When a stimulus is presented, its sensory trace decays rapidly, lasting for approximately 1000 ms. This brief and labile memory, referred as iconic memory, serves as a buffer before information is transferred to working memory and executive control. Here we explored the effect of different factors—geometric, spatial, and experience—with respect to the access and the maintenance of information in iconic memory and the progressive distortion of this memory. We studied performance in a partial report paradigm, a design wherein recall of only part of a stimulus array is required. Subjects had to report the identity of a letter in a location that was cued in a variable delay after the stimulus onset. Performance decayed exponentially with time, and we studied the different parameters (time constant, zero-delay value, and decay amplitude) as a function of the different factors. We observed that experience (determined by letter frequency) affected the access to iconic memory but not the temporal decay constant. On the contrary, spatial position affected the temporal course of delay. The entropy of the error distribution increased with time reflecting a progressive morphological distortion of the iconic buffer. We discuss our results on the context of a model of information access to executive control and how it is affected by learning and attention.

*iconic memory*. This sensory buffer, which precedes the formation of short-term or working memory, was extensively studied by Sperling in the sixties, using the

*partial report paradigm*(Sperling, 1960). Sperling showed that when observers saw briefly presented displays composed of several alphanumeric characters, only a few (3 to 5) elements could be remembered. This was consistent with the limits of short-term memory that had been known since, at least, the early experiments of Cattell (1886). However, observers had a much better memory when required to identify a specific subset of the characters at a short interval after the removal of the visual display (partial report). This indicated the existence of a high capacity initial memory of the stimulus display that decayed approximately 1000 ms after stimulus presentation. Since then, numerous studies have studied Iconic Memory (Coltheart, 1980; Loftus, Duncan, & Gehrig, 1992; Lu, Neuse, Madigan, & Dosher, 2005), addressing its characteristics such as their duration (Averbach & Sperling, 1961), content (Chow, 1986; Turvey & Kravetz, 1970), maintenance and extinction (Averbach & Coriell, 1961; Dember & Purcell, 1967), and models of information transfer to working memory (Gegenfurtner & Sperling, 1993; Loftus et al., 1992). The emergent picture from these studies is that iconic memory is extremely short (less than a second), has a great capacity of storage, and is labile (i.e., it can be disrupted by a competing stimulus).

*p*′(right hemi-field): 0.35 ± 0.02;

*p*′(Left Hemi-field): 0.27 ± 0.02; paired

*t*-test:

*t*= 3.7,

*p*< 0.01,

*df*= 18). This spatial population bias in search tasks is well known in the literature and has been related to asymmetries in the spatial allocation of attention, which are related to reading strategies (Efron & Yund, 1996; Goldstein & Babkoff, 2001; Nazir, Ben-Boutayab, Decoppet, Deutsch, & Frost, 2004). Although this bias reflects a trend in the population, the spatial performance map may vary substantially from subject to subject.

*p*′ as a function of letter identity—for each individual subject—collapsing the data across all ISIs and spatial locations. The dispersion in performance across different letters was striking, decreasing almost 3-fold from the most visible letter (letter

*A,*60%) to the less visible letter (letter

*K,*20%). This result was highly consistent across subjects and could not be accounted by a response bias (

*p*′ is the performance corrected for false-positives) since the false-positives (the probability of responding to the letter x when it was not presented) were smaller than 3% for all letters. For instance, for 10 of our 19 subjects, the letter A was the most visible letter. The probability that this results from chance (assuming equal probability of responding to each letter for all subjects) is

*p*< 10

^{−15}. In the next section, we investigate which aspects of letter identity (form, frequency, neighborhood similarity, etc.) account for this observation.

*R*

^{2}= 0.043, paired

*t*-test,

*t*= 1.58,

*p*= 0.13

*df*= 18).

*confusion matrix*

**C**. This matrix is calculated first by measuring the number of responses (

*j*) to the stimulus (

*i*) and then estimating the probability that this results from chance under the null hypothesis that all letters are responded with equal probability. Thus,

**C**(

*i, j*) is a measure of the probability that responding to the letter

*i*given that the target was the letter

*j*does not result by chance. All pairs for which

**C**(

*i, j*) < 0.05 are shown in Figure 2B. From an inspection of the most frequent errors (Table 1), it becomes evident that the confusion matrix is mainly dominated by morphologic resemblance; for instance, letter J was mainly confused with letter I, and similarly, the other confused pairs M–W, E–B, W–M, P–R, and I–J correspond to letters with high morphological resemblance.

