@inproceedings{10595, abstract = {A recent line of work has analyzed the theoretical properties of deep neural networks via the Neural Tangent Kernel (NTK). In particular, the smallest eigenvalue of the NTK has been related to the memorization capacity, the global convergence of gradient descent algorithms and the generalization of deep nets. However, existing results either provide bounds in the two-layer setting or assume that the spectrum of the NTK matrices is bounded away from 0 for multi-layer networks. In this paper, we provide tight bounds on the smallest eigenvalue of NTK matrices for deep ReLU nets, both in the limiting case of infinite widths and for finite widths. In the finite-width setting, the network architectures we consider are fairly general: we require the existence of a wide layer with roughly order of $N$ neurons, $N$ being the number of data samples; and the scaling of the remaining layer widths is arbitrary (up to logarithmic factors). To obtain our results, we analyze various quantities of independent interest: we give lower bounds on the smallest singular value of hidden feature matrices, and upper bounds on the Lipschitz constant of input-output feature maps.}, author = {Nguyen, Quynh and Mondelli, Marco and Montufar, Guido F}, booktitle = {Proceedings of the 38th International Conference on Machine Learning}, editor = {Meila, Marina and Zhang, Tong}, location = {Virtual}, pages = {8119--8129}, publisher = {ML Research Press}, title = {{Tight bounds on the smallest eigenvalue of the neural tangent kernel for deep ReLU networks}}, volume = {139}, year = {2021}, } @inproceedings{10597, abstract = {We thank Emmanuel Abbe and Min Ye for providing us the implementation of RPA decoding. D. Fathollahi and M. Mondelli are partially supported by the 2019 Lopez-Loreta Prize. N. Farsad is supported by Discovery Grant from the Natural Sciences and Engineering Research Council of Canada (NSERC) and Canada Foundation for Innovation (CFI), John R. Evans Leader Fund. S. A. Hashemi is supported by a Postdoctoral Fellowship from NSERC.}, author = {Fathollahi, Dorsa and Farsad, Nariman and Hashemi, Seyyed Ali and Mondelli, Marco}, booktitle = {2021 IEEE International Symposium on Information Theory}, isbn = {978-1-5386-8210-4}, location = {Virtual, Melbourne, Australia}, pages = {1082--1087}, publisher = {Institute of Electrical and Electronics Engineers}, title = {{Sparse multi-decoder recursive projection aggregation for Reed-Muller codes}}, doi = {10.1109/isit45174.2021.9517887}, year = {2021}, } @inproceedings{10598, abstract = { We consider the problem of estimating a signal from measurements obtained via a generalized linear model. We focus on estimators based on approximate message passing (AMP), a family of iterative algorithms with many appealing features: the performance of AMP in the high-dimensional limit can be succinctly characterized under suitable model assumptions; AMP can also be tailored to the empirical distribution of the signal entries, and for a wide class of estimation problems, AMP is conjectured to be optimal among all polynomial-time algorithms. However, a major issue of AMP is that in many models (such as phase retrieval), it requires an initialization correlated with the ground-truth signal and independent from the measurement matrix. Assuming that such an initialization is available is typically not realistic. In this paper, we solve this problem by proposing an AMP algorithm initialized with a spectral estimator. With such an initialization, the standard AMP analysis fails since the spectral estimator depends in a complicated way on the design matrix. Our main contribution is a rigorous characterization of the performance of AMP with spectral initialization in the high-dimensional limit. The key technical idea is to define and analyze a two-phase artificial AMP algorithm that first produces the spectral estimator, and then closely approximates the iterates of the true AMP. We also provide numerical results that demonstrate the validity of the proposed approach. }, author = {Mondelli, Marco and Venkataramanan, Ramji}, booktitle = {Proceedings of The 24th International Conference on Artificial Intelligence and Statistics}, editor = {Banerjee, Arindam and Fukumizu, Kenji}, issn = {2640-3498}, location = {Virtual, San Diego, CA, United States}, pages = {397--405}, publisher = {ML Research Press}, title = {{Approximate message passing with spectral initialization for generalized linear models}}, volume = {130}, year = {2021}, } @inproceedings{10599, abstract = {A two-part successive syndrome-check decoding of polar codes is proposed with the first part successively refining the received codeword and the second part checking its syndrome. A new formulation of the successive-cancellation (SC) decoding algorithm is presented that allows for successively refining the received codeword by comparing the log-likelihood ratio value of a frozen bit with its predefined value. The syndrome of the refined received codeword is then checked for possible errors. In case there are no errors, the decoding process is terminated. Otherwise, the decoder continues to refine the received codeword. The proposed method is extended to the case of SC list (SCL) decoding by terminating the decoding process when the syndrome of the best candidate in the list indicates no errors. Simulation results show that the proposed method reduces the time-complexity of SC and SCL decoders and their fast variants, especially at high signal-to-noise ratios.