Farming requires a balance of providing enough inputs, like water and fertilizer, to ensure sufficient yield while minimizing the consequences of overusing these resources, which can be costly, scarce, or lead to adverse environmental effects, such as eutrophication caused by nutrient run off. Current soil monitoring practices involve manually collecting soil samples that are sent to off site labs to be tested. In this talk, I’ll discuss the design of a platform for in-situ soil nutrient monitoring, which will enable the collection of high spatial resolutiIon information about multiple soil parameters.

Multi-agent reinforcement learning (RL) has found widespread success in multi-agent systems with decision-making agents. Its applications span diverse domains, ranging from playing games to navigating robotics to making economic policy. Unfortunately, the curse of dimensionality and/or multiagency poses scalability challenges for classical RL methods, particularly when dealing with large state spaces and/or numerous agents. This talk addresses the scalability challenges in the framework of independent learning (i.e., no coordination among agents), and focuses on Markov potential games for multi-agent RL. Firstly, we introduce an independent policy gradient method, where each agent updates its policy myopically using policy gradients. When the policy gradient is evaluated exactly, this method demonstrates non-asymptotic convergence to a near-Nash policy, with a polynomial dependency on the number of agents. Interestingly, the iteration complexity is not explicitly dependent on the state space size. Secondly, when the exact policy gradient is unavailable, we propose the function approximation of the value function to handle large state spaces and present polynomial sample complexity of our method in finding a near-Nash policy, showcasing the efficacy of our method in large-scale Markov games. Finally, we conclude with the game-agonistic convergence property for a variant of our method, finding a near-Nash policy in more than two types of games.

We consider joint trajectory generation and tracking control for under-actuated robotic systems. A common solution is to use a layered control architecture, where the top layer uses a simplified model of system dynamics for trajectory generation, and the low layer ensures approximate tracking of this trajectory via feedback control. While such layered control architectures are standard and work well in practice, selecting the simplified model used for trajectory generation typically relies on engineering intuition and experience. First, we propose a layered control architecture for exact reference tracking control using a suitable augmented Lagrangian reformulation of a linear optimal control problem and solve via primal-dual optimization without assuming a simplified model. We show that the resulting controller is an optimal linear feedback controller stabilizing system trajectories around dynamically feasible reference trajectories. We discuss implications of this result for quantifying sub-optimality of existing layered architectures for robot navigation, considering the context of a cartpole, a unicycle, and a quadrotor control problem in simulation. Next, we propose an alternative data-driven approach to dynamics-aware trajectory generation and show that applying a penalty method to a global nonlinear optimal control problem results in a relaxed layered decomposition of the overall problem into trajectory planning and feedback control layers. Crucially, the resulting trajectory optimization is dynamics-aware, in that it is modified with a tracking penalty regularizer encoding the dynamic feasibility of the generated trajectory. We show that this tracking penalty regularizer can be learned from system rollouts for independently-designed low layer feedback control policies. Furthermore, we showcase our framework on two quadrotor hardware platforms to illustrate that our approach i) avoids conservatism from ad hoc maximum speed and acceleration constraints and, ii) outperforms baselines in practical considerations such as aerodynamic drag and controller saturation.

In tribute to Boris Polyak, this talk focuses on analyzing the convergence of the Heavy-Ball method, an optimization technique that has gained immense attraction in the realm of Machine Learning. The talk extensively uses Stochastic Approximation, a 1950’s paradigm currently used in Reinforcement Learning. By framing Heavy-Ball as a stationary point problem, this work provides a generic framework that guarantees convergence under convex and certain class of non-convex objectives. This work shows convergence guarantees in both gradient-free methods and block coordinate updates, which have great computational savings in large-scale optimization problems.

