-
Semiclassical analysis on principal bundles
Authors:
Mihajlo Cekić,
Thibault Lefeuvre
Abstract:
Let $G$ be a compact Lie group. We introduce a semiclassical framework, called Borel-Weil calculus, to investigate $G$-equivariant (pseudo)differential operators acting on $G$-principal bundles over closed manifolds. In this calculus, the semiclassical parameters correspond to the highest roots in the Weyl chamber of the group $G$ that parametrize irreducible representations, and operators are pse…
▽ More
Let $G$ be a compact Lie group. We introduce a semiclassical framework, called Borel-Weil calculus, to investigate $G$-equivariant (pseudo)differential operators acting on $G$-principal bundles over closed manifolds. In this calculus, the semiclassical parameters correspond to the highest roots in the Weyl chamber of the group $G$ that parametrize irreducible representations, and operators are pseudodifferential in the base variable, with values in Toeplitz operators on the flag manifold associated to the group. This monograph unfolds two main applications of our calculus.
Firstly, in the realm of dynamical systems, we obtain explicit sufficient conditions for rapid mixing of volume-preserving partially hyperbolic flows obtained as extensions of an Anosov flow to a $G$-principal bundle (for an arbitrary $G$). In particular, when $G = \mathrm{U}(1)$, we prove that the flow on the extension is rapid mixing whenever the Anosov flow is not jointly integrable, and the circle bundle is not torsion. When $G$ is semisimple, we prove that ergodicity of the extension is equivalent to rapid mixing. Secondly, we study the spectral theory of sub-elliptic Laplacians obtained as horizontal Laplacians of a $G$-equivariant connection on a principal bundle. When $G$ is semisimple, we prove that the horizontal Laplacian is globally hypoelliptic as soon as the connection has a dense holonomy group in $G$. Notably, this result encompasses all flat bundles with a dense monodromy group in $G$. We also prove a quantum ergodicity result for flat principal bundles with dense holonomy group if the base space has Anosov geodesic flow (e.g. in negative sectional curvature).
We believe that this monograph will serve as a cornerstone for future investigations applying the Borel-Weil calculus across different fields.
△ Less
Submitted 23 May, 2024;
originally announced May 2024.
-
PEAVS: Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers' Opinion Scores
Authors:
Lucas Goncalves,
Prashant Mathur,
Chandrashekhar Lavania,
Metehan Cekic,
Marcello Federico,
Kyu J. Han
Abstract:
Recent advancements in audio-visual generative modeling have been propelled by progress in deep learning and the availability of data-rich benchmarks. However, the growth is not attributed solely to models and benchmarks. Universally accepted evaluation metrics also play an important role in advancing the field. While there are many metrics available to evaluate audio and visual content separately…
▽ More
Recent advancements in audio-visual generative modeling have been propelled by progress in deep learning and the availability of data-rich benchmarks. However, the growth is not attributed solely to models and benchmarks. Universally accepted evaluation metrics also play an important role in advancing the field. While there are many metrics available to evaluate audio and visual content separately, there is a lack of metrics that offer a quantitative and interpretable measure of audio-visual synchronization for videos "in the wild". To address this gap, we first created a large scale human annotated dataset (100+ hrs) representing nine types of synchronization errors in audio-visual content and how human perceive them. We then developed a PEAVS (Perceptual Evaluation of Audio-Visual Synchrony) score, a novel automatic metric with a 5-point scale that evaluates the quality of audio-visual synchronization. We validate PEAVS using a newly generated dataset, achieving a Pearson correlation of 0.79 at the set level and 0.54 at the clip level when compared to human labels. In our experiments, we observe a relative gain 50% over a natural extension of Fréchet based metrics for Audio-Visual synchrony, confirming PEAVS efficacy in objectively modeling subjective perceptions of audio-visual synchronization for videos "in the wild".
△ Less
Submitted 10 April, 2024;
originally announced April 2024.
-
Calderón problem for systems via complex parallel transport
Authors:
Mihajlo Cekić
Abstract:
We consider the Calderón problem for systems with unknown zeroth and first order terms, and improve on previously known results. More precisely, let $(M, g)$ be a compact Riemannian manifold with boundary, let $A$ be a connection matrix on $E = M \times \mathbb{C}^r$ and let $Q$ be a matrix potential. Let $Λ_{A, Q}$ be the Dirichlet-to-Neumann map of the associated connection Laplacian with a pote…
▽ More
We consider the Calderón problem for systems with unknown zeroth and first order terms, and improve on previously known results. More precisely, let $(M, g)$ be a compact Riemannian manifold with boundary, let $A$ be a connection matrix on $E = M \times \mathbb{C}^r$ and let $Q$ be a matrix potential. Let $Λ_{A, Q}$ be the Dirichlet-to-Neumann map of the associated connection Laplacian with a potential. Under the assumption that $(M, g)$ is isometrically contained in the interior of $(\mathbb{R}^2 \times M_0, c(e \oplus g_0))$, where $(M_0, g_0)$ is an arbitrary compact Riemannian manifold with boundary, $e$ is the Euclidean metric on $\mathbb{R}^2$, and $c > 0$, we show that $Λ_{A, Q}$ uniquely determines $(A, Q)$ up to natural gauge invariances. This generalises earlier work of Eskin to anisotropic geometries. Moreover, we introduce new concepts of complex ray transform and complex parallel transport problem, and study their fundamental properties and relations to the Calderón problem.
△ Less
Submitted 2 June, 2024; v1 submitted 17 September, 2023;
originally announced September 2023.
-
Correspondence between Pestov and Weitzenböck identities
Authors:
Mihajlo Cekić,
Thibault Lefeuvre,
Andrei Moroianu,
Uwe Semmelmann
Abstract:
The aim of this note is to establish the correspondence between the twisted localized Pestov identity on the unit tangent bundle of a Riemannian manifold and the Weitzenböck identity for twisted symmetric tensors on the manifold.
The aim of this note is to establish the correspondence between the twisted localized Pestov identity on the unit tangent bundle of a Riemannian manifold and the Weitzenböck identity for twisted symmetric tensors on the manifold.
