Fig 1.
In all three experiments, the task involved two-choice discrimination between stimulus configurations oriented clockwise and counterclockwise from the vertical. The upper and lower panels show examples of low- and high-energy stimuli respectively for each experiment. In all three examples, the correct choice is “counterclockwise”. (A) Task used by Herce Castañón et al. (2019) [13]. The stimulus consisted of an array of eight noisy Gabor patches with the task involving judgements of mean orientation relative to horizontal. Energy manipulations involved jointly changing the contrast of Gabors as well the variability of orientations across the array. (B) Task used by Koizumi et al. (2015) [4]. The stimulus consisted of two superimposed sinusoidal gratings overlaid by a noise mask. The task was to determine the orientation of the grating with the higher contrast (dominant grating). Increases in energy involved jointly increasing the contrast of the dominant and the non-dominant gratings. (C) The stimulus was a single Gabor patch overlaid with noise and the task was to determine its orientation. Energy was manipulated by jointly changing the contrast and noise level in the patch.
Fig 2.
Confidence-accuracy dissociations in CNNs.
For all experiments and networks (custom 4-layer CNN, VGG-19 and ResNet-50), accuracy was matched across energy conditions, but confidence significantly increases with energy levels. The violin plots show the kernel density estimates of the data distribution. *p<0.05; **p<0.01; ***p<0.001; ****p<0.0001; n.s., not significant.
Fig 3.
Energy manipulations increase both the separability and variance of the networks’ output layer activations.
For all three experiments, the separability between the distributions of evidence for the two stimulus categories, as well as the variance of the evidence distributions, increased with energy levels. For each network, the figure shows the distance between the S1 and S2 evidence distributions and their standard deviations (SD) across the 25 model instances. We note that these networks appear to represent evidence on their own unique internal axis. Therefore, to optimally visualize differences in their evidence distributions, each network’s distributions have been plotted on their own unique scale. The kernel density plots show the distribution of activations aggregated across all 25 network instances. *p<0.05; **p<0.01; ***p<0.001; ****p<0.0001; n.s., not significant.
Fig 4.
Changes in the separation and spread of evidence distributions induced by energy, contrast, and variability manipulations.
The plots show A) the average distance between the mean activations for S1 and S2 stimuli (left) and B) the average standard deviation (SD) of activations (right) in the final layer of the shallow CNNs for Experiments 1–3 in response to changes in stimulus energy, contrast and variability. Note that the energy results in both panels are equivalent to the 4-layer CNN results from Fig 3. For all experiments, increasing energy and contrast levels increases the separation between the two stimulus categories, while increasing variability decreases the separability between the two stimulus categories. On the other hand, increasing stimulus energy and variability increases the spread of evidence distributions, while increasing contrast decreases the spread of evidence for all experiments except Experiment 3 (where increasing contrast increases the spread of evidence). These results suggest contrast and noise changes selectively drive changes in separation and variance of evidence distributions respectively. The violin plots show the kernel density estimates of the data distribution. *p<0.05; **p<0.01; ***p<0.001; ****p<0.0001; n.s., not significant.
Fig 5.
Dissociations between meta-d’ and d’.
We tested the three networks (4-layer CNN, VGG-19 and ResNet-50) on the task paradigm from Maniscalco et al. (2016) [18] which demonstrated a dissociation between d’ and meta-d’ when the contrast of one stimulus category (S1) remains fixed while the contrast of the other stimulus is increased in discrete steps (S2) For this design, meta-d’ increases with d’ as expected for trials in which the observer responds "S2", but meta-d’ decreases with d’ for trials where the observer responds "S1". Maniscalco et al. (2016) [18] showed that this behavioral effect can be explained by a model incorporating the positive evidence bias. Here, we simulated this task paradigm for the stimuli in Experiments 1–3. The responses generated by our networks show that they can indeed generate the meta-d’/d’ dissociations observed in humans for at least two out of three experiments. While the 4-layer network fails to reproduce this behavior for Experiment 2, VGG-19 and ResNet-50 fail to produce this behavior for Experiment 3, suggesting that these dissociations may depend on specific interactions between the stimuli and the networks.
Fig 6.
Confidence-accuracy dissociations in a color discrimination task.
(A) The stimulus consisted of an array of eight colored circles. The task was to determine whether the mean color across the eight patches was more blue or red. In this example, the mean color is more blue than red. Energy manipulations involved joint changes to the intensity of color (the amount of “blueness” or “redness” of the patches as well the variance in color across the array. (B) The networks’ accuracy was matched across energy conditions, but confidence significantly increased with energy levels. (C) The separability between the stimulus categories as well as the variance of the evidence distributions increased with energy levels for all three networks. The panels on the top-left for each network show the average distance between the S1 and S2 evidence distributions across the 25 model instances. The panels on the bottom-left show the average standard deviation (SD) across the two distributions across all model instances. The panels on the right show the distribution of activations aggregated across all 25 network instances. *p<0.05; **p<0.01; ***p<0.001, ****p<0.0001; n.s., not significant.