Feature learning is thought to be one of the fundamental reasons for the success of deep neural networks. It is rigorously known that in two-layer fully-connected neural networks under certain conditions, one step of gradient descent on the first layer followed by ridge regression on the second layer can lead to feature learning; characterized by the appearance of a separated rank-one component — spike — in the spectrum of the feature matrix. However, with a constant gradient descent step size, this spike only carries information from the linear component of the target function and therefore learning non-linear components is impossible. We show that with a learning rate that grows with the sample size, such training in fact introduces multiple rank-one components, each corresponding to a specific polynomial feature. We further prove that the limiting large-dimensional and large sample training and test errors of the updated neural networks are fully characterized by these spikes. By precisely analyzing the improvement in the loss, we demonstrate that these non-linear features can enhance learning.

This talk is based on our recent work (https://arxiv.org/abs/2310.07891) with Donghwan Lee, Prof. Hamed Hassani, and Prof. Edgar Dobriban.

Wide bandgap semiconductors such as gallium oxide (Ga₂O₃) have attracted much attention for their use in next-generation high-power electronics. Although Ga₂O₃ substrates are routinely grown along various crystal orientations, the influence of such orientations on device performance has been seldom reported. In this study, I will present 2D/3D vertical diodes on β-Ga₂O_3, fabricated and optimized by varying substrate planar orientation, 2D material and electric contacts. I will discuss how the quality of our devices was validated using high-temperature dependent measurements, atomic-force microscopy (AFM) techniques and technology computer-aided design (TCAD) simulations.

Telecommunications and polarimetry both require the active control of the polarization of light. Currently, this is done by combining intrinsically anisotropic materials with tunable isotropic materials into heterostructures using complicated fabrication techniques due to the lack of scalable materials that possess both properties. Tunable birefringent and dichromic materials are scarce and rarely available in high-quality thin films over wafer scales. Recently, we reported semiconducting, highly aligned, single-walled carbon nanotubes (SWCNTs) over 4” wafers with normalized birefringence and dichroism values of 0.09 and 0.58, respectively. The real and imaginary parts of the refractive index of these SWCNT films are tuned by up to 5.9% and 14.3% in the infrared at 2200 nm and 1660 nm, respectively, using electrostatic doping. Our results suggest that aligned SWCNTs are among the most anisotropic and tunable optical materials known and opens new avenues for their application in integrated photonics and telecommunications.

In contemporary reinforcement learning, constrained policy optimization is a prominent methodology for finding optimal policies under constraints. Its applications span diverse domains, ranging from robot navigation to cancer screening. However, the challenge of oscillating policy iterates in training due to the minimax structure within Lagrangian-based policy gradient methods poses a significant hurdle. To mitigate this oscillation, this talk first introduces regularization into Lagrangian-based constrained minimax optimization and presents a regularized policy gradient-based primal-dual method with sublinear convergence of policy iterates to an optimum. Next, this talk introduces optimism from learning in games to develop an optimistic gradient-based primal-dual method. This method demonstrates remarkable linear convergence of policy iterates to an optimum, showcasing the effectiveness of optimism even in scenarios where the minimax optimization lacks convexity. Ultimately, our results shed light on the role of regularization and optimism in mitigating oscillation in constrained policy learning.

In this talk, we will first introduce two simple statistical models: high-dimensional linear regression and random features regression. Next, we will demonstrate that despite their simplicity, these models offer valuable insights into some of the most puzzling behaviors of deep neural networks, such as double descent and feature learning. Finally, we will investigate distribution shifts in random features models to demystify the recently observed phenomena of accuracy-on-the-line and agreement-on-the-line in deep neural networks.

Quantum computers promise to revolutionize nearly every aspect of society through drug development, materials discovery and optimization, and the exponential speedup of certain computational tasks. However, current devices do not have sufficient numbers of qubits (quantum bits) or high enough control fidelities (accuracies) for fault-tolerant quantum computing. Silicon quantum dot (QD) qubits have long coherence times on the order of milliseconds, small feature sizes, and are compatible with standard semiconductor fabrication technologies making them a very promising candidate for intermediate scale quantum computing. Here, we present silicon QD devices and an in-house quantum control infrastructure. This talk presents our general framework for automated single and two-qubit gate calibration necessary for advanced gate optimization protocols using a custom hardware-software stack. Demonstrating single and 2-qubit gates exceeding fault-tolerant fidelities (>99 %) in silicon quantum processors is necessary for a general-purpose quantum computator. We present a systematic approach to identify and suppress noise sources using optimized microwave pulse shapes and magnetic fields.

