CSML Seminar Series

  • Control with Learning On the Fly: First Toy Problems

    Thu, Oct 10, 2019, 4:30 pm
    How can we control a system without knowing beforehand what the controls do? In particular, how should we balance the imperatives to "explore" (learn what the controls do) and "exploit" (use what we've learned so far to make the system do what we want)? We won't have enough data to apply deep learning. The talk poses several toy problems and solves some of them.
  • Deep Neural Networks for Estimation and Inference: Application to Causal Effects and Other Semiparametric Estimands

    Mon, Oct 14, 2019, 12:30 pm
    We study deep neural networks and their use in semiparametric inference. We prove valid inference after first-step estimation with deep learning, a result new to the literature. We provide new rates of convergence for deep feedforward neural nets and, because our rates are sufficiently fast (in some cases minimax optimal), obtain valid semiparametric inference. Our estimation rates and semiparametric inference results handle the current standard architecture: fully connected feedforward neural networks (multi-layer perceptrons), with the now-common rectified linear unit activation function and a depth explicitly diverging with the sample size. We discuss other architectures as well, including fixed-width, very deep networks. We establish nonasymptotic bounds for these deep nets for nonparametric regression, covering the standard least squares and logistic losses in particular. We then apply our theory to develop semiparametric inference, focusing on treatment effects, expected welfare, and decomposition effects for concreteness. Inference in many other semiparametric contexts can be readily obtained. We demonstrate the effectiveness of deep learning with a Monte Carlo analysis and an empirical application to direct mail marketing.
  • Beyond Supervised Learning for Biomedical Imaging

    Wed, Oct 2, 2019, 4:30 pm
    Many biomedical imaging tasks, such as 3D reconstruction, denoising, detection, registration, and segmentation, are ill-posed inverse problems. In this talk, I will present a flexible machine learning-based framework that has allowed us to derive efficient solutions for a variety of such problems, without relying on heavy supervision. I will primarily employ image registration as a concrete application and present the details of VoxelMorph, our unsupervised learning-based image registration tool. I will show empirical results obtained by co-registering thousands of brain MRI scans where VoxelMorph has yielded state-of-the-art accuracy with runtimes that are orders of magnitude faster than conventional tools. Finally, I will present some recent results where we used VoxelMorph to learn conditional deformable templates that can reveal population variation as a function of factors of interest, such as aging or genetics. Our code is freely available at https://github.com/voxelmorph/voxelmorph.
  • Machine Learning and the Physical World

    Mon, Dec 10, 2018, 4:00 pm

    Machine learning is a data driven endeavor, but real world systems are physical and mechanistic. In this talk we will review approaches to integrating machine learning with real world systems. Our focus will be on emulation (otherwise known as surrogate modeling).

  • Seminar: Latent statistical structure in large-scale neural data: how to find it, and when to believe it

    Tue, Oct 16, 2018, 4:30 pm

    One central challenge in neuroscience is to understand how neural populations represent and produce the remarkable computational abilities of our brains.  Indeed, neuroscientists increasingly form scientific hypotheses that can only be studied at the level of the neural population, and exciting new large-scale datasets have followed.  Capitalizing on this trend, however, requires two major efforts from applied statistical and machine learning researchers: (i) methods for finding latent structure in this data, and (ii) methods for statistically validating that structure.  First, I will discu


Subscribe to CSML Seminar Series