Memory and incidental learning for visual frozen noise sequences

Abstract

Five experiments explored short-term memory and incidental learning for random visual spatio-temporal sequences. In each experiment, human observers saw samples of 8 Hz temporally-modulated 1D or 2D contrast noise sequences whose members were either uncorrelated across an entire 1-s long stimulus sequence, or comprised two frozen noise sequences that repeated identically between a stimulus' first and second 500 ms halves ("Repeated" noise). Presented with randomly intermixed stimuli of both types, observers judged whether each sequence repeated or not. Additionally, a particular exemplar of Repeated noise (a frozen or "Fixed Repeated" noise) was interspersed multiple times within a block of trials. As previously shown with auditory frozen noise stimuli (Agus, Thorpe, & Pressnitzer, 2010) recognition performance (d') increased with successive presentations of a Fixed Repeated stimulus, and exceeded performance with regular Repeated noise. However, unlike the case with auditory stimuli, learning of random visual stimuli was slow and gradual, rather than fast and abrupt. Reverse correlation revealed that contrasts occupying particular temporal positions within a sequence had disproportionately heavy weight in observers' judgments. A subsequent experiment suggested that this result arose from observers' uncertainty about the temporal mid-point of the noise sequences. Additionally, discrimination performance fell dramatically when a sequence of contrast values was repeated, but in reverse ("mirror image") order. This poor performance with temporal mirror images is strikingly different from vision's exquisite sensitivity to spatial mirror images.

Keywords: Frozen noise; Incidental learning; Memory; Mirror image; Reverse correlation.

Figure 1

Examples of Experiment 1's three kinds of stimuli (Noise(N), Repeated Noise (RN), and Fixed Repeated Noise (FixRN). Note that the second half of an RN stimulus recycles the stimulus’ first half; note also that an entire sequence FixRN repeats identically from one trial to a later trial. Top panel: 1D noise. Bottom panel: 2D noise.

Figure 2

Mean values of d’ for Experiment 1 (panels A and B), Experiment 2 (panel C) and Experiment 3 (panel D). Panels A, C and D plot d’ for 1D noise and Panel B plots d‘ for 2D noise The hit rates for RN stimuli and fixRN stimuli were both referenced to false alarm rates for N stimuli. Error bars represent +/− 1 s.e.m.

Figure 3

Trial-wise performance in Experiment 1. In both figures, percentage of hits is plotted as a function of trial for both RN and FixRN stimuli. Left panel: 1D noise. Right panel: 2D noise. Lines roughly following the data points in each panel were produced by smoothing the raw data (circles) with a moving rectangular window three trials wide. Error bars represent +/− 1 s.d. for the smoothed data, estimated by bootstrap simulations. Straight lines correspond to the best fitting (least squares) linear fit to each set of raw data.

Figure 4

Mean (closed symbols) and variance (open symbols) kernels estimated by reverse correlation for 1D noise stimuli in Experiments 1 and 2. Panel A: kernels for Experiment 1. Panel B: kernels for Experiment 2. Error bars on symbols represent +/− 1 s.d., estimated by bootstrap simulations. Gray bands denote +/− 2 s.d. confidence regions for each kernel type, estimated by bootstrap simulations.

Figure 5

A) d’ for the FixRN stimuli for individual observers in Experiment 1, plotted as a function of d’ for RN stimuli; B) Total summed contrast (signed) for the FixRN stimuli shown to individual observers in Experiment 1, plotted as a function of d’; C) The FixRN stimuli shown to observers in Experiment 1, sorted in order of d’. Each pixel corresponds to a single frame, with time progressing from top to bottom.

Figure 6

A) Trial-wise performance in Experiment 2. Percentage of hits is plotted as a function of trial for both RN and FixRN stimuli. Lines roughly following the data points were produced by smoothing the raw data (circles) with a three-trial-wide roving window. Error bars represent +/− 1 s.d. for the smoothed data, estimated by bootstrap simulations. Straight lines correspond to the best fitting (least squares) linear fit to each set of raw data. B) Individual observer d's for each of the four FixRN exemplars used in Experiment 2. Symbol numbers correspond to individual observers. Filled symbols correspond to the mean across observers for each FixRN sample. Error bars represent +/− 1 s.e.m.

