MIT

One of the great challenges of neural networks is to understand how they work.  For example: does a neuron encode a meaningful signal on its own?  Or is a neuron simply an undistinguished and arbitrary component of a feature vector space?  The tension between the neuron doctrine and the population coding hypothesis is one of the classical debates in neuroscience. It is a difficult debate to settle without an ability to monitor every individual neuron in the brain.

Within artificial neural networks we can examine every neuron. Beginning with the simple proposal that an individual neuron might represent one internal concept, we conduct studies relating deep network neurons to human-understandable concepts in a concrete, quantitative way: Which neurons? Which concepts? Are neurons more meaningful than an arbitrary feature basis? Do neurons play a causal role? We examine both simplified settings and state-of-the-art networks in which neurons learn how to represent meaningful objects within the data without explicit supervision.

Following this inquiry in computer vision leads us to insights about the computational structure of practical deep networks that enable several new applications, including semantic manipulation of objects in an image; understanding of the sparse logic of a classifier; and quick, selective editing of generalizable rules within a fully trained generative network.  It also presents an unanswered mathematical question: why is such disentanglement so pervasive?

In the talk, we challenge the notion that the internal calculations of a neural network must be hopelessly opaque. Instead, we propose to tear back the curtain and chart a path through the detailed structure of a deep network by which we can begin to understand its logic.

One of the great challenges of neural networks is to understand how they work.  For example: does a neuron encode a meaningful signal on its own?  Or is a neuron simply an undistinguished and arbitrary component of a feature vector space?  The tension between the neuron doctrine and the population coding hypothesis is one of the classical debates in neuroscience. It is a difficult debate to settle without an ability to monitor every individual neuron in the brain.

Within artificial neural networks we can examine every neuron. Beginning with the simple proposal that an individual neuron might represent one internal concept, we conduct studies relating deep network neurons to human-understandable concepts in a concrete, quantitative way: Which neurons? Which concepts? Are neurons more meaningful than an arbitrary feature basis? Do neurons play a causal role? We examine both simplified settings and state-of-the-art networks in which neurons learn how to represent meaningful objects within the data without explicit supervision.

Following this inquiry in computer vision leads us to insights about the computational structure of practical deep networks that enable several new applications, including semantic manipulation of objects in an image; understanding of the sparse logic of a classifier; and quick, selective editing of generalizable rules within a fully trained generative network.  It also presents an unanswered mathematical question: why is such disentanglement so pervasive?

In the talk, we challenge the notion that the internal calculations of a neural network must be hopelessly opaque. Instead, we propose to tear back the curtain and chart a path through the detailed structure of a deep network by which we can begin to understand its logic.

--

A link for this talk will be sent to our mailing list a day or two in advance.  If you are not on the list and wish to be sent a link, please contact @email.

Elastic strings are among the simplest one-dimensional continuum bodies and have a rich mechanical and mathematical theory dating back to the derivation of their equations of motion by Euler and Lagrange. In classical treatments, the string is either completely extensible (tensile force produces elongation) or completely inextensible (every segment has a fixed length, regardless of the motion). However, common experience is that a string can be stretched (is extensible), and after a certain amount of tensile force is applied the stretch of the string is maximized (becomes inextensible). In this talk, we discuss a model for these stretch-limited elastic strings, in what way they model elastic behavior, the well-posedness and asymptotic stability of certain simple motions, and (many) open questions.

An insightful exercise might be to ask what is the most important idea in linear algebra. Our first answer would not be eigenvalues or linearity, it would be “matrix factorizations.”  We will discuss a blueprint to generate  53 inter-related matrix factorizations (times 2) most of which appear to be new. The underlying mathematics may be traced back to Cartan (1927), Harish-Chandra (1956), and Flensted-Jensen (1978) . We will discuss the interesting history. One anecdote is that Eugene Wigner (1968) discovered factorizations such as the svd in passing in a way that was buried and only eight authors have referenced that work. Ironically Wigner referenced Sigurður Helgason (1962) but Wigner did not recognize his results in Helgason's book. This work also extends upon and completes open problems posed by Mackey²&Tisseur (2003/2005).

Classical results of Random Matrix Theory concern exact formulas from the Hermite, Laguerre, Jacobi, and Circular distributions. Following an insight from Freeman Dyson (1970), Zirnbauer (1996) and Duenez (2004/5) linked some of these classical ensembles to Cartan's theory of Symmetric Spaces. One troubling fact is that symmetric spaces alone do not cover all of the Jacobi ensembles. We present a completed theory based on the generalized Cartan distribution. Furthermore, we show how the matrix factorization obtained by the generalized Cartan decomposition G=K₁AK₂ plays a crucial role in sampling algorithms and the derivation of the joint probability density of A.

