度问题与运动协同的控制理论 · English

第五章 肌肉协同:非负矩阵分解与生物力学约束之争

从肌电到协同

关节之下还有肌肉。如果神经系统真要逐块控制几十块肌肉,问题的规模比控制关节还大。一个流行的假说是:神经系统并不逐块控制,而是调用少数几个肌肉协同——每个协同是一组按固定比例共同激活的肌肉,动作由这几个协同按不同的时程叠加而成 (Bizzi et al. 2008)。

最简单的形式是同步协同:记录 mm 块肌肉在 nn 个时刻的肌电包络,排成非负矩阵 VV(m×nm \times n),用 kk 个协同近似 V≈WH,W≥0(m×k),H≥0(k×n),V \approx W H,\qquad W \ge 0\ (m \times k),\ H \ge 0\ (k \times n), WW 的每一列是一个协同(各肌肉的权重),HH 的每一行是它随时间的激活。非负的约束来自生理:肌肉只能拉,不能推,肌电包络也不为负。李与承 1999 年在《自然》上用这种非负矩阵分解(NMF)把人脸图像分解为「部件」(Lee and Seung 1999),2001 年给出了简单的乘性更新算法 (Lee and Seung 2001)。

证据

蛙的踢腿。 达韦拉、萨尔蒂尔与比齐 2003 年记录了完整(未去传入)的蛙在向不同方向踢腿时多块后肢肌肉的肌电,用时变协同分析:每个协同是一段肌肉激活的时空模式,可以在每次动作中按一个幅度缩放、按一个时间平移,几个协同相加重建肌电;提取算法是非负矩阵分解的推广 (d’Avella et al. 2003)。达韦拉与比齐 2005 年在牛蛙 13 块后肢肌肉的跳跃、游泳与行走中发现,一些协同为几种行为共用,另一些为特定行为专有 (d’Avella and Bizzi 2005)。张等人 2005 年比较同一批蛙在去传入前后提取的协同,用来区分中枢与感觉反馈对协同的贡献 (Cheung et al. 2005)。在猫的姿势任务中,廷与麦克弗森 2005 年发现少数几个协同就能解释对支撑面平移的肌肉反应与力的产生 (Ting and Macpherson 2005)。

框架把达韦拉等人 2003 年的工作与 NMF 直接联系起来。需要说明的是,那项研究用的是时变协同及其专门的算法,而标准的李—承 NMF 提取的是同步协同;两者是同一思想的不同模型,结论不能直接互换。

VAF 阈值的问题

用 NMF 提取协同,最常见的问题是「取几个」。常用的做法是看解释方差(VAF,1−∥V−WH∥2/∥V∥21 – \|V – WH\|^2/\|V\|^2,注意分母是未中心化的平方和),选使 VAF 超过某个阈值(如 90%)的最小 kk,或者选 VAF 曲线的「拐点」。特雷施、张与达韦拉 2006 年在模拟与实验数据上比较了包括 NMF 在内的多种矩阵分解算法识别协同的表现 (Tresch et al. 2006)。下面的程序在一个已知答案的例子上检查这些规则。

推导

程序生成 12 块肌肉 × 120 个时刻的合成肌电:4 个已知的非负协同乘以平滑的非负激活,再加上标准差为信号均值 10% 的噪声(负值截为零)。对 k=1k = 1 到 6,用李—承乘性更新(150 次迭代,3 个随机起点取最好)求 WW、HH 与 VAF;对 k=4k = 4,把估计的协同与真实协同按最佳排列配对,计算余弦相似度。

# Synchronous muscle synergies by non-negative matrix factorisation (Lee & Seung's
# multiplicative updates, Frobenius cost). Synthetic EMG: 12 muscles x 120 samples
# built from 4 known non-negative synergies with smooth non-negative activations, plus
# rectified noise (SD 10 % of the mean signal). For k = 1..6: variance accounted for
# (VAF, uncentred), best of 3 random starts; for k = 4: recovery of the true synergies.
import math, random
from itertools import permutations