△ Less
Submitted 9 May, 2023; v1 submitted 1 May, 2023;
originally announced May 2023.
-
Stability estimates for the Holonomy Inverse Problem
Authors:
Mihajlo Cekić,
Thibault Lefeuvre
Abstract:
On a Riemannian manifold $(M, g)$ with Anosov geodesic flow, the problem of recovering a connection from the knowledge of traces of its holonomies along primitive closed geodesics is known as the holonomy inverse problem. In this paper, we prove Hölder type stability estimates for this inverse problem: 1) locally, near generic connections; 2) globally, for line bundles, and for vector bundles sati…
▽ More
On a Riemannian manifold $(M, g)$ with Anosov geodesic flow, the problem of recovering a connection from the knowledge of traces of its holonomies along primitive closed geodesics is known as the holonomy inverse problem. In this paper, we prove Hölder type stability estimates for this inverse problem: 1) locally, near generic connections; 2) globally, for line bundles, and for vector bundles satisfying a certain low-rank assumption over negatively curved base $(M, g)$. The proofs are based on a combination of microlocal analysis along with a new non-Abelian approximate Livsic Theorem in hyperbolic dynamics.
△ Less
Submitted 22 April, 2024; v1 submitted 21 March, 2023;
originally announced March 2023.
-
On the ergodicity of unitary frame flows on Kähler manifolds
Authors:
Mihajlo Cekić,
Thibault Lefeuvre,
Andrei Moroianu,
Uwe Semmelmann
Abstract:
Let $(M,g,J)$ be a closed Kähler manifold with negative sectional curvature and complex dimension $m := \dim_{\mathbb{C}} M \geq 2$. In this article, we study the unitary frame flow, that is, the restriction of the frame flow to the principal $\mathrm{U}(m)$-bundle $F_{\mathbb{C}}M$ of unitary frames. We show that if $m \geq 6$ is even, and $m \neq 28$, there exists $λ(m) \in (0, 1)$ such that if…
▽ More
Let $(M,g,J)$ be a closed Kähler manifold with negative sectional curvature and complex dimension $m := \dim_{\mathbb{C}} M \geq 2$. In this article, we study the unitary frame flow, that is, the restriction of the frame flow to the principal $\mathrm{U}(m)$-bundle $F_{\mathbb{C}}M$ of unitary frames. We show that if $m \geq 6$ is even, and $m \neq 28$, there exists $λ(m) \in (0, 1)$ such that if $(M, g, J)$ has negative $λ(m)$-pinched holomorphic sectional curvature, then the unitary frame flow is ergodic and mixing. The constants $λ(m)$ satisfy $λ(6) = 0.9330...$, $\lim_{m \to +\infty} λ(m) = \tfrac{11}{12} = 0.9166...$, and $m \mapsto λ(m)$ is decreasing. This extends to the even-dimensional case the results of Brin-Gromov who proved ergodicity of the unitary frame flow on negatively-curved compact Kähler manifolds of odd complex dimension.
△ Less
Submitted 14 January, 2023;
originally announced January 2023.
-
Resonant forms at zero for dissipative Anosov flows
Authors:
Mihajlo Cekić,
Gabriel P. Paternain
Abstract:
We study resonant differential forms at zero for transitive Anosov flows on $3$-manifolds. We pay particular attention to the dissipative case, that is, Anosov flows that do not preserve an absolutely continuous measure. Such flows have two distinguished Sinai-Ruelle-Bowen $3$-forms, $Ω_{\text{SRB}}^{\pm}$, and the cohomology classes $[ι_{X}Ω_{\text{SRB}}^{\pm}]$ (where $X$ is the infinitesimal ge…
▽ More
We study resonant differential forms at zero for transitive Anosov flows on $3$-manifolds. We pay particular attention to the dissipative case, that is, Anosov flows that do not preserve an absolutely continuous measure. Such flows have two distinguished Sinai-Ruelle-Bowen $3$-forms, $Ω_{\text{SRB}}^{\pm}$, and the cohomology classes $[ι_{X}Ω_{\text{SRB}}^{\pm}]$ (where $X$ is the infinitesimal generator of the flow) play a key role in the determination of the space of resonant $1$-forms. When both classes vanish we associate to the flow a $\textit{helicity}$ that naturally extends the classical notion associated with null-homologous volume preserving flows. We provide a general theory that includes horocyclic invariance of resonant $1$-forms and SRB-measures as well as the local geometry of the maps $X\mapsto [ι_{X}Ω_{\text{SRB}}^{\pm}]$ near a null-homologous volume preserving flow. Next, we study several relevant classes of examples. Among these are thermostats associated with holomorphic quadratic differentials, giving rise to quasi-Fuchsian flows as introduced by Ghys. For these flows we compute explicitly all resonant $1$-forms at zero, we show that $[ι_{X}Ω_{\text{SRB}}^{\pm}]=0$ and give an explicit formula for the helicity. In addition we show that a generic time change of a quasi-Fuchsian flow is semisimple and thus the order of vanishing of the Ruelle zeta function at zero is $-χ(M)$, the same as in the geodesic flow case. In contrast, we show that if $(M,g)$ is a closed surface of negative curvature, the Gaussian thermostat driven by a (small) harmonic $1$-form has a Ruelle zeta function whose order of vanishing at zero is $-χ(M)-1$.
△ Less
Submitted 11 November, 2022;
originally announced November 2022.
-
Isospectral connections, ergodicity of frame flows, and polynomial maps between spheres
Authors:
Mihajlo Cekić,
Thibault Lefeuvre
Abstract:
We show that on closed negatively curved Riemannian manifolds with simple length spectrum, the spectrum of the Bochner Laplacian determines both the isomorphism class of the vector bundle and the connection up to gauge under a low-rank assumption. We also show that flows of frames on low-rank frame bundles extending the geodesic flow in negative curvature are ergodic whenever the bundle admits no…
▽ More
We show that on closed negatively curved Riemannian manifolds with simple length spectrum, the spectrum of the Bochner Laplacian determines both the isomorphism class of the vector bundle and the connection up to gauge under a low-rank assumption. We also show that flows of frames on low-rank frame bundles extending the geodesic flow in negative curvature are ergodic whenever the bundle admits no holonomy reduction. This is achieved by exhibiting a link between these problems and the classification of polynomial maps between spheres in real algebraic geometry.