Joint work with Sungwoo Jeong.

An insightful exercise might be to ask what is the most important idea in linear algebra. Our first answer would not be eigenvalues or linearity, it would be “matrix factorizations.” We will discuss a blueprint to generate 53 inter-related matrix factorizations (times 2) most of which appear to be new. The underlying mathematics may be traced back to Cartan (1927), Harish-Chandra (1956), and Flensted-Jensen (1978) . We will discuss the interesting history. One anecdote is that Eugene Wigner (1968) discovered factorizations such as the SVD in passing in a way that was buried and only eight authors have referenced that work. Ironically Wigner referenced Sigurður Helgason (1962) but Wigner did not recognize his results in Helgason's book. This work also extends upon and completes open problems posed by Mackey² & Tisseur (2003/2005).

Classical results of Random Matrix Theory concern exact formulas from the Hermite, Laguerre, Jacobi, and Circular distributions. Following an insight from Freeman Dyson (1970), Zirnbauer (1996) and Duenez (2004/5) linked some of these classical ensembles to Cartan's theory of Symmetric Spaces. One troubling fact is that symmetric spaces alone do not cover all of the Jacobi ensembles. We present a completed theory based on the generalized Cartan distribution. Furthermore, we show how the matrix factorization obtained by the generalized Cartan decomposition G=K₁AK₂ plays a crucial role in sampling algorithms and the derivation of the joint probability density of A.

Joint work with Sungwoo Jeong

--

A link for this talk will be sent to our mailing list a day or two in advance.  If you are not on the list and wish to be sent a link, please contact @email.

Recent developments on the black hole information paradox have shown that Euclidean wormholes — so called “replica wormholes’’  — can dominate the von Neumann entropy as computed by a gravitational path integral, and that inclusion of these wormholes results in a unitary Page curve. This development raises some puzzles from the perspective of factorization, and has raised questions regarding what the gravitational path integral is computing. In this talk, I will focus on understanding the relationship between the gravitational path integral and the partition function via the gravitational free energy (more generally the generating functional). A proper computation of the free energy requires a replica trick distinct from the usual one used to compute the entropy. I will show that in JT gravity there is a regime where the free energy computed without replica wormholes is pathological. Interestingly, the inclusion of replica wormholes is not quite sufficient to resolve the pathology: an alternative analytic continuation is required. I will discuss the implications of this for various interpretations of the gravitational path integral (e.g. as computing an ensemble average) and also mention some parallels with spin glasses. 

Scientific machine learning is a burgeoning discipline for mixing machine learning into scientific simulation. Use cases of this field include automated discovery of physical equations and accelerating physical simulators. However, making the analyses of this field automated will require building a set of tools that handle stiff and ill-conditioned models without requiring user tuning. The purpose of this talk is to demonstrate how the methods and tools of scientific machine learning can be consolidated to give a single high performance and robust software stack. We will start by describing universal differential equations, a flexible mathematical object which is able to represent methodologies for equation discovery, 100-dimensional differential equation solvers, and discretizations of physics-informed neural networks. Then we will showcase how adjoint sensitivity analysis on the universal differential equation solving process gives rise to efficient and stiffly robust training methodologies for a large variety of scientific machine learning problems. With this understanding of differentiable programming we will describe how the Julia SciML Software Organization is utilizing this foundation to provide high performance tools for deploying battery powered airplanes, improving the energy efficiency of buildings, allow for navigation via the Earth's magnetic field, and more.

A link for this talk will be sent to our mailing list a day or two in advance.  If you are not on the list and wish to be sent a link, please send email to @email.

In this talk, I will discuss some results on the structure of the cohomology of the moduli space of stable SL_n Higgs bundles on a curve. 

One consequence is a new proof of the Hausel-Thaddeus conjecture proven previously by Groechenig-Wyss-Ziegler via p-adic integration.

We will also discuss connections to the P=W conjecture if time permits. Based on joint work with Junliang Shen.

There is a very simple algorithm for the inference of posteriors for probability models on trees. This algorithm, known as "Belief Propagation" is widely used in coding theory, in machine learning, in evolutionary inference, among many other areas. The talk will be devoted to the analysis of Belief Propagation in some of the simplest probability models. We will highlight the interplay between Belief Propagation, linear estimators (statistics), the Kesten-Stigum bound (probability) and Replica Symmetry Breaking (statistical physics). We will show how the analysis of Belief Propagation allowed proof phase transitions for phylogenetic reconstruction in evolutionary biology and developed optimal algorithms for inference of block models. Finally, we will discuss the computational complexity of this 'simple' algorithm.