def nmf(V, k, rng, iters=150):
    m, n = len(V), len(V[0]); eps = 1e-12
    W = [[rng.random() for _ in range(k)] for _ in range(m)]
    H = [[rng.random() for _ in range(n)] for _ in range(k)]
    for _ in range(iters):
        WtV = [[sum(W[i][a] * V[i][j] for i in range(m)) for j in range(n)]
               for a in range(k)]
        WtW = [[sum(W[i][a] * W[i][b] for i in range(m)) for b in range(k)]
               for a in range(k)]
        H = [[H[a][j] * WtV[a][j] / (sum(WtW[a][b] * H[b][j] for b in range(k)) + eps)
              for j in range(n)] for a in range(k)]
        VHt = [[sum(V[i][j] * H[a][j] for j in range(n)) for a in range(k)]
               for i in range(m)]
        HHt = [[sum(H[a][j] * H[b][j] for j in range(n)) for b in range(k)]
               for a in range(k)]
        W = [[W[i][a] * VHt[i][a] / (sum(W[i][b] * HHt[b][a] for b in range(k)) + eps)
              for a in range(k)] for i in range(m)]
    return W, H

def vaf(V, W, H):
    k = len(H); sse = sst = 0.0
    for i, row in enumerate(V):
        for j, v in enumerate(row):
            r = v - sum(W[i][a] * H[a][j] for a in range(k)); sse += r * r; sst += v * v
    return 1 - sse / sst

def best_fit(V, k, rng, starts=3):
    return max((nmf(V, k, rng) for _ in range(starts)), key=lambda f: vaf(V, *f))

def unit(v):
    s = math.sqrt(sum(x * x for x in v)); return [x / s for x in v]

def synthetic(rng, m=12, n=120, k=4, noise=0.10):
    W = [[max(0.0, rng.gauss(0, 1)) ** 2 for _ in range(k)] for _ in range(m)]
    H = []
    for a in range(k):                           # smooth bumps of activation
        c = [rng.uniform(0, n) for _ in range(3)]
        H.append([sum(math.exp(-0.5 * ((j - ci) / 8) ** 2) for ci in c)
                  for j in range(n)])
    V = [[sum(W[i][a] * H[a][j] for a in range(k)) for j in range(n)] for i in range(m)]
    mean = sum(map(sum, V)) / (m * n)
    return W, [[max(0.0, v + rng.gauss(0, noise * mean)) for v in row] for row in V]

if __name__ == "__main__":
    rng = random.Random(2003); Wtrue, V = synthetic(rng)
    print("k   VAF      (synthetic EMG generated from 4 synergies + 10 % noise)")
    fits = {}
    for k in range(1, 7):
        fits[k] = best_fit(V, k, rng); print(f"{k}   {vaf(V, *fits[k]):.4f}")
    pick = next(k for k in range(1, 7) if vaf(V, *fits[k]) >= 0.90)
    pick95 = next(k for k in range(1, 7) if vaf(V, *fits[k]) >= 0.95)
    gain = next(k for k in range(1, 6)
                if vaf(V, *fits[k + 1]) - vaf(V, *fits[k]) < 0.01)
    print(f"smallest k with VAF >= 0.90: {pick}; >= 0.95: {pick95};"
          f" next synergy adds < 1 point: {gain}")
    W = fits[4][0]
    est = [unit([W[i][a] for i in range(12)]) for a in range(4)]
    true = [unit([Wtrue[i][a] for i in range(12)]) for a in range(4)]
    def score(p):
        return sum(sum(x * y for x, y in zip(true[a], est[p[a]])) for a in range(4))
    perm = max(permutations(range(4)), key=score)
    sims = [sum(x * y for x, y in zip(true[a], est[perm[a]])) for a in range(4)]
    print("k = 4: cosine similarity of each true synergy with its best-matched estimate:")
    print("  " + "  ".join(f"{s:.3f}" for s in sims))
# k   VAF      (synthetic EMG generated from 4 synergies + 10 % noise)
# 1   0.8196
# 2   0.9653
# 3   0.9937
# 4   0.9980
# 5   0.9980
# 6   0.9985
# smallest k with VAF >= 0.90: 2; >= 0.95: 2; next synergy adds < 1 point: 3
# k = 4: cosine similarity of each true synergy with its best-matched estimate:
#   0.992  0.925  0.999  0.893

