Figure 1.
Sequential delayed match-to-sample task.
(A) Event sequence. First, a gray fixation spot appears in the center of the screen. Once the monkey fixates on the spot, a sample image replaces the fixation spot for 0.5–1.0 s, after which the spot is restored. After a variable delay-period, the image/spot sequence is repeated 0, 1, 2, or 3 times with nonmatching patterns. Finally, the original (matching) pattern reappears; the monkey has to release the bar within 2 s to get a drop of water as a reward. (ITI, inter-trial interval). (B) Stimuli.
Figure 2.
Response variance explained by stimulus identity, demonstrating that these cells were stimulus selective.
(A) Percentage of variance explained by stimulus identity in the sample, nonmatch, and match task phases for the population of 35 TE neurons, and for the population of 11 perirhinal cortex neurons. In box plots, the middle line indicates the median. The notches indicate the 95% confidence interval for the median. The whiskers extend to the most extreme data point which is no more than 1.5 times the inter-quartile range from the box. Population distributions across both TE and perirhinal cortex explained the same amount of variance in the sample, nonmatch and match phases. However, in some cells, the behavioral phase significantly influenced the response magnitude. (B) percentage of variance explained by sample stimulus identity in a 200-ms delay period before nonmatch stimulus presentation (Delay(NM)), and before match stimulus presentation (Delay(M)).
Figure 3.
Correlations of response deviations.
(A, B) Correlations between sample versus match (filled circles, solid line) and sample versus nonmatch response deviations (open circles, broken line) for one TE (A) and one perirhinal neuron (B). The correlation for sample vs. match is significantly different than for sample versus nonmatch deviations in (A) (Z statistic = 2.11, df = 366, p = 0.018). (C, D) Mean±SE of variance of response deviations explained by TE (C) and perirhinal (D) populations. Differences between sample vs. match and other comparisons (either gray versus either white bar) were significant only in TE (paired t-test, p<0.05).
Figure 4.
The matched filter model calculates the correlation between a stored pattern and an incoming pattern. A local memory trace is stored in synaptic weights (Wkm) of TE neurons (blue) when a behaviorally determined learning command (red) is sent to the neurons. The image is encoded by a rich array of neurons in the visual encoder (gray circles). The Ckm represents the output from a sparse subpopulation of the encoder cells. The green and brown circles represent two overlapping subpopulations that project to different TE neurons. The yellow blocks represent noise between the encoder and the TE neurons; the αkm are independent, multiplicative noise; the βkm are independent additive noise; the gold block (δ) represents noise common to all neurons constant throughout one trial (e.g., due to arousal). Thus, on each trial for each stimulus presentation, a new value of α and β are drawn for each synapse, but only a single value of δ is drawn. The xkm represents the input to the synapses, after the noise has been added, and Rk represents the output of the k-th TE neuron.
Figure 5.
Predictions of responses by the deterministic model simulation.
(A, B) Predictions of responses for TE neurons with inputs from encoder stage estimates (Equations 1–10) (A), and by the model for perirhinal neurons (Equation 3 was applied) (B). Left column shows predictions of sample responses compared to the actual sample responses. Right column shows predictions for nonmatch responses. Each colored dot represents data points for each neuron, and each pattern with a colored outline indicates the mean response versus mean predicted response to the pattern for the neuron. Variance explained is high in all cases, because the cells in both areas are stimulus selective.
Figure 6.
Predictions of response deviations by the deterministic model simulation.
(A, B) Predictions of the deviations for TE (A) and perirhinal (B) neurons. Left column shows predictions of the deviations from the mean in the sample responses compared to the actual sample response deviations. Right column shows predictions for nonmatch response deviations. Variance explained is low in TE, but zero for perirhinal, which is a rough match to our data. (Format as in Figure 5.)
Figure 7.
Performance on the DMS task of the deterministic matched filter model.
