## Week 1

See the Conference program page

During Weeks 2-5, the program will host 1 or 2 talks on Monday, Wednesday and Fridays, followed by coffee. The list of talks will be updated shortly. All the talks will be held in the ISB 102 seminar room.

### Week 2

**Monday:**

- 12:00-1:00: Student presentations
- 1:00-1:30: Coffee

**Wednesday: **

- 12:00-1:00: ArXiv discussion
- 1:00-1:30: Coffee

**Friday**

- 12:00-1:00:
**Rich Caruana** - Title: Friends Don’t Let Friends Use Black-Box Models in Science: The Importance of Interpretability in Machine Learning
- Abstract: If you’re going to use machine learning in science, it would be nice to be able to understand what your models have learned --- it isn’t really science until you understand it. Unfortunately, historically there has been a tradeoff between accuracy and intelligibility: accurate models such as deep neural nets, boosted tress and random forests are not very intelligible, and intelligible models such as linear regression and decision trees usually are less accurate. But this is changing. We have developed a learning method based on generalized additive models with pairwise interactions (GA2Ms) that is as accurate as full complexity models on many datasets, yet even more interpretable than linear regression. In this talk I’ll show how interpretable, high-accuracy machine learning is making what was previously hidden, visible and helping us discover what our models have learned and uncover flaws lurking in our data. Code for GA2Ms is available at https://github.com/microsoft/interpret
- 1:30-2:00: Coffee

### Week 3

Monday:

- 12:00-1:00:
**Daisuke Nagai** - Title: Physics-based vs. Data-driven Approach in the Era of Multi-Wavelength Astronomical Surveys
- Abstract: A plethora of multi-wavelength astronomical surveys are underway to create high definition (HD) maps of the universe across the electromagnetic spectrum. These multi-band HD images will produce detailed maps of dark matter, gas and stars in the universe and provide a new platform for studying the inner workings of our cosmos, from the smallest constituents of matter to the largest k nown structures. Major challenges facing this cosmic frontier include (1) extreme dynamic range, (2) complexity of astrophysical processes, and (3) detecting small signals in massive, complex and noisy data. Solving this multi-scale, multi-physics and big data problem requires creative and effective use of emerging tools and techniques, including computer simulations, theoretical modeling and machine learning. In this talk, I will discuss our research program to advance the use of complex systems, such as galaxy clusters and cosmic web, as laboratories for cosmology and astrophysics.

- 1:00-1:30: Coffee

Wednesday:

- 12:00-12:30:
**Chirag Modi** - Title: Learning Halo & Stellar Mass field with applications to Cosmological Reconstruction
- Abstract : In this talk, I will discuss different ways of learning halo and stellar mass fields with specific application to reconstruction of cosmological field, though these methods can be more generically applicable. Our reconstruction relies on modeling the observed data field from initial conditions in a differentiable manner and following its gradients in backward pass. Since traditional methods such as Friends-of-Friends to model these fields are non-differentiable and have high resolution requirements, I will investigate a number of differentiable approaches to do so. For intensity mapping data,, briefly I will show that the Lagrangian bias model is sufficient. For the main part of my talk, I will focus on the discrete, point-like galaxy survey data. We propose a simple framework to model halo(galaxy) positions and masses using fully connected neural networks and show that our reconstruction improves over standard methods in both real and redshift space. I also discuss how this data modeling can be improved and the data likelihood learnt using other deep network architectures. Along these lines, I will also introduce FlowPM, a PM code completely written in Tensorflow that allows one to use extensive Tensorflow machinery for machine learning.
- 12:30-1:00:
**Biwei Dai** - Title: From dark matter to baryons with physics based machine learning model
- Abstract: Running cosmological hydrodynamic simulations to model baryons and produce mock data is computationally expensive. In this talk I propose physics based models to improve the matter field from dark matter only simulations, and to produce hydro outputs from dark matter distribution. In the first part of the talk, I will introduce a scheme to calibrate fast simulations to mimic the precision of the hydrodynamic simulations or high resolution N-body simulations. The scheme is based on a gradient descent of either effective gravitational potential, which mimics the short range force, or of effective enthalpy, which mimics gas hydrodynamics and feedback. I will show that the model can improve the matter power spectrum, halo internal structures and weak lensing observables. It can also be used to add baryonic effects to full N-body simulations. In the second part of the talk, I will generalize the scheme and use it to generate hydro outputs. The model takes dark matter particles as input, moves the particles based on local density, and then transforms the particle density field to produce hydro observables. The particle movement and field transformation can be regarded as baryonic process such as gas cooling, star formation and AGN feedback. I will show that this model is able to produce maps of stellar mass, tSZ, kSZ, X-ray and HI of various redshifts. Comparing to other machine learning models such as GAN and CNN, this physics based model has the advantage that it has much fewer parameters (typically less than 10) to optimize.
- 1:00-1:30: Coffee

Friday

- 12:00-1:00:
**Ekin Dogus Cubuk** - Title: Learning order parameters in physics and symmetries in images.
- Abstract: I will talk about our research at the intersection of physics and deep learning. On the deep learning side, we have been taking advantage of symmetries already present in natural images to train more accurate and more robust models on images, ranging from classification and object detection to speech recognition, where the symmetries are found via meta-learning (RL, evolution, and meta-gradients). On the physics side, we have been using machine learning to find predictive order parameters for modeling solid state systems, ranging from supercooled liquids to poly-crystals. If time permits, I will talk about our newly open-sourced molecular dynamics package, jax-md, which aims to apply some of the meta-learning approaches to the simulation of solids.
- 1:00-1:30:Coffee