真实的协同数是 4,但 VAF 在 k=1k = 1 时已是 0.820,k=2k = 2 时 0.965。按「VAF ≥ 0.90」或「≥ 0.95」的规则都只选 2 个;按「再加一个协同使 VAF 增加不到 1 个百分点」的规则选 3 个。只有把阈值提高到约 0.995 才能选出 4 个,而这时阈值的选择又取决于噪声水平。k=4k = 4 时,估计的协同与真实协同的余弦相似度为 0.992、0.925、0.999、0.893:形状大体找回来了,但并不完美。这说明,VAF 阈值给出的协同数主要反映阈值本身与噪声水平,不能直接读成神经系统使用的模块数。图五左还画出把每块肌肉在时间上打乱后的数据:它的 VAF 曲线同样很高,只比原数据略低,这是未中心化 VAF 的基线问题,下文程序 p09 会再遇到。

协同是不是神经的

低维的肌电是神经协同的证据,还是别的什么的产物?特雷施与雅尔克 2009 年在《神经生物学新见》上专门讨论了支持与反对肌肉协同的论据 (Tresch and Jarc 2009)。其中最尖锐的反对意见来自生物力学。

证据

库奇与巴列罗—奎瓦斯 2012 年。 他们指出,即使每块肌肉都由神经系统独立控制,生物力学与任务的约束本身也会使肌电变成低维的 (Kutch and Valero-Cuevas 2012)。两类约束分开讨论:在运动任务中,关节数远少于肌肉数,肢体运动只能在一个低维子空间中改变肌肉长度,如果每块肌肉都独立地抵抗被拉长,模拟的肌电也是低维的;在力的任务中,如果实验只在固定大小的力下改变方向,可行的肌肉激活模式就被限制在一个低维集合中。在离体手指(7 条肌腱)的力任务中,2 个「协同」就解释了 80% 以上的变异;在 14 块肌肉的腿部模型中则需要 7 个。他们明确说明,这些结果没有否定神经协同假说,并提出一个流程:先判断任务以运动还是力为主,再用生物力学模型检验观察到的低维性能否由约束解释;只有在采样了完整的可行力集合(所有方向与大小)之后仍然存在的低维性,才可以归于神经系统。

证据

记录了哪些肌肉也有影响。 斯蒂尔、特雷施与佩罗 2013 年用一个 30 块肌肉的上肢模型模拟等长用力任务,发现只记录少数肌肉时,提取出的协同与用全部肌肉提取的协同差别很大:只用 5 块肌肉时归一化相似度平均为 0.57,约 11 块以上才超过 0.8;VAF 反而被高估(5 块肌肉时平均 0.98,全部 30 块时 0.88),所以只看 VAF 会选出过少的协同 (Steele et al. 2013)。

下面的程序按库奇与巴列罗—奎瓦斯的思路构造一个「没有任何神经协同」的例子:8 块平面肌肉,拉力方向与最大力固定,在 16 个方向上产生大小固定为 0.5 的端点力(每个方向 10 次)。每块肌肉都独立控制,激活模式有两种来源:(a) 最小努力解(使激活平方和最小)加 5% 噪声;(b) 在全部可行解中随机抽样(命中—奔跑抽样)。作为基准,把每块肌肉的数据在试次之间独立打乱(破坏任务结构),以及只「记录」8 块中的 5 块。

推导

# Can "synergies" appear without any neural synergy? (After Kutch & Valero-Cuevas 2012.)
# Eight planar muscles with fixed pulling directions and strengths produce an isometric
# endpoint force of fixed magnitude 0.5 in 16 directions (10 trials each). Every muscle
# is controlled INDEPENDENTLY. (a) minimum-effort activations + 5 % noise; (b) random
# feasible activations (hit-and-run sampling of the whole solution set, 0 <= a <= 1).
# Baselines: each muscle shuffled across trials (task structure destroyed), and data
# recorded from only 5 of the 8 muscles. NMF code as in p08.
import math, random
ANG = (10, 55, 100, 150, 200, 240, 290, 330); FMAX = (1.0, .8, .9, .7, 1.0, .6, .8, .9)
MM = [[f * math.cos(math.radians(a)) for a, f in zip(ANG, FMAX)],
      [f * math.sin(math.radians(a)) for a, f in zip(ANG, FMAX)]]