}, author = {Hashemi, Seyyed Ali and Mondelli, Marco and Cioffi, John and Goldsmith, Andrea}, booktitle = {Proceedings of the 55th Asilomar Conference on Signals, Systems, and Computers}, isbn = {9781665458283}, issn = {1058-6393}, location = {Virtual, Pacific Grove, CA, United States}, pages = {943--947}, publisher = {Institute of Electrical and Electronics Engineers}, title = {{Successive syndrome-check decoding of polar codes}}, doi = {10.1109/IEEECONF53345.2021.9723394}, volume = {2021-October}, year = {2021}, } @article{10606, abstract = {Cell division orientation is thought to result from a competition between cell geometry and polarity domains controlling the position of the mitotic spindle during mitosis. Depending on the level of cell shape anisotropy or the strength of the polarity domain, one dominates the other and determines the orientation of the spindle. Whether and how such competition is also at work to determine unequal cell division (UCD), producing daughter cells of different size, remains unclear. Here, we show that cell geometry and polarity domains cooperate, rather than compete, in positioning the cleavage plane during UCDs in early ascidian embryos. We found that the UCDs and their orientation at the ascidian third cleavage rely on the spindle tilting in an anisotropic cell shape, and cortical polarity domains exerting different effects on spindle astral microtubules. By systematically varying mitotic cell shape, we could modulate the effect of attractive and repulsive polarity domains and consequently generate predicted daughter cell size asymmetries and position. We therefore propose that the spindle position during UCD is set by the combined activities of cell geometry and polarity domains, where cell geometry modulates the effect of cortical polarity domain(s).}, author = {Godard, Benoit G and Dumollard, Remi and Heisenberg, Carl-Philipp J and Mcdougall, Alex}, issn = {2050-084X}, journal = {eLife}, publisher = {eLife Sciences Publications}, title = {{Combined effect of cell geometry and polarity domains determines the orientation of unequal division}}, doi = {10.7554/eLife.75639}, volume = {10}, year = {2021}, } @article{10607, abstract = {The evidence linking innate immunity mechanisms and neurodegenerative diseases is growing, but the specific mechanisms are incompletely understood. Experimental data suggest that microglial TLR4 mediates the uptake and clearance of α-synuclein also termed synucleinophagy. The accumulation of misfolded α-synuclein throughout the brain is central to Parkinson's disease (PD). The distribution and progression of the pathology is often attributed to the propagation of α-synuclein. Here, we apply a classical α-synuclein propagation model of prodromal PD in wild type and TLR4 deficient mice to study the role of TLR4 in the progression of the disease. Our data suggest that TLR4 deficiency facilitates the α-synuclein seed spreading associated with reduced lysosomal activity of microglia. Three months after seed inoculation, more pronounced proteinase K-resistant α-synuclein inclusion pathology is observed in mice with TLR4 deficiency. The facilitated propagation of α-synuclein is associated with early loss of dopamine transporter (DAT) signal in the striatum and loss of dopaminergic neurons in substantia nigra pars compacta of TLR4 deficient mice. These new results support TLR4 signaling as a putative target for disease modification to slow the progression of PD and related disorders.