John T. Halloran
ML Scientist, Amazon
| About |
|
I am a machine learning scientist at Amazon, where I focus on methods to effectively train deep generative models. Before Amazon, I was a postdoc at UC Davis. My previous research focused primarily on two
different approaches to accelerate the training of
widely used machine learning models:
I received the UC Davis award for
Excellence in Postdoctoral Research, and a Genome
Training Grant while at UW. |
| Recent News |
|
| Selected Publications |
| See my publications page for the full list. |
|
| Software |
| ProteoTorch: Deep Semi-Supervised Learning for Accurate Recalibration of Shotgun Proteomics Data |
ProteoTorch faithfully implements the
semi-supervised learning framework (with
cross-validation) pioneered by the C++ Percolator software, in a
light and flexible Python package with an emphasis
on speed. By default, deep neural network
classifiers are used to accurately recalibrate PSMs
and features collected from a database-search. A
host of other classifiers are also available,
including ultrafast support vector machine (SVM)
implementations.
|
| GPU-Optimized Logistic Regression and SVM training algorithms |
The following contains highly optimized
GPU-training primal algorithms for logistic regression (in LIBLINEAR for
sparse features) and SVMs (in Percolator for dense
features). The underlying CPU-centric algorithms
considered natively resist GPU-optimizations, due to
the heavy sequential dependence of variables in the
underlying algorithm (i.e., the trust-region second-order
algorithm, TRON) and reliance on random access.
Thus, the code enables GPU-optimizations by
extensively using the following principles: 1) decouple CPU and GPU variable
dependencies, (2) minimize transfer latency between
the CPU and GPU, (3) saturate the GPU optimally by
using routines which allow memory coalescing, and
(4) maximize concurrency between the CPU and GPU.
|
| Jensen: A General Toolkit for Large-Scale Machine Learning and Convex Optimization |
Written in C++, Jensen is an easily-customizable
toolkit for production-level machine learning and
convex optimization. Fast, flexible, and light on
external dependencies (only CMake is necessary to
build the source), Jensen natively supports a large
number of popular loss functions,
state-of-the-art optimization algorithms, and
machine learning applications.
Documentation and code
examples are
described here.
The software is freely
available here and supported on Unix, OSX, and
Windows operating systems.
|
| Optimized Percolator Software |
The following repository contains SVM solvers
highly optimized for large-scale Percolator
analysis: bitbucket.org/jthalloran/percolator_upgrade.
Both solvers, Trust Region Newton (TRON) and
L2-SVM-MFN*, support multithreading and are
optimized for single-threaded use. Further details
may be found in our
paper, A
Matter of Time: Faster Percolator Analysis via Efficient SVM
Learning for Large-Scale Proteomics.
|
| The DRIP Toolkit |
The DRIP Toolkit (DTK), for searching a tandem mass spectra using a dynamic Bayesian network (DBN) for Rapid Identification of Peptides (DRIP), is now available! DTK supports parameter estimation for low-resolution MS2 searches, multithreading on a single machine, utilities easing cluster use, instantiating/decoding/plotting DRIP PSMs in the python shell, and in-browser analysis of identified spectra via the Lorikeet plugin. Further information and documentation regarding the toolkit's use is available in the DRIP Toolkit documentation.
Details of the DRIP model may be found
in here.
|
| HMM-DNN tutorial for GMTK |
I've written a short tutorial for the Graphical Models Toolkit (GMTK), with all pertinent files available in this tarball. The following is a copy of the tutorial's documentation. This tutorial covers training an HMM in GMTK via generative training (expectation maximization), discriminative training (maximum mutual information), and training an HMM/DNN hybrid. |
| Contact Info |
|
| Miscellaneous |
|