def nmf(V, k, rng, iters=150):
    m, n = len(V), len(V[0]); eps = 1e-12
    W = [[rng.random() for _ in range(k)] for _ in range(m)]
    H = [[rng.random() for _ in range(n)] for _ in range(k)]
    for _ in range(iters):
        WtV = [[sum(W[i][a] * V[i][j] for i in range(m)) for j in range(n)]
               for a in range(k)]
        WtW = [[sum(W[i][a] * W[i][b] for i in range(m)) for b in range(k)]
               for a in range(k)]
        H = [[H[a][j] * WtV[a][j] / (sum(WtW[a][b] * H[b][j] for b in range(k)) + eps)
              for j in range(n)] for a in range(k)]
        VHt = [[sum(V[i][j] * H[a][j] for j in range(n)) for a in range(k)]
               for i in range(m)]
        HHt = [[sum(H[a][j] * H[b][j] for j in range(n)) for b in range(k)]
               for a in range(k)]
        W = [[W[i][a] * VHt[i][a] / (sum(W[i][b] * HHt[b][a] for b in range(k)) + eps)
              for a in range(k)] for i in range(m)]
    sse = sum((V[i][j] - sum(W[i][a] * H[a][j] for a in range(k))) ** 2
              for i in range(m) for j in range(n))
    return 1 - sse / sum(v * v for row in V for v in row)

def min_effort(f, lo=0.0, hi=1.0):           # min |a|^2 s.t. M a = f, lo <= a <= hi
    lam = [f[0] * 2, f[1] * 2]
    for _ in range(60):
        z = [MM[0][i] * lam[0] + MM[1][i] * lam[1] for i in range(8)]
        a = [min(hi, max(lo, v)) for v in z]
        act = [lo < v < hi for v in z]
        r = [sum(MM[k][i] * a[i] for i in range(8)) - f[k] for k in range(2)]
        J = [[sum(MM[p][i] * MM[q][i] for i in range(8) if act[i]) for q in range(2)]
             for p in range(2)]
        det = J[0][0] * J[1][1] - J[0][1] * J[1][0]
        lam = [lam[0] - (J[1][1] * r[0] - J[0][1] * r[1]) / det,
               lam[1] - (-J[1][0] * r[0] + J[0][0] * r[1]) / det]
    return a

def null_basis():
    G = [[sum(MM[p][i] * MM[q][i] for i in range(8)) for q in range(2)] for p in range(2)]
    d = G[0][0] * G[1][1] - G[0][1] ** 2
    Gi = [[G[1][1] / d, -G[0][1] / d], [-G[1][0] / d, G[0][0] / d]]
    P = [[float(i == j) - sum(MM[p][i] * Gi[p][q] * MM[q][j] for p in range(2)
                              for q in range(2)) for j in range(8)] for i in range(8)]
    basis = []
    for col in zip(*P):
        v = list(col)
        for b in basis:
            c = sum(x * y for x, y in zip(v, b)); v = [x - c * y for x, y in zip(v, b)]
        s = math.sqrt(sum(x * x for x in v))
        if s > 1e-8:
            basis.append([x / s for x in v])
    return basis

def hit_and_run(a, basis, rng, steps):
    for _ in range(steps):
        z = [rng.gauss(0, 1) for _ in basis]
        d = [sum(zk * b[i] for zk, b in zip(z, basis)) for i in range(8)]
        lo, hi = -math.inf, math.inf
        for ai, di in zip(a, d):
            if abs(di) > 1e-12:
                t1, t2 = (0 - ai) / di, (1 - ai) / di
                lo, hi = max(lo, min(t1, t2)), min(hi, max(t1, t2))
        t = rng.uniform(lo, hi); a = [ai + t * di for ai, di in zip(a, d)]
    return a