△ Less
Submitted 1 December, 2023; v1 submitted 22 September, 2022;
originally announced September 2022.
-
Towards Brin's conjecture on frame flow ergodicity: new progress and perspectives
Authors:
Mihajlo Cekić,
Thibault Lefeuvre,
Andrei Moroianu,
Uwe Semmelmann
Abstract:
We report on some recent progress achieved in [arXiv:2111.14811] on the ergodicity of the frame flow of negatively-curved Riemannian manifolds. We explain the new ideas leading to ergodicity for nearly $0.25$-pinched manifolds and give perspectives for future work.
We report on some recent progress achieved in [arXiv:2111.14811] on the ergodicity of the frame flow of negatively-curved Riemannian manifolds. We explain the new ideas leading to ergodicity for nearly $0.25$-pinched manifolds and give perspectives for future work.
△ Less
Submitted 13 May, 2022; v1 submitted 19 April, 2022;
originally announced April 2022.
-
Local lens rigidity for manifolds of Anosov type
Authors:
Mihajlo Cekić,
Colin Guillarmou,
Thibault Lefeuvre
Abstract:
The lens data of a Riemannian manifold with boundary is the collection of lengths of geodesics with endpoints on the boundary together with their incoming and outgoing vectors. We show that negatively-curved Riemannian manifolds with strictly convex boundary are locally lens rigid in the following sense: if $g_0$ is such a metric, then any metric $g$ sufficiently close to $g_0$ and with same lens…
▽ More
The lens data of a Riemannian manifold with boundary is the collection of lengths of geodesics with endpoints on the boundary together with their incoming and outgoing vectors. We show that negatively-curved Riemannian manifolds with strictly convex boundary are locally lens rigid in the following sense: if $g_0$ is such a metric, then any metric $g$ sufficiently close to $g_0$ and with same lens data is isometric to $g_0$, up to a boundary-preserving diffeomorphism. More generally, we consider the same problem for a wider class of metrics with strictly convex boundary, called metrics of Anosov type. We prove that the same rigidity result holds within that class in dimension $2$ and in any dimension, further assuming that the curvature is non-positive.
△ Less
Submitted 21 July, 2023; v1 submitted 5 April, 2022;
originally announced April 2022.
-
Neuro-Inspired Deep Neural Networks with Sparse, Strong Activations
Authors:
Metehan Cekic,
Can Bakiskan,
Upamanyu Madhow
Abstract:
While end-to-end training of Deep Neural Networks (DNNs) yields state of the art performance in an increasing array of applications, it does not provide insight into, or control over, the features being extracted. We report here on a promising neuro-inspired approach to DNNs with sparser and stronger activations. We use standard stochastic gradient training, supplementing the end-to-end discrimina…
▽ More
While end-to-end training of Deep Neural Networks (DNNs) yields state of the art performance in an increasing array of applications, it does not provide insight into, or control over, the features being extracted. We report here on a promising neuro-inspired approach to DNNs with sparser and stronger activations. We use standard stochastic gradient training, supplementing the end-to-end discriminative cost function with layer-wise costs promoting Hebbian ("fire together," "wire together") updates for highly active neurons, and anti-Hebbian updates for the remaining neurons. Instead of batch norm, we use divisive normalization of activations (suppressing weak outputs using strong outputs), along with implicit $\ell_2$ normalization of neuronal weights. Experiments with standard image classification tasks on CIFAR-10 demonstrate that, relative to baseline end-to-end trained architectures, our proposed architecture (a) leads to sparser activations (with only a slight compromise on accuracy), (b) exhibits more robustness to noise (without being trained on noisy data), (c) exhibits more robustness to adversarial perturbations (without adversarial training).
△ Less
Submitted 8 July, 2022; v1 submitted 26 February, 2022;
originally announced February 2022.
-
Self-supervised Speaker Recognition Training Using Human-Machine Dialogues
Authors:
Metehan Cekic,
Ruirui Li,
Zeya Chen,
Yuguang Yang,
Andreas Stolcke,
Upamanyu Madhow
Abstract:
Speaker recognition, recognizing speaker identities based on voice alone, enables important downstream applications, such as personalization and authentication. Learning speaker representations, in the context of supervised learning, heavily depends on both clean and sufficient labeled data, which is always difficult to acquire. Noisy unlabeled data, on the other hand, also provides valuable infor…
▽ More
Speaker recognition, recognizing speaker identities based on voice alone, enables important downstream applications, such as personalization and authentication. Learning speaker representations, in the context of supervised learning, heavily depends on both clean and sufficient labeled data, which is always difficult to acquire. Noisy unlabeled data, on the other hand, also provides valuable information that can be exploited using self-supervised training methods. In this work, we investigate how to pretrain speaker recognition models by leveraging dialogues between customers and smart-speaker devices. However, the supervisory information in such dialogues is inherently noisy, as multiple speakers may speak to a device in the course of the same dialogue. To address this issue, we propose an effective rejection mechanism that selectively learns from dialogues based on their acoustic homogeneity. Both reconstruction-based and contrastive-learning-based self-supervised methods are compared. Experiments demonstrate that the proposed method provides significant performance improvements, superior to earlier work. Dialogue pretraining when combined with the rejection mechanism yields 27.10% equal error rate (EER) reduction in speaker recognition, compared to a model without self-supervised pretraining.
△ Less
Submitted 17 February, 2022; v1 submitted 7 February, 2022;
originally announced February 2022.
