the hmeasure package

Measuring Classification Performance

My hmeasure package averages 800 CRAN downloads per month and is among my most popular contributions. An upcoming Python release and more upcoming developments can be tracked here.  

the datastream.io engine

Anomaly Detection on Data Streams

My startup Mentat Innovations specialised in the ultimate "needle-in-a-haystack" problem, that of identifying anomalous events in a sea of data, without any labels or human supervision. As a result of our work in the area, we released a very popular Python library which has received over 450 stars in github, datastream.io. 

the wsl package

Weakly Supervised Learning

This is my most recent work, exploring midway solutions between supervised and unsupervised learning, in recognition of the fact that real-life problems often hold a mix of partial, noisy, imperfect labels, that cannot easily fit into vanilla machine learning interfaces. A CRAN package is coming up - meanwhile, the code is here.

the smle framework

Streaming learning

Online learning offers great computational advantages by avoiding the need to store or revisit the data history when fresh data arrives. My work in streaming updates for classification algorithms during my PhD led to a general-purpose algorithmic suite for maximum likelihood parameter estimation in data streams. 

the single package

Functional connectivity in fMRI studies is typically estimated by looking at the degree to which brain activity across physically disparate regions is correlated: put simply, does one part of the brain tend to light up whenever another part of the brain does? This connectivity graph can change during an fMRI experiment, as different parts of the brain collaborate to solve various tasks. A disadvantage of the state of the art in this area was that the analysis had to take place after the fact, once all the data had been collected, which gave no access to the connectivity graph in real-time, as the experiment is being performed. Our old CRAN package is currently being revamped and will become available again soon. 

the mns package

The MNS algorithm was proposed as a way to understand the variability of the brains of different subjects, while still respecting the biological fact that a large part of their functional structure is expected to be shared. Drawing inspiration from hierarchical mixed models, a classical technique designed to tease apart inter-subject from intra-subject variability, we applied a similar idea to correlation matrices, resulting in functional connectivity maps where connections between brain regions are characterised as either shared by the population under study or variable across it.