}, author = {Venezia, Serena and Kaufmann, Walter and Wenning, Gregor K. and Stefanova, Nadia}, issn = {1873-5126}, journal = {Parkinsonism & Related Disorders}, pages = {59--65}, publisher = {Elsevier}, title = {{Toll-like receptor 4 deficiency facilitates α-synuclein propagation and neurodegeneration in a mouse model of prodromal Parkinson's disease}}, doi = {10.1016/j.parkreldis.2021.09.007}, volume = {91}, year = {2021}, } @inproceedings{10609, abstract = {We study Multi-party computation (MPC) in the setting of subversion, where the adversary tampers with the machines of honest parties. Our goal is to construct actively secure MPC protocols where parties are corrupted adaptively by an adversary (as in the standard adaptive security setting), and in addition, honest parties’ machines are compromised. The idea of reverse firewalls (RF) was introduced at EUROCRYPT’15 by Mironov and Stephens-Davidowitz as an approach to protecting protocols against corruption of honest parties’ devices. Intuitively, an RF for a party P is an external entity that sits between P and the outside world and whose scope is to sanitize P ’s incoming and outgoing messages in the face of subversion of their computer. Mironov and Stephens-Davidowitz constructed a protocol for passively-secure two-party computation. At CRYPTO’20, Chakraborty, Dziembowski and Nielsen constructed a protocol for secure computation with firewalls that improved on this result, both by extending it to multi-party computation protocol, and considering active security in the presence of static corruptions. In this paper, we initiate the study of RF for MPC in the adaptive setting. We put forward a definition for adaptively secure MPC in the reverse firewall setting, explore relationships among the security notions, and then construct reverse firewalls for MPC in this stronger setting of adaptive security. We also resolve the open question of Chakraborty, Dziembowski and Nielsen by removing the need for a trusted setup in constructing RF for MPC. Towards this end, we construct reverse firewalls for adaptively secure augmented coin tossing and adaptively secure zero-knowledge protocols and obtain a constant round adaptively secure MPC protocol in the reverse firewall setting without setup. Along the way, we propose a new multi-party adaptively secure coin tossing protocol in the plain model, that is of independent interest.}, author = {Chakraborty, Suvradip and Ganesh, Chaya and Pancholi, Mahak and Sarkar, Pratik}, booktitle = {27th International Conference on the Theory and Application of Cryptology and Information Security}, isbn = {978-3-030-92074-6}, issn = {1611-3349}, location = {Virtual, Singapore}, pages = {335--364}, publisher = {Springer Nature}, title = {{Reverse firewalls for adaptively secure MPC without setup}}, doi = {10.1007/978-3-030-92075-3_12}, volume = {13091}, year = {2021}, } @article{10613, abstract = {Motivated by the recent preprint [\emph{arXiv:2004.08412}] by Ayala, Carinci, and Redig, we first provide a general framework for the study of scaling limits of higher-order fields. Then, by considering the same class of infinite interacting particle systems as in [\emph{arXiv:2004.08412}], namely symmetric simple exclusion and inclusion processes in the d-dimensional Euclidean lattice, we prove the hydrodynamic limit, and convergence for the equilibrium fluctuations, of higher-order fields. In particular, the limit fields exhibit a tensor structure. Our fluctuation result differs from that in [\emph{arXiv:2004.08412}], since we considered-dimensional Euclidean lattice, we prove the hydrodynamic limit, and convergence for the equilibrium fluctuations, of higher-order fields. In particular, the limit fields exhibit a tensor structure. Our fluctuation result differs from that in [\emph{arXiv:2004.08412}], since we consider a different notion of higher-order fluctuation fields.}, author = {Chen, Joe P. and Sau, Federico}, issn = {1024-2953}, journal = {Markov Processes And Related Fields}, keywords = {interacting particle systems, higher-order fields, hydrodynamic limit, equilibrium fluctuations, duality}, number = {3}, pages = {339--380}, publisher = {Polymat Publishing}, title = {{Higher-order hydrodynamics and equilibrium fluctuations of interacting particle systems}}, volume = {27}, year = {2021}, } @article{10628, abstract = {The surface states of 3D topological insulators in general have negligible quantum oscillations (QOs) when the chemical potential is tuned to the Dirac points. In contrast, we find that topological Kondo insulators (TKIs) can support surface states with an arbitrarily large Fermi surface (FS) when the chemical potential is pinned to the Dirac point. We illustrate that these FSs give rise to finite-frequency QOs, which can become comparable to the extremal area of the unhybridized bulk bands. We show that this occurs when the crystal symmetry is lowered from cubic to tetragonal in a minimal two-orbital model. We label such surface modes as 'shadow surface states'. Moreover, we show that the sufficient next-nearest neighbor out-of-plane hybridization leading to shadow surface states can be self-consistently stabilized for tetragonal TKIs. Consequently, shadow surface states provide an important example of high-frequency QOs beyond the context of cubic TKIs.