Future wireless networks are envisioned to be highly-complex, large-scale systems. To effectively manage the resources within these networks, it is necessary to tackle challenging, high-dimensional constrained optimization problems. Instead of relying on heuristic solutions, data-driven methods that leverage machine learning can be employed to learn superior wireless resource allocation algorithms. However, achieving autonomy in wireless networking requires the development of learning-based solutions that respect performance constraints, are able to handle the irregular data structure of wireless networks, and adapt to unexpected situations. In this talk, I present methods at the intersection of constrained optimization, graph representation learning, and network information theory, which enable wireless network management solutions that provide performance guarantees, transfer to arbitrary network sizes and topologies, adjust to unforeseen circumstances, and perform well with limited training data.

We discuss a relationship between rate-distortion and optimal transport (OT) theory, even though they seem to be unrelated at first glance. In particular, we show that a function defined via an extremal entropic OT distance is equivalent to the rate-distortion function. We numerically verify this result, as well as previous results that connect non-entropic OT to optimal scalar quantization. Thus, we unify solving scalar quantization and rate-distortion functions in an alternative fashion by using their respective optimal transport solvers.

Broadband perfect absorption in the visible (VIS) to near-infrared (NIR) range is important for solar energy harvesting, photodetection, stealth, and imaging systems. Achieving strong absorption in thin structures is necessary for lightweight devices with lower cost, flexibility, and versatility. In this talk, I’ll explain the fundamentals and challenges of light absorption in ultrathin geometries. I will further discuss the benefits of high index, high loss van der Waals dielectrics as alternatives to traditional absorber materials. Finally, I will discuss how nanophotonic effects in arrays of nanostructures made from a high index van der Waals material enable absorption of 97% of solar radiation in the VIS-NIR range.

Isoperimetry and Talagrand’s inequality for concentration of measure have been powerful tools in probability theory with widespread applications. In today’s tutorial form talk, we will explore one of the applications called the bin-packing problem, which is a classical problem in computer science known to be NP-hard. The problem is as follows: given a set of indivisible items with different sizes and a set of bins with a fixed capacity, how can we pack the items into the bins in the most efficient way possible? This problem has numerous real-world applications, including logistics, manufacturing, and resource allocation. In this tutorial, we will start by discussing the importance of the bin-packing problem and discover how Talagrand’s concentration of measure is used to obtain tighter estimates on the packing efficiency. Finally, we will discuss the implications of these results for the analysis of heuristic and approximation algorithms.

FPGA (Field-Programmable-Gate-Array) is an interesting hardware platform that has a flexibility of processors and a performance of hardware. To enable wider FPGA utilization in academia/industry, this work focuses on resolving a well-known issue: a long compilation time. In the presentation, I will illustrate the limitations of the previous approaches and how our new approach can provide a fast and flexible FPGA development environment. Even if you are not familiar with FPGA or computer architecture, please come and visit because the presentation is modified to target a wider range of audience!

In this talk, we will set up the imitation learning problem in the context of continuous control, and demonstrate how distribution shift fundamentally hinders out-of-the-box attempts at establishing provable generalization guarantees. Therefore, the key statistical learning problem we want to answer is: “when does low training error along the expert’s trajectory imply low test error on the learned trajectory?” We will then show how nailing down a notion of an expert demonstrator’s ability to recover from small errors translates to a sufficient condition for controlling the distribution shift. This naturally leads to the introduction of TaSIL: Taylor Series Imitation Learning. We will show why, in addition to standard regression on the expert’s outputs, regression on the expert’s higher order derivatives naturally pops up as a way to provably mitigate distribution shift.