### Week 4

Monday:

- 12:00-1:00:
**Francois Lanusse**- Title: Deep Learning for Modern Cosmological Surveys: from Image Processing to Physical Inference
- Abstract: The upcoming generation of cosmological surveys such as LSST will aim to shed some much needed light on the physical nature of dark energy and dark matter by mapping the Universe in great detail and on an unprecedented scale. While this implies a great potential for discoveries, it also involves new and outstanding challenges at every step of the science analysis, from image processing to the cosmological inference. In this talk, I will present how these challenges can be addessed with some of the latest developments in Deep Learning, in particular graph neural networks, deep generative models, and neural density estimation. At the image level, I will demonstrate how deep convolutional networks can be used for low level data-reduction, for instance to efficiently detect rare events or help reliably measure galaxy properties, allowing us to make optimal use of increasingly large and complex datasets. Another important aspect of the analysis of modern surveys is our ability to generate realistic mocks of the observations. In situtations where physical models either do not exist or are intractable, I will present how deep generative models can be used as an alternative, with as an example, learning to generate realistic galaxy intrinsic alignments inside large volume cosmological simulations. Finally, I will present how neural density estimation can be used for perform dimensionality reduction and inference, allowing us to build complex and very sensitive summary statistics of the data, and to use them in a consistent Bayesian framework to help constrain our cosmological model.

- 1:00-1:30: Coffee

Wednesday:

- 12:00-12:30:
**Michelle Lochner** - Title: Probabilities in Machine Learning
- Abstract: Machine learning has already proven to be an invaluable tool in many fields of science. However, when it is incorporated into sophisticated scientific analyses, careful attention must be paid to how to determine and propagate uncertainties from these algorithms, to avoid biases and incorrect results. Bayesian statistics is increasingly being included into machine learning approaches to improve robustness and quantify uncertainty. In this talk I will use a fun toy example to demonstrate the impact that biased probabilities from machine learning can have on scientific analyses, as well as discuss some techniques to improve uncertainty estimates from machine learning.
- 12:30 - 1:00:
**Marc Huertas-Company**- Title: The Hubble sequence at z~0 in the Illustris TNG simulation
- Abstract: I will very briefly show some of the main results of this paper : in which we analyze the optical morphologies of simulated galaxies using a simple CNN.

- 1:00-1:30: Coffee

Friday

- 12:00-1:00:
**John Tamanas** - Title: Neural Density Estimation
- Abstract: Neural density estimators (NDEs) are probability density models parametrized by neural networks. In this talk, I will cover how NDEs work and ways they can be used to improve inference and data generation methods. Finally, I will end with an interactive example where we will perform likelihood-free inference on a high-dimensional parameter space in supersymmetry.
- 1:00-1:30: Coffee

### Week 5

Monday:

- 12:00-1:00: Student Seminars
- 1:00-1:30: Coffee

Wednesday:

- 12:00-1:00: Student Seminars
- 1:00-1:30: Coffee

Friday

- 12:00-1:30: Seminars
- 1:30-2:00: Coffee

### Week 6: Fellows Presentations.

All talks will be held in ISB 102.

**Monday: **No seminar

Tuesday: No seminar

**Wednesday:**

**11:00-11:45: Humna Awan**: Extracting 3D Galaxy Shape From 2D Image Properties using Random Forests**11:45-12:30: Dezső Ribli**: Domain adaptation with generative adversarial networks in astrophysics**12:30-1:15: Ting-Yun Cheng**: Better than BigGAN? - The application of Vector Quantized Variational Autoencoder from Google DeepMind in Astrophysics**1:15-2:00: Emma Platts**: FRBs and KDEs: Constraining the Dispersion Measure of the Galactic Halo

**Thursday**

**11:00-11:45: Lorenzo Zanisi**: Comparing simulations and observations with deep autoregressive generative models**11:45-12:30: Kate Storey-Fisher**: Anomaly Detection in Hyper Suprime-Cam Images with Generative Adversarial Networks**12:30-1:15: Maddie Lucey**: Declumping the Red Clump**1:15-2:00: Connor Bottrell**: Stream and shell identification in HyperSuprime-Cam PDR2 with deep learning: Toward detailed characterization of galaxy tidal features**2:00-2:45: Mike Walmsley**: Hunting 1 million AGN with Photometry

**Friday**

**11:45-12:30: Claire Guilloteau**: Generative models for emulating synthetic sky images**12:30-1:15: Szymon Nakoneczny**: Toward 1% calibration of CMB Lensing Cluster Mass Estimate with Machine Learning**1:15-2:00: Daniel Muthukrishna**: Real-time transient anomaly detection with recurrent neural networks**2:00-2:45: Paulo H. Barchi**: Extracting the Slime Mold Graph from the Cosmic Web**2:45-3:30: Sara Webb**: Exploring millions of light curves with unsupervised learning and anomaly detection- Closing remarks