Correction for “Emergent neuronal mechanisms mediating covert attention in convolutional neural networks,” by Sudhanshu Srivastava, William Yang Wang, and Miguel P. Eckstein, which published November 13, 2025; 10.1073/pnas.2411909122 (Proc. Natl. Acad. Sci. U.S.A. 122, e2411909122).
The authors note that, due to a production error, Fig. 6 appeared incorrectly. The labels within Fig. 6H did not include their colors corresponding to the scatterplot. The online version has been corrected.
Fig. 6.
Neuronal Cueing across layers and cueing mechanisms. (A) Layerwise neuronal cueing for 80% (orange) and 50% (blue) cue validity. (B) Response distributions pre- and post-ReLU of an example flat-layer TC neuron when the target is present at the left location (orange) and absent (blue). The Top row shows responses when the cue is absent, the Bottom row shows responses when the cue is present at the left location. (C) Linear RF obtained via linear regression predicting neuronal response for a location-summation neuron with cue on the Left (above blue columns) and the Right (above green columns). Bar plots depict mean positive and negative RF weights for each location and condition. (D) Linear RFs predicting neuronal response for a location-opponent neuron with cue on the Left (above blue columns) and the Right (above green columns). Note that the RF for the left location of the location-opponent neuron (TL-) is excitatory to the vertical distractor and inhibitory to the tilted target. (E) Linear RFs and associated weights for the network’s output neuron representing target presence. Error bars indicate SE over 10 reruns. (F) Linear RFs and associated weights for BIO model (log of weighted likelihood ratio). (G) Histograms of neuronal cueing by mechanism types. The different colors correspond to different mechanism types consistent with the legend of graph (H). (H) Scatterplots depicting the relationship between the dense layer units’ neuronal cueing and their weights projected onto the output layer for each mechanism type in the dense layer. (I) Simplified circuit diagram depicting how the three dense-layer attention mechanisms (location summation, opponency, and summation/opponency) receive input from the three flat neuron types, and feed into the CNN output neuron. Arrow widths between dense and output layers depict the mean absolute regression weight for each dense neuron mechanism type for predicting the output responses. Arrow widths between the flat and dense layer depict the mean absolute regression weights for each flat neuron type for predicting the average response of each dense mechanism type (see SI Appendix, section S.25 for details). (J) Valid and invalid hit rates of the CNNs (blue) and linear regression (orange).