if __name__ == "__main__":
    rng = random.Random(2012); basis = null_basis()
    dirs = [2 * math.pi * j / 16 for j in range(16)]
    me, rnd = [], []
    for th in dirs:
        f = [0.5 * math.cos(th), 0.5 * math.sin(th)]
        a0 = min_effort(f)
        a = min_effort(f, 0.02, 0.98)                   # interior start point
        for _ in range(10):
            me.append([max(0.0, v * (1 + rng.gauss(0, 0.05))) for v in a0])
            a = hit_and_run(a, basis, rng, 25); rnd.append(a)
    def shuffled(D):
        return [list(c) for c in zip(*[rng.sample([r[i] for r in D], len(D))
                                       for i in range(8)])]
    sets = {"(a) minimum effort": me, "(a) shuffled": shuffled(me),
            "(a) 5 muscles only": [r[:5] for r in me], "(b) random feasible": rnd,
            "(b) shuffled": shuffled(rnd)}
    print("VAF by number of NMF components k (best of 2 starts), 160 trials")
    print("data                      k=1     k=2     k=3     k=4")
    for name, D in sets.items():
        V = [list(c) for c in zip(*D)]
        v = [max(nmf(V, k, rng) for _ in range(2)) for k in (1, 2, 3, 4)]
        print(f"{name:22s}  " + "  ".join(f"{x:.3f}" for x in v))
    print(f"solution set per force: 8 muscles - 2 force components ="
          f" {len(basis)} dimensions")
# VAF by number of NMF components k (best of 2 starts), 160 trials
# data                      k=1     k=2     k=3     k=4
# (a) minimum effort      0.424  0.693  0.853  0.950
# (a) shuffled            0.483  0.594  0.689  0.778
# (a) 5 muscles only      0.551  0.816  0.959  0.992
# (b) random feasible     0.793  0.837  0.869  0.899
# (b) shuffled            0.784  0.825  0.860  0.892
# solution set per force: 8 muscles - 2 force components = 6 dimensions

最小努力的数据中,3 个成分解释 85.3%,4 个解释 95.0%:如果按 90% 的阈值,会报告「4 个协同」,而这 8 块肌肉事实上是独立控制的。低维性来自任务(固定大小的力、16 个方向)加上代价函数(最小努力);打乱之后,4 个成分只解释 77.8%。只记录 5 块肌肉时,3 个成分就解释 95.9%——与斯蒂尔等人的发现一致,少记录几块肌肉会让数据显得更「协同」。随机可行解的情形则不同:每个力对应一个 6 维的解集,随机抽样的数据与打乱后的数据几乎一样(4 个成分 0.899 对 0.892),也就是说,单靠力学约束不一定产生低维结构,是约束加上某种选择规则(如最小努力)产生了它。还要注意未中心化 VAF 的基线很高:打乱的随机数据 1 个成分就有 0.784。

注

这些程序说明了什么。 它们不说明肌肉协同不存在,只说明:VAF 达到某个阈值、协同数小于肌肉数,这两件事本身不足以证明神经系统使用固定的协同模块。要论证神经起源,需要排除任务约束与优化代价的解释,例如采样完整的可行力集合 (Kutch and Valero-Cuevas 2012)、比较去传入前后的结构 (Cheung et al. 2005),或者检验协同结构在损伤后的变化。克拉克等人 2010 年发现,脑卒中后患侧下肢在行走中可提取的模块数减少,常表现为健康人的几个模块合并,模块数越少行走表现越差 (Clark et al. 2010);这类临床差异同样需要在控制了生物力学与速度等因素之后解读。

假说 · 尚未验证 · 协同是对代价的解,不是固定的模块

作者的诠释:至少在一部分任务中,「协同」可能是神经系统对特定任务与代价(如努力最小)求解的结果,而不是事先存储、只能按比例调用的固定模块。可以区分两者的预测: 如果协同是固定模块,改变任务约束(例如从固定大小的力改为大小也变化的力、或者改变肢体姿势)时,协同的肌肉组成应当保持不变,只改变激活时程;如果协同是代价的解,组成应当随约束系统地改变,并且能被生物力学模型预先算出。能推翻它的结果: 在采样了完整可行力集合、并改变姿势的实验中,同一组固定协同仍以高相似度重现,而生物力学最优解预测的组成变化没有出现。

本章小结

肌肉协同假说认为动作由少数几组按固定比例激活的肌肉叠加而成,常用非负矩阵分解从肌电中提取;达韦拉等人 2003 年分析的是蛙踢腿的时变协同,用的是 NMF 的推广算法。VAF 阈值给出的协同数主要反映阈值与噪声:真有 4 个协同时,90% 阈值只选出 2 个。库奇与巴列罗—奎瓦斯说明,任务与生物力学约束本身就能产生低维肌电:8 块独立控制的肌肉在最小努力解下,4 个成分解释 95.0%;只记录 5 块肌肉时 3 个成分解释 95.9%。这些结果不否定神经协同,但要求在下结论之前排除约束与代价的解释。下一章转向学习:练习怎样改变自由度的使用,注意的方向又有什么影响。

