5D Library › Attention Networks, the Working-Memory Bottleneck and Self-Determination Theory · 中文
Chapter 1 Attention Is Not a Single Tap: The Three-Network Model and the Attention Network Test
Everyday language speaks of attention as something that can be “focused” or “scattered”, as though it were a tap with an on–off switch and a flow rate. Cognitive neuroscience takes a different view: attention consists of at least several systems that can be told apart functionally and are partly separate anatomically; each does a different job, and each can fail in its own way.
Posner and Petersen’s three networks
Evidence
The original framework. In the Annual Review of Neuroscience in 1990, Posner and Petersen proposed that the attention system is anatomically separate from the processing systems it acts upon, and consists of several networks carrying out different functions: orienting to sensory events, detecting signals (later called executive attention), and maintaining an alert state (Posner and Petersen 1990). Revision twenty years on. In 2012 Petersen and Posner reviewed how the framework had developed: the alerting network is associated with the noradrenergic system of the locus coeruleus and with right frontal and parietal regions; the orienting network involves dorsal and ventral frontoparietal regions and is modulated mainly by the cholinergic system; the executive control network includes the anterior cingulate, anterior insula and lateral prefrontal cortex and is closely tied to the dopamine system. They also noted that later work split executive control into frontoparietal and cingulo-opercular subnetworks (Petersen and Posner 2012).
What “associated with” means needs to be stated plainly. The mapping between neurotransmitters and networks refers to principal modulators: pharmacological and imaging studies find that manipulating a given transmitter system affects the measures of a given network most clearly. It does not mean that alerting “is” noradrenaline, nor that other transmitters play no part in that network. Each network comprises several brain regions, and those regions are modulated by several transmitters at once.
Note
“Dopamine = motivation” is a popular exaggeration. The dopamine system is involved in executive control, reward learning and willingness to exert effort, but it is neither a “pleasure chemical” nor motivation itself. Reviewing animal and human work, Berridge and Robinson argued that dopamine is closer to lending cues “incentive salience” (making them “wanted”) than to producing the hedonic experience of “liking” (Berridge and Robinson 1998). Reducing complex questions of motivation to “not enough dopamine” or “dopamine hijacked by phones” is inaccurate and can encourage inappropriate self-medication. This book does not discuss any drug.
The Attention Network Test and subtraction scores
To measure all three networks within one task, Fan, McCandliss, Sommer, Raz and Posner designed the Attention Network Test (ANT) in 2002. Participants see a row of arrows above or below the centre of the screen and report the direction of the middle arrow. Before the arrows appear there may be no cue, a cue asterisk at the centre, cues at both possible locations at once (a double cue), or a cue at the location where the arrows are about to appear (a spatial cue); the flanking arrows either side of the target point the same way (congruent) or the opposite way (incongruent).
The efficiency of each network is expressed as a difference in reaction times:
The alerting score measures how much a timing signal (“it’s coming”) helps; the orienting score how much knowing “where it will appear” helps; the conflict score how much extra time is needed to overcome incongruent flankers, so a larger value means executive control is working harder on this task.
Evidence
The original study. In 40 adults, Fan and colleagues found that this half-hour test yielded individual estimates for each network, and that the three scores correlated little with one another, from which they concluded that the networks are relatively independent; at the same time, the alerting and orienting conditions changed the size of the flanker interference, so the networks also interact (Fan et al. 2002).
Calculation
The program below first writes a generative model of one person’s reaction times (parameters illustrative) and then scores it with the subtraction formulas above. It compares four “fictional people”: a reference person, someone 60 ms slower across the board, someone with lower tonic (sustained) alertness, and someone with weaker conflict control.