-
On the ergodicity of the frame flow on even-dimensional manifolds
Authors:
Mihajlo Cekić,
Thibault Lefeuvre,
Andrei Moroianu,
Uwe Semmelmann
Abstract:
It is known that the frame flow on a closed $n$-dimensional Riemannian manifold with negative sectional curvature is ergodic if $n$ is odd and $n \neq 7$. In this paper we study its ergodicity for $n \geq 4$ even and $n = 7$, and we show that: if $n \equiv 2$ mod $4$, or $n=4$, the frame flow is ergodic if the manifold is $\sim 0.3$-pinched, if $n \equiv 0$ mod $4$, it is ergodic if the manifold i…
▽ More
It is known that the frame flow on a closed $n$-dimensional Riemannian manifold with negative sectional curvature is ergodic if $n$ is odd and $n \neq 7$. In this paper we study its ergodicity for $n \geq 4$ even and $n = 7$, and we show that: if $n \equiv 2$ mod $4$, or $n=4$, the frame flow is ergodic if the manifold is $\sim 0.3$-pinched, if $n \equiv 0$ mod $4$, it is ergodic if the manifold is $\sim 0.6$-pinched, except in the three dimensions $n=7,8,134$, where the respective pinching conditions are $0.4962...$, $0.6212...$, and $0.5788...$. In particular, if $n = 4$ or $n \equiv 2$ mod $4$, this almost solves a long-standing conjecture of Brin asserting that $1/4$-pinched even-dimensional manifolds have an ergodic frame flow.
△ Less
Submitted 5 February, 2023; v1 submitted 29 November, 2021;
originally announced November 2021.
-
Generic injectivity of the X-ray transform
Authors:
Mihajlo Cekić,
Thibault Lefeuvre
Abstract:
In dimensions $\geq 3$, we prove that the X-ray transform of symmetric tensors of arbitrary degree is generically injective with respect to the metric on closed Anosov manifolds and on manifolds with spherical strictly convex boundary, no conjugate points and a hyperbolic trapped set. This has two immediate corollaries: local spectral rigidity, and local marked length spectrum rigidity (building o…
▽ More
In dimensions $\geq 3$, we prove that the X-ray transform of symmetric tensors of arbitrary degree is generically injective with respect to the metric on closed Anosov manifolds and on manifolds with spherical strictly convex boundary, no conjugate points and a hyperbolic trapped set. This has two immediate corollaries: local spectral rigidity, and local marked length spectrum rigidity (building on earlier work by Guillarmou, Knieper and the second author [arXiv:1806.04218], [arXiv:1909.08666]), in a neighbourhood of a generic Anosov metric. In both cases, this is the first work going beyond the negatively curved assumption or dimension $2$.
Our method, initiated in [arXiv:2008.09191] and fully developed in the present paper, is based on a perturbative argument of the $0$-eigenvalue of elliptic operators via microlocal analysis which turn the analytic problem of injectivity into an algebraic problem of representation theory. When the manifold is equipped with a Hermitian vector bundle together with a unitary connection, we also show that the twisted X-ray transform of symmetric tensors (with values in that bundle) is generically injective with respect to the connection. This property turns out to be crucial when solving the $\textit{holonomy inverse problem}$, as studied in a subsequent article [arXiv:2105.06376].
△ Less
Submitted 24 January, 2024; v1 submitted 11 July, 2021;
originally announced July 2021.
-
The RSNA-ASNR-MICCAI BraTS 2021 Benchmark on Brain Tumor Segmentation and Radiogenomic Classification
Authors:
Ujjwal Baid,
Satyam Ghodasara,
Suyash Mohan,
Michel Bilello,
Evan Calabrese,
Errol Colak,
Keyvan Farahani,
Jayashree Kalpathy-Cramer,
Felipe C. Kitamura,
Sarthak Pati,
Luciano M. Prevedello,
Jeffrey D. Rudie,
Chiharu Sako,
Russell T. Shinohara,
Timothy Bergquist,
Rong Chai,
James Eddy,
Julia Elliott,
Walter Reade,
Thomas Schaffter,
Thomas Yu,
Jiaxin Zheng,
Ahmed W. Moawad,
Luiz Otavio Coelho,
Olivia McDonnell
, et al. (78 additional authors not shown)
Abstract:
The BraTS 2021 challenge celebrates its 10th anniversary and is jointly organized by the Radiological Society of North America (RSNA), the American Society of Neuroradiology (ASNR), and the Medical Image Computing and Computer Assisted Interventions (MICCAI) society. Since its inception, BraTS has been focusing on being a common benchmarking venue for brain glioma segmentation algorithms, with wel…
▽ More
The BraTS 2021 challenge celebrates its 10th anniversary and is jointly organized by the Radiological Society of North America (RSNA), the American Society of Neuroradiology (ASNR), and the Medical Image Computing and Computer Assisted Interventions (MICCAI) society. Since its inception, BraTS has been focusing on being a common benchmarking venue for brain glioma segmentation algorithms, with well-curated multi-institutional multi-parametric magnetic resonance imaging (mpMRI) data. Gliomas are the most common primary malignancies of the central nervous system, with varying degrees of aggressiveness and prognosis. The RSNA-ASNR-MICCAI BraTS 2021 challenge targets the evaluation of computational algorithms assessing the same tumor compartmentalization, as well as the underlying tumor's molecular characterization, in pre-operative baseline mpMRI data from 2,040 patients. Specifically, the two tasks that BraTS 2021 focuses on are: a) the segmentation of the histologically distinct brain tumor sub-regions, and b) the classification of the tumor's O[6]-methylguanine-DNA methyltransferase (MGMT) promoter methylation status. The performance evaluation of all participating algorithms in BraTS 2021 will be conducted through the Sage Bionetworks Synapse platform (Task 1) and Kaggle (Task 2), concluding in distributing to the top ranked participants monetary awards of $60,000 collectively.
△ Less
Submitted 12 September, 2021; v1 submitted 5 July, 2021;
originally announced July 2021.