参考文献

Bizzi, Emilio, Vincent C. K. Cheung, Andrea d’Avella, Philippe Saltiel, and Matthew Tresch. 2008. “Combining Modules for Movement.” Brain Research Reviews 57 (1): 125–33. https://doi.org/10.1016/j.brainresrev.2007.08.004.
Cheung, Vincent C. K., Andrea d’Avella, Matthew C. Tresch, and Emilio Bizzi. 2005. “Central and Sensory Contributions to the Activation and Organization of Muscle Synergies During Natural Motor Behaviors.” Journal of Neuroscience 25 (27): 6419–34. https://doi.org/10.1523/JNEUROSCI.4904-04.2005.
Clark, David J., Lena H. Ting, Felix E. Zajac, Richard R. Neptune, and Steven A. Kautz. 2010. “Merging of Healthy Motor Modules Predicts Reduced Locomotor Performance and Muscle Coordination Complexity Post-Stroke.” Journal of Neurophysiology 103 (2): 844–57. https://doi.org/10.1152/jn.00825.2009.
d’Avella, Andrea, and Emilio Bizzi. 2005. “Shared and Specific Muscle Synergies in Natural Motor Behaviors.” Proceedings of the National Academy of Sciences of the United States of America 102 (8): 3076–81. https://doi.org/10.1073/pnas.0500199102.
d’Avella, Andrea, Philippe Saltiel, and Emilio Bizzi. 2003. “Combinations of Muscle Synergies in the Construction of a Natural Motor Behavior.” Nature Neuroscience 6 (3): 300–308. https://doi.org/10.1038/nn1010.
Kutch, Jason J., and Francisco J. Valero-Cuevas. 2012. “Challenges and New Approaches to Proving the Existence of Muscle Synergies of Neural Origin.” PLoS Computational Biology 8 (5): e1002434. https://doi.org/10.1371/journal.pcbi.1002434.
Lee, Daniel D., and H. Sebastian Seung. 1999. “Learning the Parts of Objects by Non-Negative Matrix Factorization.” Nature 401 (6755): 788–91. https://doi.org/10.1038/44565.
Lee, Daniel D., and H. Sebastian Seung. 2001. “Algorithms for Non-Negative Matrix Factorization.” In Advances in Neural Information Processing Systems 13, edited by Todd K. Leen, Thomas G. Dietterich, and Volker Tresp. MIT Press.
Steele, Katherine M., Matthew C. Tresch, and Eric J. Perreault. 2013. “The Number and Choice of Muscles Impact the Results of Muscle Synergy Analyses.” Frontiers in Computational Neuroscience 7: 105. https://doi.org/10.3389/fncom.2013.00105.
Ting, Lena H., and Jane M. Macpherson. 2005. “A Limited Set of Muscle Synergies for Force Control During a Postural Task.” Journal of Neurophysiology 93 (1): 609–13. https://doi.org/10.1152/jn.00681.2004.
Tresch, Matthew C., Vincent C. K. Cheung, and Andrea d’Avella. 2006. “Matrix Factorization Algorithms for the Identification of Muscle Synergies: Evaluation on Simulated and Experimental Data Sets.” Journal of Neurophysiology 95 (4): 2199–212. https://doi.org/10.1152/jn.00222.2005.
Tresch, Matthew C., and Anthony Jarc. 2009. “The Case for and Against Muscle Synergies.” Current Opinion in Neurobiology 19 (6): 601–7. https://doi.org/10.1016/j.conb.2009.09.002.

Fan, Y. (2026). Embodied Cognition, Bernstein's Degrees-of-Freedom Problem and the Control Theory of Motor Synergies: From the Uncontrolled Manifold Hypothesis and Internal Forward Models to Fitts's Law, Read Beside Classical Chinese Thought on Skill and the Way [具身认知、伯恩斯坦自由度问题与运动协同的控制理论] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23828057 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).