(A) Performance of the deterministic matched filter model using the data from the recorded sample of 35 TE neurons. The left column shows the eight stimuli presented to the model as the sample, and the top row shows the eight stimuli presented as the test image. At the intersection of each row and column is the average response of all the estimates across 35 TE neurons using the matched filter model. The upper-left to lower-right diagonal shows the matched filter outputs for the eight sample-match pairs. The off diagonals show the matched filter outputs for the 56 sample-nonmatch pairs. The model gave the best discrimination performance with the threshold set to 6.15 spikes per 400 ms epoch, i.e., the model made the fewest mistakes. The blue values show correct matches (hits), and the green responses show the correct nonmatches (correct rejections). The orange values show misses, and the red values show false alarms. This model got 32/64 = 50% of the trials correct. (B) Performance of the matched filter model for perirhinal neurons with inputs from encoder stage estimate. At the intersection of each row and column is the average response of all the estimates across 11 perirhinal neurons using the matched filter model. With the threshold set to 4.55 spikes per 400 ms epoch, the model achieved its best performance, getting 35/64 = 55% of the trials correct.
Figure 8.
Example of matched filter output computation.
The top row shows input images. The memory trace of the model is simulated with the discrete Fourier Transform of the input, plus noise (middle row). (Noise is not very noticeable, because of the logarithmic scaling). Bottom row shows the product of the memory trace and the match and nonmatch inputs. The output power is shown on the left.
Figure 9.
Performance of the matched filter across full set of stimulus pairs by the stochastic model simulation.
The left column shows the 8 stimuli presented to the model as the sample, and the top row shows the 8 stimuli presented as a match or nonmatch. The intersection of each row and column is a 16×16 pixel image made up of the responses of the 256 model TE neurons. The diagonal (with slope -1) shows the matched filter outputs for the eight sample-match pairs. The off diagonals show the matched filter outputs for the 56 sample-nonmatch pairs. The total power (normalized to 1.0 for the peak of the 64 pair set, in this example, S7-S7 sample-match) is shown above each output. With the threshold set to 0.225, the model made the fewest mistakes (false alarms, red values). The green values show correct matches, and the blue responses show the correct nonmatches. With the noise in the model adjusted to match that in the monkeys, the model got 62/64 = 97% of the trials correct. The average performance across the two monkeys was 98%.
Figure 10.
Stochastic model's noise correlations.
Correlations for noise in sample versus match and sample versus nonmatch deviations are similar to our data. (A, B) TE shows significantly higher noise correlations for sample versus match phases. (C, D) response deviations in the perirhinal cortex are much less, and more uniformly, correlated. Data are from trials with no intervening nonmatch stimuli for sample versus match or from trials with one nonmatch stimulus for sample versus nonmatch. Uniform correlations in perirhinal cortex are due to slowly varying input noise. Increased sample versus match noise correlations depend on a multiplicative interaction between memory trace and current input.
Figure 11.
Noise-driven output of a matched filter.
Top row shows a sample image made up of white noise uniformly distributed on [0, 1], the template, and the product of the two. The second row shows the effect of averaging across many such representations. Clearly, the signal-to-noise ratio is rapidly improving as the size of the pool being averaged increases, but even the output from a single sample (top row) looks somewhat like the filter. The response to the noise input has revealed some cells with selectivity for the template image. Similarly, neuronal activity seen between stimulus presentations may be the result of noisy inputs to matched filter cells.
Figure 12.
Schematic localization of recording sites.
(A) Ventral view of the brain with perirhinal cortex (medial, dots) and area TE (hatched) highlighted. Actual recording was done in parts of area TE and perirhinal cortex that are indicated in gray. (B) Coronal cross-section of a standard rhesus monkey atlas (Laboratory of Neuropsychology, NIMH; http://ln.nimh.nih.gov/) at 17 mm rostral to the interaural line (AP +17) showing a recording track into perirhinal cortex and a track into area TE. The noise correlation was found to occur for most area TE neurons that were recorded lateral to the anterior middle temporal sulcus. The noise correlation was not found in the responses of neurons recorded in perirhinal cortex in this or pervious studies recorded medial to anterior middle temporal sulcus [19],[45]. MRs with electrodes can be seen in Liu and Richmond [13]. amts, anterior middle temporal sulcus; rs, rhinal sulcus; sts, superior temporal sulcus; TE, area TE; Prh, perirhinal cortex.