-
The Holonomy Inverse Problem
Authors:
Mihajlo Cekić,
Thibault Lefeuvre
Abstract:
Let $(M,g)$ be a smooth Anosov Riemannian manifold and $\mathcal{C}^\sharp$ the set of its primitive closed geodesics. Given a Hermitian vector bundle $\mathcal{E}$ equipped with a unitary connection $\nabla^{\mathcal{E}}$, we define $\mathcal{T}^\sharp(\mathcal{E}, \nabla^{\mathcal{E}})$ as the sequence of traces of holonomies of $\nabla^{\mathcal{E}}$ along elements of $\mathcal{C}^\sharp$. This…
▽ More
Let $(M,g)$ be a smooth Anosov Riemannian manifold and $\mathcal{C}^\sharp$ the set of its primitive closed geodesics. Given a Hermitian vector bundle $\mathcal{E}$ equipped with a unitary connection $\nabla^{\mathcal{E}}$, we define $\mathcal{T}^\sharp(\mathcal{E}, \nabla^{\mathcal{E}})$ as the sequence of traces of holonomies of $\nabla^{\mathcal{E}}$ along elements of $\mathcal{C}^\sharp$. This descends to a homomorphism on the additive moduli space $\mathbb{A}$ of connections up to gauge $\mathcal{T}^\sharp: (\mathbb{A}, \oplus) \to \ell^\infty(\mathcal{C}^\sharp)$, which we call the $\textit{primitive trace map}$. It is the restriction of the well-known $\textit{Wilson loop}$ operator to primitive closed geodesics.
The main theorem of this paper shows that the primitive trace map $\mathcal{T}^\sharp$ is locally injective near generic points of $\mathbb{A}$ when $\dim(M) \geq 3$. We obtain global results in some particular cases: flat bundles, direct sums of line bundles, and general bundles in negative curvature under a spectral assumption which is satisfied in particular for connections with small curvature. As a consequence of the main theorem, we also derive a spectral rigidity result for the connection Laplacian.
The proofs are based on two new ingredients: a Livšic-type theorem in hyperbolic dynamical systems showing that the cohomology class of a unitary cocycle is determined by its trace along closed primitive orbits, and a theorem relating the local geometry of $\mathbb{A}$ with the Pollicott-Ruelle resonance near zero of a certain natural transport operator.
△ Less
Submitted 22 December, 2023; v1 submitted 13 May, 2021;
originally announced May 2021.
-
Sparse Coding Frontend for Robust Neural Networks
Authors:
Can Bakiskan,
Metehan Cekic,
Ahmet Dundar Sezer,
Upamanyu Madhow
Abstract:
Deep Neural Networks are known to be vulnerable to small, adversarially crafted, perturbations. The current most effective defense methods against these adversarial attacks are variants of adversarial training. In this paper, we introduce a radically different defense trained only on clean images: a sparse coding based frontend which significantly attenuates adversarial attacks before they reach t…
▽ More
Deep Neural Networks are known to be vulnerable to small, adversarially crafted, perturbations. The current most effective defense methods against these adversarial attacks are variants of adversarial training. In this paper, we introduce a radically different defense trained only on clean images: a sparse coding based frontend which significantly attenuates adversarial attacks before they reach the classifier. We evaluate our defense on CIFAR-10 dataset under a wide range of attack types (including Linf , L2, and L1 bounded attacks), demonstrating its promise as a general-purpose approach for defense.
△ Less
Submitted 12 April, 2021;
originally announced April 2021.
-
A Neuro-Inspired Autoencoding Defense Against Adversarial Perturbations
Authors:
Can Bakiskan,
Metehan Cekic,
Ahmet Dundar Sezer,
Upamanyu Madhow
Abstract:
Deep Neural Networks (DNNs) are vulnerable to adversarial attacks: carefully constructed perturbations to an image can seriously impair classification accuracy, while being imperceptible to humans. While there has been a significant amount of research on defending against such attacks, most defenses based on systematic design principles have been defeated by appropriately modified attacks. For a f…
▽ More
Deep Neural Networks (DNNs) are vulnerable to adversarial attacks: carefully constructed perturbations to an image can seriously impair classification accuracy, while being imperceptible to humans. While there has been a significant amount of research on defending against such attacks, most defenses based on systematic design principles have been defeated by appropriately modified attacks. For a fixed set of data, the most effective current defense is to train the network using adversarially perturbed examples. In this paper, we investigate a radically different, neuro-inspired defense mechanism, starting from the observation that human vision is virtually unaffected by adversarial examples designed for machines. We aim to reject L^inf bounded adversarial perturbations before they reach a classifier DNN, using an encoder with characteristics commonly observed in biological vision: sparse overcomplete representations, randomness due to synaptic noise, and drastic nonlinearities. Encoder training is unsupervised, using standard dictionary learning. A CNN-based decoder restores the size of the encoder output to that of the original image, enabling the use of a standard CNN for classification. Our nominal design is to train the decoder and classifier together in standard supervised fashion, but we also consider unsupervised decoder training based on a regression objective (as in a conventional autoencoder) with separate supervised training of the classifier. Unlike adversarial training, all training is based on clean images.
Our experiments on the CIFAR-10 show performance competitive with state-of-the-art defenses based on adversarial training, and point to the promise of neuro-inspired techniques for the design of robust neural networks. In addition, we provide results for a subset of the Imagenet dataset to verify that our approach scales to larger images.
△ Less
Submitted 21 December, 2020; v1 submitted 21 November, 2020;
originally announced November 2020.
-
First band of Ruelle resonances for contact Anosov flows in dimension $3$
Authors:
Colin Guillarmou,
Mihajlo Cekic
Abstract:
We show, using semiclassical measures and unstable derivatives, that a smooth vector field $X$ generating a contact Anosov flow on a $3$-dimensional manifold $\mathcal{M}$ has only finitely many Ruelle resonances in the vertical strips $\{ s\in \mathbb{C}\ |\ {\rm Re}(s)\in [-ν_{\min}+ε,-\frac{1}{2}ν_{\max}-ε]\cup [-\frac{1}{2}ν_{\min}+ε,0]\}$ for all $ε>0$, where $0<ν_{\min}\leq ν_{\max}$ are the…
▽ More
We show, using semiclassical measures and unstable derivatives, that a smooth vector field $X$ generating a contact Anosov flow on a $3$-dimensional manifold $\mathcal{M}$ has only finitely many Ruelle resonances in the vertical strips $\{ s\in \mathbb{C}\ |\ {\rm Re}(s)\in [-ν_{\min}+ε,-\frac{1}{2}ν_{\max}-ε]\cup [-\frac{1}{2}ν_{\min}+ε,0]\}$ for all $ε>0$, where $0<ν_{\min}\leq ν_{\max}$ are the minimal and maximal expansion rates of the flow (the first strip only makes sense if $ν_{\min}>ν_{\max}/2$). We also show polynomial bounds in $s$ for the resolvent $(-X-s)^{-1}$ as $|{\rm Im}(s)|\to \infty$ in Sobolev spaces, and obtain similar results for cases with a potential. This is a short proof of a particular case of the results by Faure-Tsujii in \cite{FaTs1,FaTs2,FaTs3}, using that $\dim E_u=\dim E_s=1$.
