2022 Flatiron Machine Learning X Science Summer School

Maria Avdeeva

Maria Avdeeva will co-mentor on the following topics:

1. Dimensionality reduction for large-scale tissue flows in developing embryos
2. Representation and modeling of spatiotemporal gene expression datasets

Kyunghyun Cho

1. Dissecting variance in deep learning: in this topic, a student (or a team of students) chooses one or two benchmark tasks, trains a set of deep neural nets on each of these tasks while varying as many sources of variations as possible (see, e.g., https://kyunghyuncho.me/how-to-think-of-uncertainty-and-calibration-2/) and analyzes the impact of these sources of noise on the final variance of each of different outcome metrics, such as accuracy, entropy, etc.

SueYeon Chung

1. Exploring the connection between the structure in the neural network connectivity (e.g. synaptic weights) and the geometry of neural representations, both in feedforward and recurrent neural network settings.
2. Exploring the efficiency of various bioplausible learning rules from theoretical perspectives (e.g. their match to the backpropagation, and the efficiency of the learning dynamics), with a particular focus on broadcasting global error signals, including but not limited to Global Error Vector Broadcasting (GEVB) rule.

Miles Cranmer

1. Learning new analytic models in astrophysics using symbolic regression
2. Using deep learning to accelerate symbolic regression and model discovery
3. Applications of machine learning to fluid dynamics

Domenico Di Sante

1. What physics can we learn about interacting electrons by using artificial intelligence based methods? The research will be conducted within the framework of the NeuralFRG architecture that we have recently developed at CCQ to learn the functional Renormalization Dynamics for correlated fermions (arXiv:2202.13268). The technical part of the research program will also aim at: i) restructuring the code from Python to Julia, in order to efficiently benefit from the Adjoint method for training Neural ODE solvers; ii) extension to other kinds of fRG dynamics; iii) implementation of multi-GPU training.
2. Application of dimensionality reduction algorithms to the ab-initio molecular dynamics for solid state systems. Use of deep-learning architectures to find representations of Koopman eigenfunctions from data and identify coordinate transformations that will make non-linear ab-initio molecular dynamics approximately linear. The aim is to enable nonlinear prediction of these ab-initio dynamics, their extrapolation in future and control using linear theory.

Barbara Engelhardt

1. Live-cell imaging is an inexpensive way to study the behavior of modified cells. But the methods to fully characterize behavior of cells in these videos do not exist. In this topic, we will use CAR-T live cell imaging videos to motivate the development of machine learning methods that describe the behavior of these modified cells.
2. While single-cell RNA-sequencing has allowed an unprecedented look at transcription within single cells, methods to quantify variation and associate transcription levels with genotypes including expression QTL methods rely heavily on pseudo-bulk methods. In this topic, we will build ML methods for association testing in single cell RNA-seq data that take advantage of the multiple cell read-outs per sample.

Siavash Golkar

1. Developing a framework using analytical tools (NTK, replica theory etc.) to do active learning (finding relevant training samples that maximally improve generalization) and neural architecture search.
2. How does curriculum learning affect adversarial robustness of a neural net? A possible explanation of adversarial susceptibility of neural networks is that they are trained on complicated datasets from scratch. This is akin to taking advanced classes without knowing the prerequisites and learning the wrong lessons.

Robert M. Gower

1. Testing and designing stochastic optimization methods for learning models that almost interpolate.
2. A new natural policy gradient method: Exploiting curvature for RL.

Shirley Ho

1. Explore astrophysical applications of NLP models.
2. Explore the neural tangent kernel and neural architecture search.

Pearson Miller

Pearson Miller will co-mentor on the following topics:

1.Dimensionality reduction for large-scale tissue flows in developing embryos
2. Representation and modeling of spatiotemporal gene expression datasets

Natalie Sauerwald

1. While we often think of genes as functional units of the genome, the same gene can be modified to form multiple isoforms which can vary significantly in their respective functions. This topic will apply machine learning methods to integrate isoform and structural data to better understand the functional implications of different genetic isoforms.

Anirvan Sengupta

1. Learning dynamics from data: We would like to benchmark different models for learning latent dynamics for scientific problems.

Stanislav Y. Shvartsman

Stanislav Y. Shvartsman will co-mentor on the following topics:

1. Dimensionality reduction for large-scale tissue flows in developing embryos
2. Representation and modeling of spatiotemporal gene expression datasets

David Spergel

1. Application of machine Learning in cosmological challenges

Alex Williams

1. This topic will analyze camera and microphone array recordings of Mongolian gerbil families in collaboration with labs run by professors Dan Sanes and David Schneider at NYU. Gerbils are highly social rodents — more so than typical laboratory species — and they produce elaborate vocal call sequences in their social interactions. From these video and audio data, researchers want to extract all vocalizations and identify which gerbil produced them (a problem known as speaker diarization). To achieve this, this topic plan to utilize deep convolutional networks and associated methods to quantify uncertainty in deep learning models.

Frank Zijun Zhang

1. Explainable graph neural network approach for cancer risk predictions
The aim of this topic is to make explainable predication of tumor grade classifications and risks for each tumor patient, leveraging personalized whole-genome scale somatic mutations. State-of-the-art deep learning approaches will be applied to model the coding and non-coding mutation effects, and graph neural networks will be used to integrate variant information across the genome. Finally, researchers will use model explaining techniques to distill biomedical insights that help tumor patient treatments.
2. Whole-genome genetic variation effect interrogation in autism
This topic will employ similar underlying computational strategies encompassing deep convolution neural networks and geometric deep learning and apply them to study autism. Researchers will leverage the latest in-house autism cohort collected by SFARI, the largest autism cohort in the world, and explore the genome-wide de novo mutation effects that influence a range of autism-related measures. These include levels and prognosis of autism, comorbidity and other relevant molecular and phenotypical traits.