# ANT subtraction scores on simulated mean reaction times (ms); all parameters illustrative.
# alerting = RT(no cue) - RT(double cue); orienting = RT(centre) - RT(spatial);
# conflict = RT(incongruent) - RT(congruent), averaged over cues.
# Model: a missing warning cue costs tonic_gap; a spatial cue saves orient; incongruent flankers
# cost conflict, slightly enlarged (k_alert) after any warning cue.
import random
random.seed(1)
CUES = ("none", "centre", "double", "spatial")
def mean_rt(cue, flank, p):
phasic = 0.0 if cue == "none" else 1.0
rt = p["base"] + p["tonic_gap"] * (1 - phasic) # no warning: pay the alertness gap
rt -= p["orient"] if cue == "spatial" else 0.0
if flank == "incongruent":
rt += p["conflict"] * (1 + p["k_alert"] * phasic) # alerting slightly enlarges conflict
return rt
def run_ant(p, trials=48, noise=90.0):
cells = {}
for cue in CUES:
for flank in ("congruent", "incongruent"):
m = mean_rt(cue, flank, p)
cells[cue, flank] = sum(random.gauss(m, noise) for _ in range(trials)) / trials
cue_m = {c: (cells[c, "congruent"] + cells[c, "incongruent"]) / 2 for c in CUES}
inc = sum(cells[c, "incongruent"] for c in CUES) / 4
con = sum(cells[c, "congruent"] for c in CUES) / 4
return (cue_m["none"] - cue_m["double"], cue_m["centre"] - cue_m["spatial"], inc - con,
sum(cue_m.values()) / 4)
BASE = dict(base=520.0, tonic_gap=45.0, orient=45.0, conflict=90.0, k_alert=0.15)
PROFILES = [
("reference person", {}),
("everything 60 ms slower", dict(base=580.0)),
("lower tonic alertness", dict(base=540.0, tonic_gap=80.0)),
("weaker conflict control", dict(conflict=125.0)),
]
print("profile | mean RT | alerting | orienting | conflict (ms; near-noiseless)")
for name, change in PROFILES:
p = dict(BASE, **change)
a, o, c, m = run_ant(p, trials=4000) # many trials: close to the true values
print(f"{name:24s} | {m:7.0f} | {a:8.0f} | {o:9.0f} | {c:8.0f}")
a, o, c, m = run_ant(BASE)
print(f"one realistic session for the reference person: alerting {a:.0f}, orienting {o:.0f},"
f" conflict {c:.0f}")
print("note: uniform slowing leaves all three scores unchanged; LOWER tonic alertness shows up")
print("as a LARGER alerting score, because the warning cue then helps more")
# profile | mean RT | alerting | orienting | conflict (ms; near-noiseless)
# reference person | 570 | 38 | 43 | 99
# everything 60 ms slower | 630 | 40 | 45 | 100
# lower tonic alertness | 599 | 74 | 43 | 100
# weaker conflict control | 590 | 38 | 46 | 139
# one realistic session for the reference person: alerting 17, orienting 33, conflict 104
# note: uniform slowing leaves all three scores unchanged; LOWER tonic alertness shows up
# as a LARGER alerting score, because the warning cue then helps moreThe example shows two features of subtraction scores. First, uniform slowing leaves all three scores essentially unchanged (the 1–2 ms differences in the table are simulation noise): subtraction removes whatever all conditions share, so a person who is slow overall can have exactly the same network scores as anyone else. Second, lower sustained alertness makes the alerting score larger: if a person struggles to stay ready when there is no cue, the cue helps more, and “no cue minus double cue” rises. A high alerting score is therefore not necessarily “good alertness”; ANT scores can be read only as differences between conditions. The last line of output also shows that one ordinary session of 48 trials per cell produced an alerting score (17 ms) far from the person’s true value (about 38 ms): a single measurement is very noisy.
Are the three scores reliable?
Evidence
A large psychometric appraisal. MacLeod and colleagues pooled ANT data from several studies, 1,129 healthy adults in all, and found Spearman–Brown-corrected split-half reliabilities of 0.38 for alerting and 0.55 for orienting, both low, and 0.81 for the executive (conflict) score, moderate to high. Both analyses of variance and correlational analyses indicated that the three networks measured by the ANT are not independent (MacLeod et al. 2010). They recommended careful attention to these psychometric properties whenever ANT data are collected, analysed or interpreted.
Why are difference scores often unreliable? The program below gives a common explanation.
Calculation
Each fictional person has a true network score; the between-person standard deviation (“true SD”) is set at 12 ms for alerting, 16 ms for orienting and 25 ms for conflict, and trial-to-trial reaction-time noise at 110 ms. Reliability = true between-person variance ÷ observed variance.
