A more efficient loss function for Siamese

At work, we are working with Siamese Neural Net (NN) for one shot training on telecom data. Our goal is to create a NN that can easily detect failure in Telecom Operators networks. To do so, we are building this N dimension encoding to describe the actual status of the network. With this encoding we can then evaluate what is the status of network and detect faults. This encoding as the same goal as something like word encoding (Word2Vec or others). To train this encoding we use a Siamese Network [Koch et al.] to create a one shot encoding so it would work on any network. A simplified description of Siamese network is available here. For more details about our experiment you can read the blog of my colleague who is the master brain behind this idea.
Continue Reading "Lossless Triplet loss"
Lately, at work, we had to do a lot of unsupervised classification. We basically had to distinguish N classes from a sample population. We had a rough idea of how many classes were present but nothing was sure, we discovered the Kolmogorov–Smirnov test a very efficient way to determine if two samples are significantly different from each other. I will give you a bit of context on the Kolmogorov–Smirnov test and walk you though one problem we solved with it. A bit of theory Rejecting the null hypothesis. That sounds...Continue Reading "Kolmogorov–Smirnov test"