Smoothness and low dimensional structures play central roles in improving generalization and stability in learning and statistics. The combination of these properties has led to many advances in semi-supervised learning, generative modeling, and control of dynamical systems. However, learning smooth functions is generally challenging, except in simple cases such as learning linear or kernel models. Typical methods are either too conservative, relying on crude upper bounds such as spectral normalization, too lax, penalizing smoothness on average, or too computationally intensive, requiring the solution of large-scale semi-definite programs. These issues are only exacerbated when trying to simultaneously exploit low dimensionality using, e.g., manifolds. This work proposes to overcome these obstacles by combining techniques from semi-infinite constrained learning and manifold regularization. To do so, it shows that, under typical conditions, the problem of learning a Lipschitz continuous function on a manifold is equivalent to a dynamically weighted manifold regularization problem. This observation leads to a practical algorithm based on a weighted Laplacian penalty whose weights are adapted using stochastic gradient techniques. We prove that, under mild conditions, this method estimates the Lipschitz constant of the solution, learning a globally smooth solution as a byproduct. Numerical examples illustrate the advantages of using this method to impose global smoothness on manifolds as opposed to imposing smoothness on average.

In this talk, I will present the routing of vehicles to various delivery points as the Travelling salesman problem (TSP), which is a widely studied NP-hard problem in combinatorial optimization. I will discuss a practical heuristic for the TSP that was implemented in a routing system for food delivery. The heuristic algorithm is based on space-filling curves — a mathematical construction that generates a path to cover a planar region. I will also skim through the analysis of how the space-filling curves and a concentration inequality demonstrate that the length of the optimal tour is approximately constant.

Many signals can be reasonably characterized by a small number of dominant modes that produce time-evolving behaviors, and identifying these modes is paramount to modeling systems simply. In this work, we present some algorithms and heuristics that enable discovery of these modes. Theoretical properties of the algorithms are explored and validation of its effectiveness is demonstrated on synthetic data. Finally, we combine our algorithm with heuristic techniques to discover latent states in neurological time series gathered from epileptic patients; despite the perceived complexity of neural systems, we demonstrate that the signals associated with seizures can be described by relatively few dominant modes.

Over the past ~40 years the internet’s hyperexponential growth was enabled by factors such as coexistence with telecommunications infrastructure, interoperability and the ability to scale. Quantum networks have shown coexistence with internet infrastructure but interoperability has not been demonstrated; scaling of geographically distributed Hilbert spaces has no digital analogue. We introduce these topics and discuss an Atomic Ensemble network under development at Stony Brook University and a Diamond Vacancy center network under development at TU-Delft.

Learning-enabled systems are emerging in many real-world applications where learning modules such as deep neural networks are deployed as feedback controllers, perception modules, or motion planners in complex dynamical systems. Providing formal safety guarantees for such systems is crucial since neural networks can behave in unexpected ways under small input perturbation and cause the closed-loop system to be unsafe. Due to the coupling between nonlinear dynamics and complex, large-scale learning modules, providing formal guarantees for learning-enabled systems is challenging and calls for the development of new theoretical and computational frameworks on the interface between machine learning, control, and optimization.

My research focuses on designing numerically efficient tools for providing formal guarantees for learning-enabled systems. In this talk, I will discuss how control-theoretic analysis and machine learning can be combined to address this challenge by solving hierarchical problems including (1) scalable verification of neural network input-output properties, (2) closed-loop reachability and stability analysis of neural network controlled systems, and (3) robust model predictive control of uncertain dynamical systems with safety guarantees. At the end of the talk, I will discuss how the research on the above three problems can complement each other and relate to broader research areas.