Figure 7

Trial-wise performance in Experiment 3, shown as percentage of hits plotted against successive trials for both RN and FixRN stimuli. Results in the lefthand panel are for trials on which a sequence was repeated in Forward direction (as in Experiments 1 and 2); results in the righthand panel are for trials on which a sequence was repeated in Reverse order. Lines roughly following the data points were produced by smoothing the raw data (circles) with a three-trial-wide roving window. Straight lines correspond to the best-fitting (least-squares) linear fits to the raw data. Error bars represent +/− 1 s.d. for the smoothed data, estimated by bootstrap simulations.

Figure 8

Mean (closed symbols) and variance (open symbols) kernels estimated by reverse correlation for the two types of 1D noise stimuli used in Experiment 3. Left panel: kernel for stimulus sequences whose first four items repeated in the same direction (forward) as originally presented. Right panel: kernels for stimulus sequences whose first four items were repeated in reverse order. Error bars on symbols represent +/− 1 s.d., estimated by bootstrap simulations. Gray bands denote +/− 2 s.d. confidence regions for each kernel type, estimated by bootstrap simulations.

Figure 9

Values of d’ produced by 1D and 2D versions of RN and FixRN stimuli on the first and second days of testing.

Figure 10

Trial-wise performance in Experiment 4, shown as percentage of hits plotted as a function of successive trials for RN (filled symbols) and FixRN stimuli (open symbols). Panels in the upper row show results for 1D stimuli; panels in the lower row show results for 2D stimuli. The lefthand panels show results from Day 1 of testing; the righthand panels show results from Day 2 of testing. Lines roughly following the data points were produced by smoothing the raw data with a three-trial-wide roving window. Error bars represent +/− 1 s.d. for the smoothed data, estimated by bootstrap simulations. Straight lines correspond to the best fitting (least squares) linear fit to each set of raw data.

Figure 11

Mean kernels estimated by reverse correlation for the 1D and 2D noise stimuli used in Experiment 4. Values derived from an analysis of the first day's testing are shown as filled symbols; values from the second day's testing are shown as open symbols. Left panel: mean kernel for 1D stimulus sequences. Middle panel: mean kernel for contrasts comprising the left half of a 2D stimulus sequence; right panel: mean kernel for contrasts comprising the right half of a 2D stimulus sequence. Error bars represent +/− 1 s.d., estimated by bootstrap simulations. Gray band in each panel denotes +/− 2 s.d. confidence region, estimated by bootstrap simulations.

Figure 12

d’ values associated with gaps of different duration inserted into sequences between the fourth and fifth items. Open symbols show results for RN stimuli; closed symbols are for FixRN stimuli.

Figure 13

Mean kernels estimated by reverse correlation for sequences into which various durations of gap have been inserted between items four and five of the sequence. Upper left panel: kernels estimated by averaging over all orders of gaps. Upper right panel: kernels estimated only from observers tested first with 0 msec gap (i.e., no gap). Bottom left panel: kernels estimated from observers tested first with a 133 msec gap. Bottom right panel: kernels estimated from observers tested first with a 400 msec gap. Error bars represent +/−1 s.d., estimated by bootstrap simulations. Gray band in each panel denotes +/− 2 s.d. confidence region, estimated by bootstrap simulations.

Figure 14

A) The relationship between the two halves of a non-repeat (N) stimulus sequence predicts whether the stimulus will attract a false recognition. The probability of a false recognition response to a non-repeat stimulus [P(fa|N)] is plotted against levels of difference between the summed contrast of a stimulus’ two halves. The family of curves represents results from various experiments. Error bars were generated by means of bootstrapping. The dotted line shows results averaged over all the experiments, with size of error bars shown by the gray ribbon. The leftmost data points in each curve represent the one-fourth of trials on which the difference between summed contrasts between sequence halves was smallest; the rightmost data points are for the one-fourth of trials that had the largest difference between first and second halves’ summed contrasts. B) Results averaged over all experiments, with each quartile's value expressed as a proportion of all false alarms. Error bars for values shown in the inset were calculated from the fractional uncertainty for each original estimate. C) P(fa|N) as a function of the temporal correlation between the first and second halves of the noise sequence (trials were pooled from all experiments). Line through the data is the best (least-squares) linear fit.