On the 19th of May I was very glad to take part in the RooStats tutorial organised by the AMVA4NewPhysics Network as a part of a workshop in Oviedo. RooStats is a ROOT library that uses the “RooFit” package, and provides classes to perform statistical analysis. The tutorial was attended by all the ESR from our Network, among which I was the only non-physicist. I am a statistician who does not use ROOT at all. For this reason, my attendance at the tutorial could seem Continue reading “My impressions on the RooStats Tutorial”
Have you ever wondered how Facebook suggests the tags for the picture you post on your wall, or how the photo library on your computer manages to automatically create albums containing pictures of particular people? Well, they use facial recognition software based on Convolutional Neural Network (CNN).
CNN is the most popular and effective method for object recognition, and it is a specialized kind of neural network for processing data that has a known grid-like topology. The network employs a mathematical operation Continue reading “Convolutional Neural Networks and neutrinos”
Last time we looked at how we can could fix some of the problems that were responsible for limiting the size of networks we could train. Here we will be covering some additions we can make to the models in order to further increase their power. Having learnt how to build powerful networks, we will also look into why exactly neural-networks can be so much more powerful than other methods.
Continue reading “Understanding Neural-Networks: Part IV – Improvements & Advantages”
So it’s 1986, we’ve got a mathematically sensible way of optimising our networks, but they’re still not as performant as other methods… Well, we know that adding more layers will increase their power, let’s just keep making them larger. Oh no! Now the network no longer trains! It just sits there refusing to optimise. Continue reading “Understanding Neural-Networks: Part III – Diagnosis and treatment”
Welcome back to the second part of my introduction into how neural-networks function! If you missed the first part, you can read it here.
When we left off, we’d understood that a neural network aims to form a predictive model by building a mathematical map from features in the data to a desired output. This map takes the form of layers of neurons, each applying a basic function. The map is built by altering the weights each neuron applies to the inputs. By aiming to minimise the loss function, which characterises the performance of the network, the optimal values of these weights may be learnt. We found that this can be a difficult task due to the large number of free parameters, but luckily the loss function is populated by many equally optimal minima. We simply need to reach one, and can therefore employ the gradient descent algorithm. Continue reading “Understanding Neural-Networks: Part II – Back-propagation”
Below is a short summary of the IML workshop at CERN, which Markus Stoye has also reported on in the previous post.
Day 1 was a discussion with industry experts about the state and future of ML. In the afternoon there was work on the community white-paper that the IML plans to publish. This document is meant to be a road-map for where we want HEP to be in 10 years time with regards to ML. The proto-document is Continue reading “Some More Info on the IML Workshop”
This week the first Inter-experimental LHC Machine Learning IML workshop took place at CERN. I showed my results on using deep learning for hadronic particle labeling (flavour tagging), a method that offers significant improvements in the labeling of heavy flavour jets for the CMS experiment (which I am member of). Despite deep learning as a topic is all over the media, the big CERN experiments have not used it a lot this far. In fact my application is, to my knowledge, the very first deep-learning application in CMS reconstruction.
The workshop featured several presentations on deep learning using Continue reading “Big LHC Experiments Go Deep”
And here it is, the second – but really synchronous in publication with the first – scientific deliverable of our network. Deliverable 4.1, titled “Report of the Performance of Algorithms for Data-Driven Background Shape Modeling“, is a report of studies performed by network members operating within Work Package 4, also known as “New Statistical Learning Tools for HEP Analysis“.
The research presented in this document aims at constructing a precise representation of background processes to searches for small signals in hadron collider data. Specifically, we focused on the multijet QCD background, Continue reading “AMVA4NewPhysics Deliverable 4.1: Report of the Performance of Algorithms for Data-Driven Background Shape Modeling”
It is with a certain satisfaction that I can announce today that the AMVA4NewPhysics network is in complete control of its planned schedule, and has now started to provide real research-grade output, delivering its first two scientific products of relevance. Deliverable 1.1 (from work package 1, which focuses on MVA applications to Higgs boson studies) and Deliverable 4.1 (from work package 4, which focuses on the development of entirely new Machine Learning tools with in mind their application to specific HEP Continue reading “AMVA4NewPhysics Deliverable 1.1: MVA for Higgs Boson Searches at the LHC”