关于「工作记忆的瓶颈:组块、图式与认知负荷」,《注意力网络与内在动机》给出的定量结论是什么?
结论:工作记忆的容量不是「7 ± 2」,而是在排除复述与组块化之后约 3 到 5 个组块;米勒本人强调的正是组块与比特的区别。专家的优势主要来自长时记忆中的图式,使他们能把大量项目压缩成少数几个组块;材料与图式不匹配时,优势随之消失。认知负荷理论区分内在负荷与外在负荷;相关负荷在后来的修订中被理解为分配给内在负荷的资源,而不是第三种可相加的负荷。下一章讨论当注意力在不同任务之间来回切换时,代价从哪里来。
数据口径:书中以标注框区分已发表的证据、可运行的 Python 计算(16 段程序,全部核对)、原典与作者的假说。
对照数据表
下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库
| string of 26 letters; working memory holds K | 4 chunks |
| novice (letters) | 26 chunks, overflows; letters recalled about 4 of 26 |
| level-1 acronyms | 9 chunks, overflows; letters recalled about 12 of 26 |
| level-2 schemas | 4 chunks, fits; letters recalled about 26 of 26 |
| expert parse: US-AGENCIES | DEGREES | BROADCASTERS | COUNTRIES | |
| raw information 122 bits; the expert stores 4 pointers into a library of 13 schemas, about 15 bits | |
| shuffled string: expert parse has 24 chunks, letters recalled about 6 | |
| the advantage vanishes when the material does not match the schemas (cf. random boards) |
查看完整验算程序
# Chunking as compression: K = 4 slots, one chunk per slot whatever its size. Greedy parse with a
# schema library (longest chunk first); level-2 schemas bind level-1 chunks. If the parse fits in K
# slots all is recalled, else only K chunks survive. Letters and schemas are illustrative.
import math, random
K = 4
STRING = "FBICIANSAPHDMBABBCCNNUSAUK"
LEVEL1 = {"FBI", "CIA", "NSA", "PHD", "MBA", "BBC", "CNN", "USA", "UK"}
LEVEL2 = {("FBI", "CIA", "NSA"): "US-AGENCIES", ("PHD", "MBA"): "DEGREES",
("BBC", "CNN"): "BROADCASTERS", ("USA", "UK"): "COUNTRIES"}
def parse(s, lib):
out, i = [], 0
while i < len(s):
for L in range(min(5, len(s) - i), 0, -1):
if L == 1 or s[i:i + L] in lib:
out.append(s[i:i + L]); i += L; break
return out
def bind(chunks, lib2):
out, i = [], 0
while i < len(chunks):
for key in sorted(lib2, key=len, reverse=True):
if tuple(chunks[i:i + len(key)]) == key:
out.append(lib2[key]); i += len(key); break
else:
out.append(chunks[i]); i += 1
return out
def recalled_letters(chunks, level1_of):
kept = sorted(chunks, key=lambda c: -len(level1_of(c)))[:K]
return min(len(STRING), sum(len(level1_of(c)) for c in kept))
inv = {v: "".join(k) for k, v in LEVEL2.items()}
size = lambda c: inv.get(c, c)
novice = list(STRING)
inter = parse(STRING, LEVEL1)
expert = bind(inter, LEVEL2)
print(f"string of {len(STRING)} letters; working memory holds K = {K} chunks")
for name, ch in (("novice (letters)", novice), ("level-1 acronyms", inter),
("level-2 schemas", expert)):
fits = "fits" if len(ch) <= K else "overflows"
print(f" {name:17s}: {len(ch):2d} chunks, {fits}; letters recalled about "
f"{recalled_letters(ch, size):2d} of {len(STRING)}")
print(" expert parse:", " | ".join(expert))
bits = len(STRING) * math.log2(26)
lib = len(LEVEL1) + len(LEVEL2)
print(f"raw information {bits:.0f} bits; the expert stores 4 pointers into a library of "
f"{lib} schemas, about {4 * math.log2(lib):.0f} bits")
rng = random.Random(4)
rnd = "".join(rng.sample(STRING, len(STRING))) # same letters, shuffled
r = bind(parse(rnd, LEVEL1), LEVEL2)
print(f"shuffled string: expert parse has {len(r)} chunks, letters recalled about "
f"{recalled_letters(r, size)}")
print(" the advantage vanishes when the material does not match the schemas (cf. random boards)")唯一数据源与引用
范雅福(Fan, Y.). (2026). 《注意力网络、工作记忆瓶颈与内在动机的自决定理论》第 2 章「工作记忆的瓶颈:组块、图式与认知负荷」(版本 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23761341
DOI: 10.5281/zenodo.23761341 · 阅读本章所在专著 · 5DL-B-024 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026024,
author = {Fan, Yafu},
title = {Attention Networks, the Working-Memory Bottleneck and Self-Determination Theory},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
doi = {10.5281/zenodo.23761341},
url = {https://doi.org/10.5281/zenodo.23761341}
}