△ Less
Submitted 21 April, 2021; v1 submitted 11 November, 2020;
originally announced November 2020.
-
The Ruelle zeta function at zero for nearly hyperbolic 3-manifolds
Authors:
Mihajlo Cekić,
Benjamin Delarue,
Semyon Dyatlov,
Gabriel P. Paternain
Abstract:
We show that for a generic conformal metric perturbation of a compact hyperbolic 3-manifold $Σ$ with Betti number $b_1$, the order of vanishing of the Ruelle zeta function at zero equals $4-b_1$, while in the hyperbolic case it is equal to $4-2b_1$. This is in contrast to the 2-dimensional case where the order of vanishing is a topological invariant. The proof uses the microlocal approach to dynam…
▽ More
We show that for a generic conformal metric perturbation of a compact hyperbolic 3-manifold $Σ$ with Betti number $b_1$, the order of vanishing of the Ruelle zeta function at zero equals $4-b_1$, while in the hyperbolic case it is equal to $4-2b_1$. This is in contrast to the 2-dimensional case where the order of vanishing is a topological invariant. The proof uses the microlocal approach to dynamical zeta functions, giving a geometric description of generalized Pollicott-Ruelle resonant differential forms at 0 in the hyperbolic case and using first variation for the perturbation. To show that the first variation is generically nonzero we introduce a new identity relating pushforwards of products of resonant and coresonant 2-forms on the sphere bundle $SΣ$ with harmonic 1-forms on $Σ$.
△ Less
Submitted 8 February, 2022; v1 submitted 17 September, 2020;
originally announced September 2020.
-
Generic dynamical properties of connections on vector bundles
Authors:
Mihajlo Cekić,
Thibault Lefeuvre
Abstract:
Given a smooth Hermitian vector bundle $\mathcal{E}$ over a closed Riemannian manifold $(M,g)$, we study generic properties of unitary connections $\nabla^{\mathcal{E}}$ on the vector bundle $\mathcal{E}$. First of all, we show that twisted Conformal Killing Tensors (CKTs) are generically trivial when $\dim(M) \geq 3$, answering an open question of Guillarmou-Paternain-Salo-Uhlmann. In negative cu…
▽ More
Given a smooth Hermitian vector bundle $\mathcal{E}$ over a closed Riemannian manifold $(M,g)$, we study generic properties of unitary connections $\nabla^{\mathcal{E}}$ on the vector bundle $\mathcal{E}$. First of all, we show that twisted Conformal Killing Tensors (CKTs) are generically trivial when $\dim(M) \geq 3$, answering an open question of Guillarmou-Paternain-Salo-Uhlmann. In negative curvature, it is known that the existence of twisted CKTs is the only obstruction to solving exactly the twisted cohomological equations which may appear in various geometric problems such as the study of transparent connections. The main result of this paper says that these equations can be generically solved. As a by-product, we also obtain that the induced connection $\nabla^{\mathrm{End}(\mathcal{E})}$ on the endomorphism bundle $\mathrm{End}(\mathcal{E})$ has generically trivial CKTs as long as $(M,g)$ has no nontrivial CKTs on its trivial line bundle. Eventually, we show that, under the additional assumption that $(M,g)$ is Anosov (i.e. the geodesic flow is Anosov on the unit tangent bundle), the connections are generically $\textit{opaque}$, namely there are no non-trivial subbundles of $\mathcal{E}$ which are generically preserved by parallel transport along geodesics. The proofs rely on the introduction of a new microlocal property for (pseudo)differential operators called $\textit{operators of uniform divergence type}$, and on perturbative arguments from spectral theory (especially on the theory of Pollicott-Ruelle resonances in the Anosov case).
△ Less
Submitted 19 October, 2020; v1 submitted 20 August, 2020;
originally announced August 2020.
-
Wireless Fingerprinting via Deep Learning: The Impact of Confounding Factors
Authors:
Metehan Cekic,
Soorya Gopalakrishnan,
Upamanyu Madhow
Abstract:
Can we distinguish between two wireless transmitters sending exactly the same message, using the same protocol? The opportunity for doing so arises due to subtle nonlinear variations across transmitters, even those made by the same manufacturer. Since these effects are difficult to model explicitly, we investigate learning device fingerprints using complex-valued deep neural networks (DNNs) that t…
▽ More
Can we distinguish between two wireless transmitters sending exactly the same message, using the same protocol? The opportunity for doing so arises due to subtle nonlinear variations across transmitters, even those made by the same manufacturer. Since these effects are difficult to model explicitly, we investigate learning device fingerprints using complex-valued deep neural networks (DNNs) that take as input the complex baseband signal at the receiver. We ask whether such fingerprints can be made robust to distribution shifts across time and locations due to clock drift and variations in the wireless channel. In this paper, we point out that, unless proactively discouraged from doing so, DNNs learn these strong confounding features rather than the nonlinear device-specific characteristics that we seek to learn. We propose and evaluate strategies, based on augmentation and estimation, to promote generalization across realizations of these confounding factors, using data from WiFi and ADS-B protocols. We conclude that, while DNN training has the advantage of not requiring explicit signal models, significant modeling insights are required to focus the learning on the effects we wish to capture.
△ Less
Submitted 9 March, 2021; v1 submitted 25 February, 2020;
originally announced February 2020.