}, author = {Ghazaryan, Areg and Nica, Emilian M. and Erten, Onur and Ghaemi, Pouyan}, issn = {1367-2630}, journal = {New Journal of Physics}, number = {12}, publisher = {IOP Publishing}, title = {{Shadow surface states in topological Kondo insulators}}, doi = {10.1088/1367-2630/ac4124}, volume = {23}, year = {2021}, } @inproceedings{10629, abstract = {Product graphs arise naturally in formal verification and program analysis. For example, the analysis of two concurrent threads requires the product of two component control-flow graphs, and for language inclusion of deterministic automata the product of two automata is constructed. In many cases, the component graphs have constant treewidth, e.g., when the input contains control-flow graphs of programs. We consider the algorithmic analysis of products of two constant-treewidth graphs with respect to three classic specification languages, namely, (a) algebraic properties, (b) mean-payoff properties, and (c) initial credit for energy properties. Our main contributions are as follows. Consider a graph G that is the product of two constant-treewidth graphs of size n each. First, given an idempotent semiring, we present an algorithm that computes the semiring transitive closure of G in time Õ(n⁴). Since the output has size Θ(n⁴), our algorithm is optimal (up to polylog factors). Second, given a mean-payoff objective, we present an O(n³)-time algorithm for deciding whether the value of a starting state is non-negative, improving the previously known O(n⁴) bound. Third, given an initial credit for energy objective, we present an O(n⁵)-time algorithm for computing the minimum initial credit for all nodes of G, improving the previously known O(n⁸) bound. At the heart of our approach lies an algorithm for the efficient construction of strongly-balanced tree decompositions of constant-treewidth graphs. Given a constant-treewidth graph G' of n nodes and a positive integer λ, our algorithm constructs a binary tree decomposition of G' of width O(λ) with the property that the size of each subtree decreases geometrically with rate (1/2 + 2^{-λ}).}, author = {Chatterjee, Krishnendu and Ibsen-Jensen, Rasmus and Pavlogiannis, Andreas}, booktitle = {41st IARCS Annual Conference on Foundations of Software Technology and Theoretical Computer Science}, isbn = {978-3-9597-7215-0}, issn = {1868-8969}, location = {Virtual}, publisher = {Schloss Dagstuhl - Leibniz-Zentrum für Informatik}, title = {{Quantitative verification on product graphs of small treewidth}}, doi = {10.4230/LIPIcs.FSTTCS.2021.42}, volume = {213}, year = {2021}, } @inproceedings{10630, abstract = {In the Intersection Non-emptiness problem, we are given a list of finite automata A_1, A_2,… , A_m over a common alphabet Σ as input, and the goal is to determine whether some string w ∈ Σ^* lies in the intersection of the languages accepted by the automata in the list. We analyze the complexity of the Intersection Non-emptiness problem under the promise that all input automata accept a language in some level of the dot-depth hierarchy, or some level of the Straubing-Thérien hierarchy. Automata accepting languages from the lowest levels of these hierarchies arise naturally in the context of model checking. We identify a dichotomy in the dot-depth hierarchy by showing that the problem is already NP-complete when all input automata accept languages of the levels B_0 or B_{1/2} and already PSPACE-hard when all automata accept a language from the level B_1. Conversely, we identify a tetrachotomy in the Straubing-Thérien hierarchy. More precisely, we show that the problem is in AC^0 when restricted to level L_0; complete for L or NL, depending on the input representation, when restricted to languages in the level L_{1/2}; NP-complete when the input is given as DFAs accepting a language in L_1 or L_{3/2}; and finally, PSPACE-complete when the input automata accept languages in level L_2 or higher. Moreover, we show that the proof technique used to show containment in NP for DFAs accepting languages in L_1 or L_{3/2} does not generalize to the context of NFAs. To prove this, we identify a family of languages that provide an exponential separation between the state complexity of general NFAs and that of partially ordered NFAs. To the best of our knowledge, this is the first superpolynomial separation between these two models of computation.