Colloquia and Seminars

Famously, multiset neural networks based on sum-pooling can separate all distinct multisets, and as a result can be used by message passing neural networks (MPNNs) to separate all pairs of graphs that can be separated by the 1-WL graph isomorphism test. 
However, the quality of this separation may be very weak, to the extent that the embeddings of "separable" multisets and graphs might even be considered identical when using fixed finite precision.

In this work, we propose to fully analyze the separation quality of multiset models and MPNNs via a novel adaptation of Lipschitz and Hölder continuity to parametric functions. 
We prove that common sum-based models are lower-Hölder continuous, with a Hölder exponent that decays rapidly with the network's depth. 
Our analysis leads to adversarial examples of graphs which can be separated by three 1-WL iterations, but cannot be separated in practice by standard maximally powerful MPNNs. 
To remedy this, we propose two novel MPNNs with improved separation quality, one of which is lower Lipschitz continuous. 
We show these MPNNs can easily classify our adversarial examples, and compare favorably with standard MPNNs on standard graph learning tasks.

The computations required for deep learning research have been doubling every few months, resulting in an estimated 5,000x increase from 2018 to 2022. This trend has led to unprecedented success in a range of AI tasks. In this talk I will discuss a few troubling side-effects of this trend, touching on issues of lack of inclusiveness within the research community, and an increasingly large environmental footprint. I will then present Green AI – an alternative approach to help mitigate these concerns. Green AI is composed of two main ideas: increased reporting of computational budgets, and making efficiency an evaluation criterion for research alongside accuracy and related measures. I will focus on the latter topic, discussing some of our recent efforts for reducing the costs of AI. This is joint work with Michael Hassid, Yossi Adi, Matanel Oren, Tal Remez, Jesse Dodge, Noah A. Smith, Oren Etzioni and Jonas Gehring. 

Bio: Roy Schwartz is a senior lecturer (assistant professor) at the School of Computer Science and Engineering at The Hebrew University of Jerusalem (HUJI). Roy studies natural language processing and artificial intelligence. Prior to joining HUJI, Roy was a postdoc (2016-2019) and then a research scientist (2019-2020) at the Allen institute for AI and at The University of Washington, where he worked with Noah A. Smith. Roy completed his Ph.D. in 2016 at HUJI, where he worked with Ari Rappoport. Roy’s work has appeared on the cover of the CACM magazine, and has been featured, among others, in the New York Times, Haaretz, and Ynet.

Modern programming languages rely on advanced and intricate type systems.
Expressive type systems support more language features but often result in unexpected language capabilities. For example, we know that the Java type system is Turing complete, which means it is so complex that Java compilers cannot guarantee termination. But a powerful type system can also be a blessing in disguise. Over the years, crafty programmers found ways to coerce type systems to perform basic computations at the type level. Certain program interfaces employ this form of metaprogramming to detect and report specialized program bugs before runtime.

In classical computer science, the notion of computational power is tightly coupled with automata (abstract machines) such as finite-state and Turing machines. We reason about the computability of various type systems by connecting them with well-founded classes of automata through type automata -- machines that employ program types as control and storage. By describing a bisimulation between type automata and classical automata, we precisely classify the expressiveness of a type system.

This work is comprised of a series of studies that analyze different type systems using type automata. We classify the computability of the decidable type systems fragments of Kennedy and Pierce in terms of regular and context-free tree languages. On the other hand, we prove that the Python type system defined in PEP 484 is Turing complete. In addition, we introduce novel metaprogramming techniques for an advanced interface design called fluent API. Our fluent APIs can enforce the API protocol or the grammar of an embedded domain-specific language (DSL) at compile time. We present the first fluent API design that supports all deterministic context-free API protocols and DSLs. We also demonstrate how to create elegant and sophisticated fluent APIs in functional programming languages.

Statistical enrichment tools are highly useful in biological research. Current approaches to statistical enrichment in ranked or ordered lists such as, for example, GSEA and GOrilla, are limited to the suffix (prefix) of the list. These methods assess extreme density of 1s in binary vectors on either side. Statistical significance can be assigned using, e.g, Wilcoxon Rank Sum and mHG statistics.
In this work we extend the mHG approach to also address enrichment in any index intervals of the binary vector. We define and provide a partial characterization of related distributions under a uniform null model. Our partial characterization yields useful bounds for extreme events. We provide a software tool to the community, implementing the method in Python. Finally, we analyze several example use cases and describe the results. We show, for example, that lung cancer differential expression, comparing ADC to other types, is enriched in a region of Chromosome 3. This example represents a typical use case for imHG -- obtaining enriched intervals for any set of genes of interest. We provide a Python implementation, called imHG, for finding and reporting enriched genomic intervals with any given list of genes of interest. 

Diagnosing Crohn's Disease (CD) typically involves examining 2D slices from magnetic resonance enterography (MRE). However, the anisotropic resolution of MRE complicates precise 3D measurements and visualization. The absence of automated 3D measurement systems further complicates assessment. Previous methods for generating isotropic volumes from anisotropic data often rely on extensive 3D data and focus solely on interslice resolution, leading to suboptimal outcomes due to data scarcity and inaccuracies. We propose a self-supervised multi-plane generative model using Generative Adversarial Network (GAN) architecture, incorporating multiple discriminators for different planes. We introduce a semi-automatic algorithm to predict the centerline of the terminal ileum, which is the part of the body primarily affected by CD, enhancing efficiency in 3D MRE analysis. Evaluation using 115 2D abdominal MRE datasets from Rambam Health Care Campus underscores the potential of our approach to enhance diagnostic accuracy and visualization in CD. Our semi-automatic centerline prediction reduces the radiologist's analysis time significantly. Additionally, our isotropic volume generation model can be expanded to other anatomical regions, thereby reducing MRI scan-time and costs. 

Traditional discriminative Machine-Learning approaches often formulate problems as optimization of a parameterized function mapping from a given input space to some desired output space.
While this formulation is applicable to many theoretical and practical problems, it is inherently reliant on the assumption that the dataset X is constant. The world around us is abundant with scenarios where this is either not the case, or where dropping this assumption could help achieve better solutions.
Namely, in many scenarios one can gain added benefit from starting from one step before the data is acquired - if the way data is perceived by our sensors can be modeled, with a correct model it can be parameterized. If it can be parameterized, with the right mathematical tools (e.g. Deep Learning), it can be optimized to gives us datasets more useful for achieving the downstream task at
hand. As we are solving task-specific different optimization of the acquired data, a joint optimization of the downstream task model and data acquisition model is usually called for. This type of joint optimization calls for the delicate balancing in optimization of the two systems (acquisition and downstream), bringing upon a rich and interesting range of problems, coming from a diverse set of fields in science and engineering.
In our research we identify and study a collection of such problems, and offer modeling and optimization schemes to solve them.