Group-invariant max filtering (Q6975880)

This interesting paper speaks to a common problem in modern machine learning when the number of data points is small.\N\NIf the number of data points is large then one is able to use a variety of modern tools to deal with matters such as the curse of dimensionality and estimates of dimension for example.\N\NThese techniques range for example from manifold learning to nonlinear dimension reduction and many others. Recent papers have dealt with data points of small cardinality by various techniques for example to controlling local neighborhoods of the points.\N\NThis paper proposes an interesting idea to deal with point sets of small cardinality. One idea to deal with this problem is to introduce small distortions in the data whose actions correspond to that of a group meaning the ground truth exhibits known symmetries. This encourages the data set to learn such actions, often symmetries, and they of course increase the cardinality of the set of data points. Although this idea has proved successful in many applications, it is often inefficient to train a large, symmetry-agnostic model to find a highly symmetric function. This then naturally begs the question of why not use a model which accounts for known symmetries and this is the subject of invariant machine learning.\N\NThe authors do the following. Let us be given a group \(G\) of linear isometries and a real inner product space \(V\). The authors then construct what they call max filters which are a family of \(G\)-invariant real-valued functions on \(V\). For a special case to give some insight. Take \(V = R^d\) and \(G\) finite. Then a a suitable max filter bank separates orbits, and is bilipschitz in the quotient metric. In the case where \(V = L_2(R^d )\) and \(G\) is the group of translation operators, a max filter shows stability to diffeomorphic distortion. This is similar to the scattering transform which has been introduced by Mallat. The authors show that max filters are well suited for various classification tasks, both in practice and theory.\N\NThe paper is well written with an excellent set of references.