Target letter | Answered letter | % of total errors for the target |
---|---|---|

J | I | 16.0 |

M | W | 12.7 |

R | S | 10.7 |

E | B | 10.4 |

W | M | 10.1 |

P | R | 10.1 |

N | M | 10.1 |

I | J | 9.9 |

R | A | 9.9 |

S | C | 9.7 |

*neighborhood of similarity*of a letter as:

*N*(

*x*) =

*y*∣

**C**(

*x, y*) < 0.05, i.e., for each letter

*x,*its

*neighborhood of similarity N*(

*x*) is composed of all letters (

*y*) such that the probability of responding to

*y*given that the target was

*x*results from chance—is smaller than 0.05. The number of elements in

*N*(

*x*) varies with

*x*between 0 (words that do not have highly probable error targets) and 3. We then estimated a more generalized notion of performance, considering “approximate” responses as those for which the responded letter was either the target or belonged to the

*neighborhood of similarity*. We then estimated the regression between the percent of approximate responses and the frequency of each letter ( Figure 2C) and observed that this correlation is significant (

*R*

^{2}= 0.28,

*t*-test across subjects

*p*< 0.01,

*t*= 7.3,

*df*= 18). This shows that natural frequency of the letter has an effect on performance, which becomes significant once morphologic contributions to the variance are taken into account.

*S*

_{ i}corresponds to the entropy of errors for letter

*i, T*(

*i, j*) is the probability of responding the letter

*j*given that the letter

*i*was presented, and

*p*

_{ ij}is the normalized probability across errors, i.e., the probability of responding the letter

*j*given that the letter

*i*was presented and the response corresponded to an error trial. The total entropy

*S*is calculated as the mean entropy averaging across all letters. We then calculated the entropy

*S*as a function of ISIs ( Figure 2D) and observed that, as predicted, the entropy of the error distribution increases with ISI. Qualitatively, the time constants of the progression of the entropy distribution and of the decay in performance are comparable. We will later show that this observation can be quantified, performing a second experiment with repeated experimental sessions in single subjects, which allows to determine the parameters of the exponential decay at the individual level. The minimum and the maximum values of the entropy (which is measured in bits) provide a measure of the range of clustering in the error distribution for varying ISIs: In a fully uniformly distributed error distribution among the 25 remaining letters, the entropy would correspond to ln(25) ≈ 3.22; on the contrary, in the case in which all errors would be clustered in only four neighbor letters, the entropy would be ln(4) ≈ 1.39.

*p*′ with frequency for short and long ISIs. Performance increased with frequency in a non-significantly different manner both for short and long ISIs ( Figure 3A). To quantify this observation, we performed an ANOVA with ISI (short or long) and frequency (the six categories) as main factors and subjects as a random variable. The main effects of frequency and of ISI were significant (ISI:

*F*= 58.67,

*df*= 1,

*p*< 0.01; frequency:

*F*= 3.99,

*df*= 5,

*p*< 0.01) but the interaction was not significant (

*F*= 1.19,

*df*= 5,

*p*> 0.1). This indicates that the effect of frequency in performance is comparable at short and long ISIs, i.e., that persistency in iconic memory is not determined by letter frequency. In the next experiment, we will provide further evidence for this, showing that the temporal constant of the decay is unaffected by the frequency manipulation.

*p*′ with spatial location (sorted according to performance when collapsing across all ISI values). The effect of position was considerably more pronounced for short ISI values ( Figure 3B). To quantify this observation, we performed an ANOVA with ISI (short or long) and position as main factors and subjects as a random variable. The main effects of position and of ISI were significant (ISI:

*F*= 70.68,

*df*= 1,

*p*< 0.01; position:

*F*= 141.88,

*df*= 7,

*p*< 0.01). The interaction between ISI and position was significant (

*F*= 3,

*df*= 7,

*p*< 0.01) in contrast with what we had observed for the interaction between frequency and ISI, which was not significant. This indicates that there are significant differences in the spatial distribution of performance during the few-hundred milliseconds between stimulus presentation and response.

*p*′ from the shortest to the longest SOA was 0.14, which, with a stimulus display of eight letters, correspond to 1.13 letters. Moreover, of the 24 subjects that participated in the first study, only 13 showed a reliable fit to an exponential function (set to the criterion:

*R*

^{2}> 0.6 and 15 ms <

*τ*< 800 ms). We conducted a second experiment in which we studied performance in multiple sessions of two highly practiced subjects (authors MS and MG) with two aims:

- to collect enough trials in an individual subject basis to test an exponential model of decay function and understand the effect of the experimental manipulations on the different parameters of the exponential and
- to assure that we were measuring the limits of iconic memory rather than the quality of iconic memory in inexperienced subjects.

*p*′ as a function of the delay and, for each individual subject and session, fitted this distribution to an exponential with three free parameters

*α*(the gain) indicates the change in performance between short and long delays, i.e., a measure of the information that accesses iconic memory and does not access working memory or explicit reports after a long-delay,

*β*(performance at 8) indicates the performance level for large delays, i.e., the probability that a stimulus accesses working memory, and

*τ*(the time constant) indicates the temporal constant which characterizes the exponential decay in performance, i.e., the duration of iconic memory.