}, author = {Arrighi, Emmanuel and Fernau, Henning and Hoffmann, Stefan and Holzer, Markus and Jecker, Ismael R and De Oliveira Oliveira, Mateus and Wolf, Petra}, booktitle = {41st IARCS Annual Conference on Foundations of Software Technology and Theoretical Computer Science}, isbn = {978-3-9597-7215-0}, issn = {1868-8969}, location = {Virtual}, publisher = {Schloss Dagstuhl - Leibniz-Zentrum für Informatik}, title = {{On the complexity of intersection non-emptiness for star-free language classes}}, doi = {10.4230/LIPIcs.FSTTCS.2021.34}, volume = {213}, year = {2021}, } @article{10631, abstract = {We combine experimental and theoretical approaches to explore excited rotational states of molecules embedded in helium nanodroplets using CS2 and I2 as examples. Laser-induced nonadiabatic molecular alignment is employed to measure spectral lines for rotational states extending beyond those initially populated at the 0.37 K droplet temperature. We construct a simple quantum-mechanical model, based on a linear rotor coupled to a single-mode bosonic bath, to determine the rotational energy structure in its entirety. The calculated and measured spectral lines are in good agreement. We show that the effect of the surrounding superfluid on molecular rotation can be rationalized by a single quantity, the angular momentum, transferred from the molecule to the droplet.}, author = {Cherepanov, Igor and Bighin, Giacomo and Schouder, Constant A. and Chatterley, Adam S. and Albrechtsen, Simon H. and Muñoz, Alberto Viñas and Christiansen, Lars and Stapelfeldt, Henrik and Lemeshko, Mikhail}, issn = {2469-9934}, journal = {Physical Review A}, number = {6}, publisher = {American Physical Society}, title = {{Excited rotational states of molecules in a superfluid}}, doi = {10.1103/PhysRevA.104.L061303}, volume = {104}, year = {2021}, } @article{10635, abstract = {The brain efficiently performs nonlinear computations through its intricate networks of spiking neurons, but how this is done remains elusive. While nonlinear computations can be implemented successfully in spiking neural networks, this requires supervised training and the resulting connectivity can be hard to interpret. In contrast, the required connectivity for any computation in the form of a linear dynamical system can be directly derived and understood with the spike coding network (SCN) framework. These networks also have biologically realistic activity patterns and are highly robust to cell death. Here we extend the SCN framework to directly implement any polynomial dynamical system, without the need for training. This results in networks requiring a mix of synapse types (fast, slow, and multiplicative), which we term multiplicative spike coding networks (mSCNs). Using mSCNs, we demonstrate how to directly derive the required connectivity for several nonlinear dynamical systems. We also show how to carry out higher-order polynomials with coupled networks that use only pair-wise multiplicative synapses, and provide expected numbers of connections for each synapse type. Overall, our work demonstrates a novel method for implementing nonlinear computations in spiking neural networks, while keeping the attractive features of standard SCNs (robustness, realistic activity patterns, and interpretable connectivity). Finally, we discuss the biological plausibility of our approach, and how the high accuracy and robustness of the approach may be of interest for neuromorphic computing.}, author = {Nardin, Michele and Phillips, James W. and Podlaski, William F. and Keemink, Sander W.}, issn = {2804-3871}, journal = {Peer Community Journal}, publisher = {Peer Community In}, title = {{Nonlinear computations in spiking neural networks through multiplicative synapses}}, doi = {10.24072/pcjournal.69}, volume = {1}, year = {2021}, } @misc{10644, abstract = {The purpose of this application note is to demonstrate a working example of a superconducting qubit measurement in a Bluefors cryostat using the Keysight quantum control hardware. Our motivation is twofold. First, we provide pre-qualification data that the Bluefors cryostat, including filtering and wiring, can support long-lived qubits. Second, we demonstrate that the Keysight system (controlled using Labber) provides a straightforward solution to perform these characterization measurements. This document is intended as a brief guide for starting an experimental platform for testing superconducting qubits. The setup described here is an immediate jumping off point for a suite of applications including testing quantum logical gates, quantum optics with microwaves, or even using the qubit itself as a sensitive probe of local electromagnetic fields. Qubit measurements rely on high performance of both the physical sample environment and the measurement electronics. An overview of the cryogenic system is shown in Figure 1, and an overview of the integration between the electronics and cryostat (including wiring details) is shown in Figure 2.