iic.md

From the paper:

Consider now a pair of such cluster assignment variables z and z′ for two inputs xand x′respectively. Their conditional joint distribution is given by P(z=c, z′=c′|x,x′) = Φc(x)·Φc′(x′).This equation states that z and z′ are independent when conditioned on specific inputs x and x′; however, in general they are not independent after marginalization over a dataset of input pairs(xi,x′i), i= 1, . . . , n.

I think this happens because information is lost when summing up observations. Correlation comes into picture in presence of uncertainty, which comes up when less information is available.

IIC Loss:

The official implementation is here. Below are two (quite similar) implementations which are simplified.

def IIC(z, zt, C=10):
    log = torch.log
    EPS = 1e-5
    P = (z.unsqueeze(2) * zt.unsqueeze(1)).sum(dim=0)
    P = (P + P.t())/2 / P.sum()
    P[ (P < EPS).data ] = EPS
    Pi = P.sum(dim=1).view(C,1).expand(C,C)
    Pj = P.sum(dim=0).view(1,C).expand(C,C)
    return (P * ( log(Pi) + log(Pj) - log(P) )).sum()

def IICv2(z, zt, C=10):
    log = torch.log
    EPS = 1e-5
    P = (z.unsqueeze(2) * zt.unsqueeze(1)).sum(dim=0)
    P = (P + P.t())/2 / P.sum()
    P[ (P < EPS).data ] = EPS
    Pi = P.sum(dim=1).view(C,1).expand(C,C)
    Pj = Pi.T
    return (P * ( log(Pi) + log(Pj) - log(P) )).sum()

Why degenerate solutions are avoided. Mutual information expands to I(z, z′) = H(z)−H(z|z′). Hence, maximizing this quantity trades-off minimizing the conditional cluster assignment entropy H(z|z′) and maximising individual cluster assignments entropy H(z). The smallest value of H(z|z′) is 0, obtained when the cluster assignments are exactly predictable from each other. The largest value of H(z) is lnC, obtained when all clusters are equally likely to be picked. This occurs when the data is assigned evenly between the clusters, equalizing their mass. Therefore the loss is not minimised if all samplesare assigned to a single cluster (i.e. output class is identicalfor all samples).

Meaning of mutual information. Firstly, due to the soft clustering, entropy alone could be maximised trivially by setting all prediction vectors Φ(x) to uniform distributions, resulting in no clustering. This is corrected by the conditional entropy component, which encourages deterministic one-hot predictions. For example, even for the degenerate case of identical pairs x=x′, the IIC objective encourages a deterministic clustering function (i.e.Φ(x) is a one-hot vector) as this results in null conditional entropy H(z|z′) = 0. Secondly, the objective of IIC is to find what is common between two data points that share redundancy,such as different images of the same object, explicitly encouraging distillation of the common part while ignoring the rest, i.e. instance details specific to one of the samples.This would not be possible without pairing samples.

Image clustering. IIC requires a source of paired samples (x,x′), which are often unavailable in unsupervised image clustering applications. In this case, we propose to use generated image pairs, consisting of image x and its randomly perturbed version x′=gx. The objective eq. (1) can thus be written as:maxΦ I(Φ(x),Φ(gx))