# Split-half reliability of a difference score: true between-person SD vs trial noise.
# Reliability = true variance / observed variance. SDs and noise are illustrative.
import random, math
random.seed(7)
NOISE = 110.0 # trial-to-trial RT noise (ms)
TRUE_SD = {"alerting": 12.0, "orienting": 16.0, "conflict": 25.0}
TRIALS = {"alerting": 48, "orienting": 48, "conflict": 96} # trials per condition in one half
def split_half(sd, n_trials, people=400):
halves = []
for _ in range(people):
t = random.gauss(0, sd)
h = []
for _half in range(2):
se = NOISE * math.sqrt(2 / n_trials) # SE of a difference of two means
h.append(t + random.gauss(0, se))
halves.append(h)
x = [h[0] for h in halves]; y = [h[1] for h in halves]
mx, my = sum(x) / len(x), sum(y) / len(y)
sxy = sum((a - mx) * (b - my) for a, b in zip(x, y))
r = sxy / math.sqrt(sum((a - mx) ** 2 for a in x) * sum((b - my) ** 2 for b in y))
return 2 * r / (1 + r) # Spearman-Brown to full length
def expected(sd, n_trials):
err = NOISE ** 2 * 2 / (2 * n_trials) # error variance of the full-length score
return sd ** 2 / (sd ** 2 + err)
print("network | true SD | trials per cond. | simulated split-half | expected")
for k in TRUE_SD:
sim = split_half(TRUE_SD[k], TRIALS[k])
print(f"{k:10s} | {TRUE_SD[k]:7.0f} | {TRIALS[k]:16d} | {sim:20.2f}"
f" | {expected(TRUE_SD[k], TRIALS[k]):8.2f}")
print("alerting reliability as the session gets longer (trials per condition in each half):")
sd_a = TRUE_SD["alerting"]
print(" " + " ".join(f"{n}: {expected(sd_a, n):.2f}" for n in (24, 48, 96, 192, 384)))
# network | true SD | trials per cond. | simulated split-half | expected
# alerting | 12 | 48 | 0.36 | 0.36
# orienting | 16 | 48 | 0.56 | 0.50
# conflict | 25 | 96 | 0.84 | 0.83
# alerting reliability as the session gets longer (trials per condition in each half):
# 24: 0.22 48: 0.36 96: 0.53 192: 0.70 384: 0.82
With these illustrative parameters the program’s reliabilities (expected: alerting 0.36, orienting 0.50, conflict 0.83; simulated: 0.36, 0.56, 0.84) are close to MacLeod and colleagues’ values, which shows only that the explanation is plausible: a difference inherits the measurement error of two conditions, while people may genuinely differ little on that difference. To bring the alerting score up to a reliability of 0.7, these parameters require about 192 trials per condition in each half, four times the session length assumed here.
Sleep, notifications and stress: writing only what the evidence supports
The author’s framework wanted to show which subnetwork is affected by sleep deprivation, by notification interruptions and by high stress. The evidence supports only fairly coarse statements.
Evidence
Sleep deprivation. Lim and Dinges’s meta-analysis of 70 articles and 147 cognitive tests found that short-term total sleep deprivation affects cognitive domains to different degrees: simple attention and vigilance (sustained attention) are affected most (effect size for lapses about −0.76), working memory next, complex attention (including tasks of orienting and executive attention) less, and reasoning and crystallised intelligence hardly at all (Lim and Dinges 2010). One study that used the ANT directly kept 18 men awake for 24 hours and found overall slowing and a marked drop in vigilance at night, with orienting and executive control also affected, and an interaction between sustained alertness and the other two networks (Martella et al. 2011). The two kinds of evidence emphasise different things and differ greatly in sample size. Notifications. Stothart and colleagues had students perform the Sustained Attention to Response Task (SART; 166 analysed) and, midway through, called or texted the phones of some of them while others received no notification; those who received a notification, even without answering or looking, were clearly more likely to make errors (Stothart et al. 2015). The SART measures mainly sustained attention and the inhibition of a habitual response, and does not map cleanly onto any one network. Acute stress. A meta-analysis by Shields, Sazma and Yonelinas found that acute stress generally impairs working memory and cognitive flexibility, while its effect on inhibitory control depends on the type of inhibition (Shields et al. 2016).
The safer statements, then, are these: lack of sleep first weakens sustained alertness; phone notifications interrupt sustained attention and increase errors; acute stress mainly affects executive functions such as working memory and flexible switching. Tidier mappings such as “sleep loss harms only alerting, not executive control” or “stress harms only conflict control” lack sufficient evidence at present, and this book does not make them.
Hypothesis · not established
Score-direction hypothesis (not yet established). If lack of sleep mainly lowers sustained (tonic) alertness, it should make the ANT alerting score larger (the cue helps more), not smaller. Testable prediction: in a within-person sleep-restriction study, reaction times in the no-cue condition should rise more after restriction than those in the double-cue condition. If both rise equally, or the alerting score shrinks, the hypothesis is wrong. Because the alerting score has low reliability, such a study needs more trials or more participants than usual.
Chapter summary
Attention can be divided into at least three networks, alerting, orienting and executive control, modulated principally by the noradrenergic, cholinergic and dopaminergic systems respectively; these are principal modulators, not sole mechanisms. The ANT scores networks by reaction-time differences: uniform slowing does not change the scores, and lower sustained alertness can even make the alerting score larger. Large studies show that the alerting and orienting scores have low reliability and that the three networks are not fully independent. Sleep loss, notifications and acute stress have documented costs, but finer mappings onto “which network” lack sufficient evidence. The next chapter turns to the bottleneck that follows attention: working memory.
References
Fan, Y. (2026). Attention Networks, the Working-Memory Bottleneck and Self-Determination Theory: From Posner's Tripartite Network and Motivation Crowding-Out to the Information Theory of Flow, Read Beside Neo-Confucian Reverence (Jìng) [注意力网络、工作记忆瓶颈与内在动机的自决定理论] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23761341 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).