Deep structures

Joint workshop by Alan Turing Institute and Finnish Center for Artificial Intelligence

Espoo, Finland, December 19th and 20th, 2019

Deep models have become the dominant approach of machine learning, yet we still have trouble understanding them and qualitative progress with neural networks has arguably plateaued. Simultaneously, several deep learning paradigms have emerged recently, proposing novel learning frameworks such as deep processes [1], normalizing flows [2], deep generative models [3], Bayesian deep models [4], Riemannian geometries [5], and black box (ordinary/stochastic/partial) differential equations [6,7,8] as learning machines. We find these emerging approaches extremely exciting, and want to bring together both machine learning, and statistics & numerics, and dynamical systems & stochastics researchers to tackle both current and new directions in Helsinki. Attendance is by mostly by invitation.

[1] Dunlop et al. How deep are deep Gaussian processes? JMLR’18
[2] Grathwohl et al. FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models, NeuRIPS’18
[3] Kingma et al. Auto-encoding variational Bayes, arXiv‘13
[4] Chua et al. Deep reinforcement learning in a handful of trials using probabilistic dynamics models, NeurIPS’18
[5] Bronstein et al. Geometric deep learning: going beyond euclidean data, IEEE Signal Processing ‘17
[6] Chen et al. Neural ODE, NeurIPS’18
[7] Hegde et al. Deep learning with differential GP flows, AISTATS’19
[8] Owhadi et al. Kernel flows: from learning kernels from data into the abyss, J. Comput. Phys. ’19

For any questions contact markus.o.heinonen@aalto.fi

Program

Thursday (19th Dec)
08:15 - 08:30	Samuel Kaski - Opening
08:30 - 09:30	David Duvenaud - Neural Stochastic Differential Equations for Irregularly-Sampled Time Series
09:30 - 09:45	George Papamakarios - Neural Spline Flows
09:45 - 10:15	Coffee
10:15 - 11:00	Neill Campbell - Learning Alignments and Monotonic Gaussian Process Flows
11:00 - 11:45	Carl Henrik Ek - Composite Uncertainty in Deep Gaussian Processes
12 - 13	Lunch
13:00 - 13:45	Simo Särkkä - Hilbert-Space Reduced-Rank Methods for Deep Gaussian Processes
13:45 - 14:30	Sören Hauberg - Only Bayes Should Learn a Manifold
14:30 - 14:45	Gregor Simm - GraphDG: A Generative Model for Molecular Distance Geometry
14:45 - 15:15	Coffee
15:15 - 16:00	Mauricio Alvarez - Polynomial Latent Force Models
16:00 - 16:45	Aki Vehtari - Diagnosing pre-asymptotic behavior of importance sampling and related Monte Carlo expectations
16:45 - 17:15	Markus Heinonen - Differential deep learning
17:15	Poster session
19:30	Dinner

Friday (20th Dec)
08:30 - 09:15	James Hensman - Translation insensitive convolutional kernels
09:15 - 10:00	Niklas Wahlström - Deep Learning Applied to System Identification
10:00 - 10:30	Coffee
10:30 - 11:15	Andreas Damianou - From GP to deep learning and from deep learning to GP
11:15 - 12:00	Ole Winther - Flows for quantized data - auto regressive models and variational interpretation
12 - 13	Lunch
13:00 - 13:45	Arno Solin - The unholy alliance: Encoding probabilistic priors into deep vision models
13:45 - 14:30	Jaakko Lehtinen - On Generative Adversarial Networks and the distributions they learn
14:30 - 15:00	Coffee
15:00 - 15:15	Tim Rudner - Inter-domain Deep Gaussian Processes with RKHS Fourier Features
15:15 - 15:30	ST John - Multiple dispatch in the inducing variable and multi-output framework in GPflow
15:30 - 15:45	Takuo Matsubara - Quadrature of neural networks based on theridgelet transform, and the possibility of the extension

Venue

The workshop takes place in the Palaver-room at the Dipoli-building at the Otaniemi campus in Espoo. The venue is a five minute walk from the "Aalto University" subway station connecting downtown Helsinki.

Dipoli
Otakaari 24
02150 Espoo, Finland

Organizers

The conference is jointly organized by:

Samuel Kaski, Aalto University
Mark Girolami, ATI
Markus Heinonen, Aalto University

We are grateful for funding from the Helsinki Institute for Information Technology HIIT.