-
Polarizing Front Ends for Robust CNNs
Authors:
Can Bakiskan,
Soorya Gopalakrishnan,
Metehan Cekic,
Upamanyu Madhow,
Ramtin Pedarsani
Abstract:
The vulnerability of deep neural networks to small, adversarially designed perturbations can be attributed to their "excessive linearity." In this paper, we propose a bottom-up strategy for attenuating adversarial perturbations using a nonlinear front end which polarizes and quantizes the data. We observe that ideal polarization can be utilized to completely eliminate perturbations, develop algori…
▽ More
The vulnerability of deep neural networks to small, adversarially designed perturbations can be attributed to their "excessive linearity." In this paper, we propose a bottom-up strategy for attenuating adversarial perturbations using a nonlinear front end which polarizes and quantizes the data. We observe that ideal polarization can be utilized to completely eliminate perturbations, develop algorithms to learn approximately polarizing bases for data, and investigate the effectiveness of the proposed strategy on the MNIST and Fashion MNIST datasets.
△ Less
Submitted 21 February, 2020;
originally announced February 2020.
-
Resonant spaces for volume preserving Anosov flows
Authors:
Mihajlo Cekić,
Gabriel P. Paternain
Abstract:
We consider Anosov flows on closed 3-manifolds preserving a volume form $Ω$. Following Dyatlov and Zworski (2017) we study spaces of invariant distributions with values in the bundle of exterior forms whose wavefront set is contained in the dual of the unstable bundle. Our first result computes the dimension of these spaces in terms of the first Betti number of the manifold, the cohomology class…
▽ More
We consider Anosov flows on closed 3-manifolds preserving a volume form $Ω$. Following Dyatlov and Zworski (2017) we study spaces of invariant distributions with values in the bundle of exterior forms whose wavefront set is contained in the dual of the unstable bundle. Our first result computes the dimension of these spaces in terms of the first Betti number of the manifold, the cohomology class $[ι_{X}Ω]$ (where $X$ is the infinitesimal generator of the flow) and the helicity. These dimensions coincide with the Pollicott-Ruelle resonance multiplicities under the assumption of $\textit{semisimplicity}$. We prove various results regarding semisimplicity on 1-forms, including an example showing that it may fail for time changes of hyperbolic geodesic flows. We also study non null-homologous deformations of contact Anosov flows and we show that there is always a splitting Pollicott-Ruelle resonance on 1-forms and that semisimplicity persists in this instance. These results have consequences for the order of vanishing at zero of the Ruelle zeta function. Finally our analysis also incorporates a flat unitary twist in both, the resonant spaces and the Ruelle zeta function.
△ Less
Submitted 1 December, 2020; v1 submitted 10 November, 2019;
originally announced November 2019.
-
Robust Wireless Fingerprinting via Complex-Valued Neural Networks
Authors:
Soorya Gopalakrishnan,
Metehan Cekic,
Upamanyu Madhow
Abstract:
A "wireless fingerprint" which exploits hardware imperfections unique to each device is a potentially powerful tool for wireless security. Such a fingerprint should be able to distinguish between devices sending the same message, and should be robust against standard spoofing techniques. Since the information in wireless signals resides in complex baseband, in this paper, we explore the use of neu…
▽ More
A "wireless fingerprint" which exploits hardware imperfections unique to each device is a potentially powerful tool for wireless security. Such a fingerprint should be able to distinguish between devices sending the same message, and should be robust against standard spoofing techniques. Since the information in wireless signals resides in complex baseband, in this paper, we explore the use of neural networks with complex-valued weights to learn fingerprints using supervised learning. We demonstrate that, while there are potential benefits to using sections of the signal beyond just the preamble to learn fingerprints, the network cheats when it can, using information such as transmitter ID (which can be easily spoofed) to artificially inflate performance. We also show that noise augmentation by inserting additional white Gaussian noise can lead to significant performance gains, which indicates that this counter-intuitive strategy helps in learning more robust fingerprints. We provide results for two different wireless protocols, WiFi and ADS-B, demonstrating the effectiveness of the proposed method.
△ Less
Submitted 26 August, 2019; v1 submitted 19 May, 2019;
originally announced May 2019.
-
Polyhedral billiards, eigenfunction concentration and almost periodic control
Authors:
Mihajlo Cekić,
Bogdan Georgiev,
Mayukh Mukherjee
Abstract:
We study dynamical properties of the billiard flow on convex polyhedra away from a neighbourhood of the non-smooth part of the boundary, called ``pockets''. We prove there are only finitely many immersed periodic tubes missing the pockets and moreover establish a new quantitative estimate for the lengths of such tubes. This extends well-known results in dimension $2$. We then apply these dynamical…
▽ More
We study dynamical properties of the billiard flow on convex polyhedra away from a neighbourhood of the non-smooth part of the boundary, called ``pockets''. We prove there are only finitely many immersed periodic tubes missing the pockets and moreover establish a new quantitative estimate for the lengths of such tubes. This extends well-known results in dimension $2$. We then apply these dynamical results to prove a quantitative Laplace eigenfunction mass concentration near the pockets of convex polyhedral billiards. As a technical tool for proving our concentration results on irrational polyhedra, we establish a control-theoretic estimate on a product space with an almost-periodic boundary condition. This extends previously known control estimates for periodic boundary conditions, and seems to be of independent interest.
△ Less
Submitted 17 March, 2020; v1 submitted 23 March, 2019;
originally announced March 2019.
-
Robust Adversarial Learning via Sparsifying Front Ends
Authors:
Soorya Gopalakrishnan,
Zhinus Marzi,
Metehan Cekic,
Upamanyu Madhow,
Ramtin Pedarsani
Abstract:
It is by now well-known that small adversarial perturbations can induce classification errors in deep neural networks. In this paper, we take a bottom-up signal processing perspective to this problem and show that a systematic exploitation of sparsity in natural data is a promising tool for defense. For linear classifiers, we show that a sparsifying front end is provably effective against…
▽ More
It is by now well-known that small adversarial perturbations can induce classification errors in deep neural networks. In this paper, we take a bottom-up signal processing perspective to this problem and show that a systematic exploitation of sparsity in natural data is a promising tool for defense. For linear classifiers, we show that a sparsifying front end is provably effective against $\ell_{\infty}$-bounded attacks, reducing output distortion due to the attack by a factor of roughly $K/N$ where $N$ is the data dimension and $K$ is the sparsity level. We then extend this concept to deep networks, showing that a "locally linear" model can be used to develop a theoretical foundation for crafting attacks and defenses. We also devise attacks based on the locally linear model that outperform the well-known FGSM attack. We supplement our theoretical results with experiments on the MNIST and CIFAR-10 datasets, showing the efficacy of the proposed sparsity-based defense schemes.