Condition | α | β | τ (ms) |
---|---|---|---|

Subject MG | |||

General | 0.341 ± 0.005 | 0.450 ± 0.003 | 217 ± 8 |

Low-frequency letters | 0.34 ± 0.01 | 0.41 ± 0.01 ^{a} | 189 ± 18 |

High-frequency letters | 0.34 ± 0.01 | 0.48 ± 0.01 ^{a} | 271 ± 34 |

RVF | 0.28 ± 0.03 ^{a} | 0.55 ± 0.01 ^{a} | 89 ± 30 ^{a} |

LVF | 0.45 ± 0.02 ^{a} | 0.34 ± 0.01 ^{a} | 276 ± 28 ^{a} |

| |||

Subject MS | |||

General | 0.28 ± 0.01 | 0.22 ± 0.01 | 201 ± 26 |

Low-frequency letters | 0.28 ± 0.02 | 0.16 ± 0.02 ^{a} | 354 ± 71 |

High-frequency letters | 0.30 ± 0.02 | 0.25 ± 0.01 ^{a} | 150 ± 20 |

RVF | 0.40 ± 0.02 ^{a} | 0.24 ± 0.01 | 134 ± 16 |

LVF | 0.17 ± 0.02 ^{a} | 0.20 ± 0.02 | 269 ± 105 |

*S*=

*S*

_{max}−

*α*·

*e*

^{(− t/ τ)}(see legend of Figure 4 for the parameters of the regression for each subject). In both subjects, entropy increased monotonically with ISI, with a temporal constant not significantly different to that obtained for the iconic memory decay (paired

*t*-test across sessions:

*t*= 0.21,

*p*> 0.1,

*df*= 11).

*F*= 85.68

*df*= 1

*p*< 0.05; Frequency,

*F*= 3.25

*df*= 5

*p*< 0.05; ISI × Frequency,

*F*= 1.29

*df*= 5

*p*> 0.1; Sub-MG: ISI,

*F*= 125.64

*df*= 1

*p*< 0.05; Frequency,

*F*= 2.47

*df*= 5

*p*< 0.05; ISI × Frequency,

*F*= 1.98

*df*= 5

*p*> 0.1). When we fitted this data to an exponential, we observed that the only parameter affected in this regression by the conditions was the additive constant of the exponential

*β*(see Table 2). This further suggests that while there is an increase in performance for high frequency letters, this effect does not change in time, i.e., that persistency in iconic memory is not determined by letter frequency.

*t*-test: Sub-MS:

*t*= 5.5,

*p*< 0.01,

*df*= 5; Sub-MG:

*t*= 6.5,

*p*< 0.01,

*df*= 5; see also inset Figure 4A) and showed an interaction of Position with ISI (ANOVA, Sub-MS: ISI,

*F*= 217.43

*df*= 1

*p*< 0.01; Position,

*F*= 62.11

*df*= 7

*p*< 0.01; ISI × Position,

*F*= 6.43

*df*= 7

*p*< 0.01; Sub-MG: ISI,

*F*= 157.29

*df*= 1

*p*< 0.01; Position,

*F*= 100.95

*df*= 7

*p*< 0.01; ISI × Position,

*F*= 7.18

*df*= 7

*p*< 0.01), although in one of the two subjects, this interaction is reversed (i.e., spatial changes in performance are shorter for the shorter ISIs, see figure in 1). The analysis of the parameters of the exponential also yielded a more complicated and variable picture of the position manipulation, summarized in Table 2. Thus, while the effect of frequency is to a large extent insensitive to time, the effect of position shows a strong interaction, although the specific pattern of this interaction may vary across different subjects. Such variability may be related to different strategies in the spatio-temporal allocation of attention, and more specific experiments are required to determine a precise model of the evolution of iconic memory in different locations of the visual field. Here we merely provide in 1 a tentative explanation of the difference across subjects.

*τ*

*t*-test student comparisons and ANOVA, assuming a normal distribution for the data. For linear regression analysis, we discard the data points bigger than two standard deviations (but they were shown in the figures in red).

*p*′ for positions of high, medium and low performance. In both subjects, the maximal decrease, indepedently of the spatial location, was observed at locations of intermediate performances.

*p*′ ∼ 1) and thus close to saturation and with very low variations of performance with ISI. In other locations, performance was very low (

*p*′ ∼ 0.2). In these locations, we also observed a very modest improvement for short ISIs. On the contrary, in locations in which average performance was intermediate (

*p*′ ∼ 0.5), the fraction of correct responses showed the most significant change with ISI. This result can explain the reversal of the interaction between the two subjects. For most subjects (of the naïve group and subject MS), the maximum of performance across spatial locations (right visual field) corresponded to intermediate levels of performance that are strongly affected by ISI, and the minimum of performance corresponded to very poor levels of performance thus showing a more modest effect of ISI. Subject MG, who had large amounts of practice in this task, showed saturation for the most performing locations and intermediate levels of performance for the worse locations, hence the inverted effect. It might be interesting to consider, in further studies, whether achieving close to perfect levels of performance in partial report paradigms might be achieved with sufficient extensive training.