}, author = {Lake, Russell and Simbierowicz, Slawomir and Krantz, Philip and Hassani, Farid and Fink, Johannes M}, keywords = {Application note}, pages = {9}, publisher = {Bluefors Oy}, title = {{The Bluefors dilution refrigerator as an integrated quantum measurement system}}, year = {2021}, } @misc{10645, abstract = {Superconducting qubits have emerged as a highly versatile and useful platform for quantum technological applications [1]. Bluefors and Zurich Instruments have supported the growth of this field from the 2010s onwards by providing well-engineered and reliable measurement infrastructure [2]– [6]. Having a long and stable qubit lifetime is a critical system property. Therefore, considerable effort has already gone into measuring qubit energy-relaxation timescales and their fluctuations, see Refs. [7]–[10] among others. Accurately extracting the statistics of a quantum device requires users to perform time consuming measurements. One measurement challenge is that the detection of the state-dependent response of a superconducting resonator due to a dispersively-coupled qubit requires an inherently low signal level. Consequently, measurements must be performed using a microwave probe that contains only a few microwave photons. Improving the signal-to-noise ratio (SNR) by using near-quantum limited parametric amplifiers as well as the use of optimized signal processing enabled by efficient room temperature instrumentation help to reduce measurement time. An empirical observation for fixed frequency transmons from recent literature is that as the energy-relaxation time 𝑇𝑇1 increases, so do its natural temporal fluctuations [7], [10]. This necessitates many repeated measurements to understand the statistics (see for example, Ref. [10]). In addition, as state-of-the-art qubits increase in lifetime, longer measurement times are expected to obtain accurate statistics. As described below, the scaling of the widths of the qubit energy-relaxation distributions also reveal clues about the origin of the energy-relaxation.}, author = {Simbierowicz, Slawomir and Shi, Chunyan and Collodo, Michele and Kirste, Moritz and Hassani, Farid and Fink, Johannes M and Bylander, Jonas and Perez Lozano, Daniel and Lake, Russell}, keywords = {Application note}, pages = {8}, publisher = {Bluefors Oy}, title = {{Qubit energy-relaxation statistics in the Bluefors quantum measurement system}}, year = {2021}, } @article{10655, abstract = {Adeno-associated viruses (AAVs) are widely used to deliver genetic material in vivo to distinct cell types such as neurons or glial cells, allowing for targeted manipulation. Transduction of microglia is mostly excluded from this strategy, likely due to the cells’ heterogeneous state upon environmental changes, which makes AAV design challenging. Here, we established the retina as a model system for microglial AAV validation and optimization. First, we show that AAV2/6 transduced microglia in both synaptic layers, where layer preference corresponds to the intravitreal or subretinal delivery method. Surprisingly, we observed significantly enhanced microglial transduction during photoreceptor degeneration. Thus, we modified the AAV6 capsid to reduce heparin binding by introducing four point mutations (K531E, R576Q, K493S, and K459S), resulting in increased microglial transduction in the outer plexiform layer. Finally, to improve microglial-specific transduction, we validated a Cre-dependent transgene delivery cassette for use in combination with the Cx3cr1CreERT2 mouse line. Together, our results provide a foundation for future studies optimizing AAV-mediated microglia transduction and highlight that environmental conditions influence microglial transduction efficiency. }, author = {Maes, Margaret E and Wögenstein, Gabriele M. and Colombo, Gloria and Casado Polanco, Raquel and Siegert, Sandra}, issn = {2329-0501}, journal = {Molecular Therapy - Methods and Clinical Development}, pages = {210--224}, publisher = {Elsevier}, title = {{Optimizing AAV2/6 microglial targeting identified enhanced efficiency in the photoreceptor degenerative environment}}, doi = {10.1016/j.omtm.2021.09.006}, volume = {23}, year = {2021}, } @inproceedings{10668, abstract = {Robustness to variations in lighting conditions is a key objective for any deep vision system. To this end, our paper extends the receptive field of convolutional neural networks with two residual components, ubiquitous in the visual processing system of vertebrates: On-center and off-center pathways, with an excitatory center and inhibitory surround; OOCS for short. The On-center pathway is excited by the presence of a light stimulus in its center, but not in its surround, whereas the Off-center pathway is excited by the absence of a light stimulus in its center, but not in its surround. We design OOCS pathways via a difference of Gaussians, with their variance computed analytically from the size of the receptive fields. OOCS pathways complement each other in their response to light stimuli, ensuring this way a strong edge-detection capability, and as a result an accurate and robust inference under challenging lighting conditions. We provide extensive empirical evidence showing that networks supplied with OOCS pathways gain accuracy and illumination-robustness from the novel edge representation, compared to other baselines.