"Out-of-Distribution Detection based on In-Distribution Data Patterns Memorization
with Modern Hopfield Energy"
link
This method first computes the mean of the internal layer representation of ID data
for each ID class. This mean is seen as the average of the ID activation patterns
as defined in the original paper.
The method then returns the maximum value of the dot product between the internal
layer representation of the input and the average patterns, which is a simplified
version of Hopfield energy as defined in the original paper.
Remarks:
An input perturbation is applied in the same way as in Mahalanobis score
The original paper only considers the penultimate layer of the neural
network, while we aggregate the results of multiple layers after normalizing by
the dimension of each vector (the activation vector for dense layers, and the
average pooling of the feature map for convolutional layers).
"Out-of-Distribution Detection based on In-Distribution Data Patterns Memorization with Modern Hopfield Energy" link
This method first computes the mean of the internal layer representation of ID data for each ID class. This mean is seen as the average of the ID activation patterns as defined in the original paper. The method then returns the maximum value of the dot product between the internal layer representation of the input and the average patterns, which is a simplified version of Hopfield energy as defined in the original paper.
Remarks: