High-throughput methods for quantifying neuroanatomy

Along with a team of researchers at Argonne National Laboratory, the lab is working to develop data analysis pipelines for mapping brains with X-ray microtomography. Our results demonstrate that X-ray sources can be used with image parsing techniques to rapidly quantify neuroanatomy at the mesoscale over large brain volumes without sectioning. We are currently exploring ways in which our techniques can be used to investigate intact thalamocortical pathways, as well as combined with electron microscopy to obtain multi-modal brain maps.

Related publications:

  • Dyer et al., 2016. Quantifying mesoscale neuroanatomy using X-ray microtomography (Web, Paper)

Low-dimensional signal models

Unions of subspaces (UoS) are a generalization of single subspace models that approximate data points as living on multiple subspaces, rather than assuming a global low-dimensional model (as in PCA). Modeling data with mixtures of subspaces provides a more compact and simple representation of the data, and thus can lead to better partitioning (clustering) of the data and help in compression and denoising.

Related publications:

  • E.L. Dyer, A.C. Sankaranarayanan, and R.G. Baraniuk, Greedy feature selection for subspace clustering, The Journal of Machine Learning Research 14 (1), 2487-2517, September, 2013. (Paper)
  • E.L. Dyer, T.A. Goldstein, R. Patel, K.P. Körding, and R.G. Baraniuk, Sparse self-expressive decompositions for dimensionality reduction and clustering (Paper)
  • R.J. Patel, T.A. Goldstein, E.L. Dyer, A. Mirhoseini, and R.G. Baraniuk, Deterministic column sampling for low rank approximation: Nystrom vs. Incomplete Cholesky Decomposition, SIAM Data Mining (SDM) Conference, May 2016. (Paper, Code)

Large-scale optimization

Optimization problems are ubiquitous in machine learning and neuroscience. The lab works on a few different topics in the areas of non-convex optimization and distributed learning. In a recent paper, we demonstrate how the low rank and multi-subspace structure of large datasets can be leveraged to accelerate a broad class of iterative optimization methods. In addition, are also working on new methods for non-convex and black-box optimization. In a recent paper at UAI 2016, we introduced a provable black-box approach for global optimization that learns a convex envelope from samples of the function.

Related publications:

  • A. Mirhoseini, E.L. Dyer, E. Songhori, R.G. Baraniuk, and F. Koushanfar, RankMap: A platform-aware framework for distributed learning from dense datasets, IEEE Trans. on Neural Networks and Learning Systems, 2017. (Paper, Code)
  • M Gheshlaghi Azar, E.L. Dyer, Konrad Kording, Convex Relaxation Regression (CoRR): Black-box optimization of a smooth function by learning its convex envelope, Proc. of the Conference on Uncertainity in Artificial Intelligence, 2016. (Paper)

Analyzing the activity of neuronal populations

Advances in monitoring the activity of large populations of neurons has provided new insights into the collective dynamics of neurons. I am working on methods that learn and exploit low-dimensional structure in neural activity for decoding, denoising, and deconvolution.

Related publications:

  • E.L. Dyer, M. Azar, H.L. Fernandes, M. Perich, L.E. Miller, and K.P. Körding: A cryptography-based approach to brain decoding, to appear in Nature Biomedical Engineering, 2017. (Web, Paper)
  • E.L. Dyer, C. Studer, J.T. Robinson, and R.G Baraniuk, A robust and efficient method to recover neural events from noisy and corrupted data, IEEE EMBS Neural Engineering Conference, 2013. (Paper, Code)