}, author = {Babaiee, Zahra and Hasani, Ramin and Lechner, Mathias and Rus, Daniela and Grosu, Radu}, booktitle = {Proceedings of the 38th International Conference on Machine Learning}, issn = {2640-3498}, location = {Virtual}, pages = {478--489}, publisher = {ML Research Press}, title = {{On-off center-surround receptive fields for accurate and robust image classification}}, volume = {139}, year = {2021}, } @inproceedings{10669, abstract = {We show that Neural ODEs, an emerging class of timecontinuous neural networks, can be verified by solving a set of global-optimization problems. For this purpose, we introduce Stochastic Lagrangian Reachability (SLR), an abstraction-based technique for constructing a tight Reachtube (an over-approximation of the set of reachable states over a given time-horizon), and provide stochastic guarantees in the form of confidence intervals for the Reachtube bounds. SLR inherently avoids the infamous wrapping effect (accumulation of over-approximation errors) by performing local optimization steps to expand safe regions instead of repeatedly forward-propagating them as is done by deterministic reachability methods. To enable fast local optimizations, we introduce a novel forward-mode adjoint sensitivity method to compute gradients without the need for backpropagation. Finally, we establish asymptotic and non-asymptotic convergence rates for SLR.}, author = {Grunbacher, Sophie and Hasani, Ramin and Lechner, Mathias and Cyranka, Jacek and Smolka, Scott A and Grosu, Radu}, booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence}, isbn = {978-1-57735-866-4}, issn = {2374-3468}, location = {Virtual}, number = {13}, pages = {11525--11535}, publisher = {AAAI Press}, title = {{On the verification of neural ODEs with stochastic guarantees}}, volume = {35}, year = {2021}, } @inproceedings{10670, abstract = {Imitation learning enables high-fidelity, vision-based learning of policies within rich, photorealistic environments. However, such techniques often rely on traditional discrete-time neural models and face difficulties in generalizing to domain shifts by failing to account for the causal relationships between the agent and the environment. In this paper, we propose a theoretical and experimental framework for learning causal representations using continuous-time neural networks, specifically over their discrete-time counterparts. We evaluate our method in the context of visual-control learning of drones over a series of complex tasks, ranging from short- and long-term navigation, to chasing static and dynamic objects through photorealistic environments. Our results demonstrate that causal continuous-time deep models can perform robust navigation tasks, where advanced recurrent models fail. These models learn complex causal control representations directly from raw visual inputs and scale to solve a variety of tasks using imitation learning.}, author = {Vorbach, Charles J and Hasani, Ramin and Amini, Alexander and Lechner, Mathias and Rus, Daniela}, booktitle = {35th Conference on Neural Information Processing Systems}, location = {Virtual}, title = {{Causal navigation by continuous-time neural networks}}, year = {2021}, } @inproceedings{10671, abstract = {We introduce a new class of time-continuous recurrent neural network models. Instead of declaring a learning system’s dynamics by implicit nonlinearities, we construct networks of linear first-order dynamical systems modulated via nonlinear interlinked gates. The resulting models represent dynamical systems with varying (i.e., liquid) time-constants coupled to their hidden state, with outputs being computed by numerical differential equation solvers. These neural networks exhibit stable and bounded behavior, yield superior expressivity within the family of neural ordinary differential equations, and give rise to improved performance on time-series prediction tasks. To demonstrate these properties, we first take a theoretical approach to find bounds over their dynamics, and compute their expressive power by the trajectory length measure in a latent trajectory space. We then conduct a series of time-series prediction experiments to manifest the approximation capability of Liquid Time-Constant Networks (LTCs) compared to classical and modern RNNs.}, author = {Hasani, Ramin and Lechner, Mathias and Amini, Alexander and Rus, Daniela and Grosu, Radu}, booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence}, isbn = {978-1-57735-866-4}, issn = {2374-3468}, location = {Virtual}, number = {9}, pages = {7657--7666}, publisher = {AAAI Press}, title = {{Liquid time-constant networks}}, volume = {35}, year = {2021}, }