Caricature Identification with Neural Nets
Joanne Yuan
Cambridge, Massachusetts
- 0 Collaborators
It has been found that humans can easily identify identity from caricatures -- we want to see if the same is true of Deep Neural Nets. We trained AlexNet to identify identities of non-caricature images, then tested the net on caricature images. ...learn more
Project status: Under Development
Intel Technologies
Intel Python
Overview / Usage
The overarching goal of the project is to determine the extent of similarities between the human brain and DNNs. Are the behaviors replicable? The specific aspect that we're working on is determining how DNN's handle caricatures -- while we know that humans are adept at identity recognition from caricatures, we wonder if the same is true for a DNN that has not yet been trained on such images.
Methodology / Approach
We trained AlexNet on a dataset of different identities, structured such that each folder within the dataset contained >400 images of one identity. As a baseline, we tested AlexNet on a classification task (identifying people given an image) on a held-out test dataset. As expected, it achieved high (>90% top5 accuracy). We then tested the network on a separate set of caricatures (which it had never trained on before -- the original dataset contained only natural images). We found that the network was able to reach a surprising level of accuracy (~52% top1 accuracy and ~76% top5 accuracy), showing that it was able to generalize well, even to images clearly outside its training regime. This finding suggests that the human ability to recognize caricatures is an emergent property of DNNs designed for face recognition.
Technologies Used
- Python
- PyTorch