△ Less
Submitted 25 May, 2021; v1 submitted 24 October, 2018;
originally announced October 2018.
-
The Calderón problem for the fractional Schrödinger equation with drift
Authors:
Mihajlo Cekić,
Yi-Hsuan Lin,
Angkana Rüland
Abstract:
We investigate the Calderón problem for the fractional Schrödinger equation with drift, proving that the unknown drift and potential in a bounded domain can be determined simultaneously and uniquely by an infinite number of exterior measurements. In particular, in contrast to its local analogue, this nonlocal problem does \emph{not} enjoy a gauge invariance. The uniqueness result is complemented b…
▽ More
We investigate the Calderón problem for the fractional Schrödinger equation with drift, proving that the unknown drift and potential in a bounded domain can be determined simultaneously and uniquely by an infinite number of exterior measurements. In particular, in contrast to its local analogue, this nonlocal problem does \emph{not} enjoy a gauge invariance. The uniqueness result is complemented by an associated logarithmic stability estimate under suitable apriori assumptions. Also uniqueness under finitely many \emph{generic} measurements is discussed. Here the genericity is obtained through \emph{singularity theory} which might also be interesting in the context of hybrid inverse problems. Combined with the results from \cite{GRSU18}, this yields a finite measurements constructive reconstruction algorithm for the fractional Calderón problem with drift. The inverse problem is formulated as a partial data type nonlocal problem and it is considered in any dimension $n\geq 1$.
△ Less
Submitted 18 December, 2018; v1 submitted 9 October, 2018;
originally announced October 2018.
-
Harmonic determinants and unique continuation
Authors:
Mihajlo Cekić
Abstract:
We give partial answers to the following question: if $F$ is an $m$ by $m$ matrix on $\mathbb{R}^n$ satisfying a second order linear elliptic equation, does $\det F$ satisfy the strong unique continuation property? We give counterexamples in the case when the operator is a general non-diagonal operator and also for some diagonal operators. Positive results are obtained when $n = 1$ and any $m$, wh…
▽ More
We give partial answers to the following question: if $F$ is an $m$ by $m$ matrix on $\mathbb{R}^n$ satisfying a second order linear elliptic equation, does $\det F$ satisfy the strong unique continuation property? We give counterexamples in the case when the operator is a general non-diagonal operator and also for some diagonal operators. Positive results are obtained when $n = 1$ and any $m$, when $n = 2$ for the Laplace-Beltrami operator and also twisted with a Yang-Mills connection. Reductions to special cases when $n = 2$ are obtained. The last section considers an application to the Calderón problem in 2D based on recent techniques.
△ Less
Submitted 24 March, 2018;
originally announced March 2018.
-
Calderón problem for Yang-Mills connections
Authors:
Mihajlo Cekić
Abstract:
We consider the problem of identifying a unitary Yang-Mills connection $\nabla$ on a Hermitian vector bundle from the Dirichlet-to-Neumann (DN) map of the connection Laplacian $\nabla^*\nabla$ over compact Riemannian manifolds with boundary. We establish uniqueness of the connection up to a gauge equivalence in the case of trivial line bundles in the smooth category and for the higher rank case in…
▽ More
We consider the problem of identifying a unitary Yang-Mills connection $\nabla$ on a Hermitian vector bundle from the Dirichlet-to-Neumann (DN) map of the connection Laplacian $\nabla^*\nabla$ over compact Riemannian manifolds with boundary. We establish uniqueness of the connection up to a gauge equivalence in the case of trivial line bundles in the smooth category and for the higher rank case in the analytic category, by using geometric analysis methods and essentially only one measurement.
Moreover, by using a Runge-type approximation argument along curves to recover holonomy, we are able to uniquely determine both the bundle structure and the connection, but at the cost of having more measurements. Also, we prove that the DN map is an elliptic pseudodifferential operator of order one on the restriction of the vector bundle to the boundary, whose full symbol determines the complete Taylor series of an arbitrary connection, metric and an associated potential at the boundary.
△ Less
Submitted 12 June, 2018; v1 submitted 5 April, 2017;
originally announced April 2017.
-
The Calderón problem for connections
Authors:
Mihajlo Cekić
Abstract:
In this paper we consider the problem of identifying a connection $\nabla$ on a vector bundle up to gauge equivalence from the Dirichlet-to-Neumann map of the connection Laplacian $\nabla^*\nabla$ over conformally transversally anisotropic (CTA) manifolds. This was proved in \cite{LCW} for line bundles in the case of the transversal manifold being simple -- we generalise this result to the case wh…
▽ More
In this paper we consider the problem of identifying a connection $\nabla$ on a vector bundle up to gauge equivalence from the Dirichlet-to-Neumann map of the connection Laplacian $\nabla^*\nabla$ over conformally transversally anisotropic (CTA) manifolds. This was proved in \cite{LCW} for line bundles in the case of the transversal manifold being simple -- we generalise this result to the case where the transversal manifold only has an injective ray transform. Moreover, the construction of suitable Gaussian beam solutions on vector bundles is given for the case of the connection Laplacian and a potential, following the works of \cite{CTA}. This in turn enables us to construct the Complex Geometrical Optics (CGO) solutions and prove our main uniqueness result. We also reduce the problem to a new non-abelian X-ray transform for the case of simple transversal manifolds and higher rank vector bundles. Finally, we prove the recovery of a flat connection in general from the DN map, up to gauge equivalence, using an argument relating the Cauchy data of the connection Laplacian and the holonomy.
△ Less
Submitted 8 October, 2017; v1 submitted 10 October, 2016;
originally announced October 2016.