A fundamental question in designing lossy data compression schemes is how well one can do in comparison with the rate-distortion function, which describes the known theoretical limits of lossy compression. Motivated by the empirical success of deep neural network (DNN) compressors on large, real-world data, we investigate methods to estimate the rate-distortion function on such data, which would allow comparison of DNN compressors with optimality. While one could use the empirical distribution of the data and apply the Blahut-Arimoto algorithm, this approach presents several computational challenges and inaccuracies when the datasets are large and high-dimensional, such as the case of modern image datasets. Instead, we reformulate the rate-distortion objective, and solve the resulting functional optimization problem using neural networks. We show that the resulting rate-distortion estimator, called NERD, is a strongly consistent estimator, and provide evidence that NERD can accurately estimate the rate-distortion function on popular image datasets. Using our estimate, we show that the rate-distortion achievable by DNN compressors are within several bits of the rate-distortion function for real-world datasets. Additionally, NERD provides access to the rate-distortion achieving channel, as well as samples from its output marginal. Therefore, using recent results in reverse channel coding, we describe how NERD can be used to construct an operational one-shot lossy compression scheme with guarantees on the achievable rate and distortion. Experimental results demonstrate competitive performance with DNN compressors.

Deterministic Policy Gradient (DPG) removes a level of randomness from standard randomized-action Policy Gradient (PG), and demonstrates substantial empirical success for tackling complex dynamic problems involving Markov decision processes. At the same time, though, DPG loses its ability to learn in a model-free (i.e., actor-only) fashion, frequently necessitating the use of critics in order to obtain consistent estimates of the associated policy-reward gradient. In this work, we introduce Zeroth-order Deterministic Policy Gradient (ZDPG), which approximates policy-reward gradients via two-point stochastic evaluations of the -function, constructed by properly designed low-dimensional action-space perturbations. Exploiting the idea of random horizon rollouts for obtaining unbiased estimates of the -function, ZDPG lifts the dependence on critics and restores true model-free policy learning, while enjoying built-in and provable algorithmic stability. Additionally, we present new finite sample complexity bounds for ZDPG, which improve upon existing results by up to two orders of magnitude. Our findings are supported by several numerical experiments, which showcase the effectiveness of ZDPG in a practical setting, and its advantages over both PG and Baseline PG.

Hexagonal boron nitride (h-BN) is a van der Waals material that hosts defect-based quantum emitters (QEs) at room temperature. Recent observations suggest the existence of multiple distinct defect structures hosting QEs. Theoretical proposals suggest vacancies, their complexes and substitutional atoms as likely defect candidates. However, experimental identification of the QEs’ electronic structure is lacking, and key details of the QEs’ charge and spin properties remain unknown. In this talk, we discuss probing the optical dynamics of QEs in h-BN using photon emission statistics and photoluminescence spectroscopy with the goal of predicting the electronic level structure. We probe the optical dynamics of the QEs at various excitation powers and wavelengths and propose an electronic level structure which could give rise to the experimental observations.

Graph neural networks (GNNs) use graph convolutions to exploit network invariances and learn meaningful features from network data. However, on large-scale graphs convolutions incur in high computational cost, leading to scalability limitations. Leveraging the graphon — the limit object of a graph — in this paper we consider the problem of learning a graphon neural network (WNN) — the limit object of a GNN — by training GNNs on graphs sampled Bernoulli from the graphon. Under smoothness conditions, we show that: (i) the expected distance between the learning steps on the GNN and on the WNN decreases asymptotically with the size of the graph, and (ii) when training on a sequence of growing graphs, gradient descent follows the learning direction of the WNN. Inspired by these results, we propose a novel algorithm to learn GNNs on large-scale graphs that, starting from a moderate number of nodes, successively increases the size of the graph during training. This algorithm is benchmarked on both a recommendation system and a decentralized control problem where it is shown to retain comparable performance, to its large-scale counterpart, at a reduced computational cost.

Quantum computers promise to revolutionize nearly every aspect of society through drug development, materials discovery and optimization, and the exponential speedup of certain computational tasks. However, current devices do not have sufficient numbers of qubits (quantum bits) or high enough control fidelities (accuracies) for fault-tolerant quantum computing. Silicon quantum dot (QD) qubits have long coherence times on the order of milliseconds, small feature sizes, and are compatible with standard semiconductor fabrication technologies making them a very promising candidate for intermediate scale quantum computing. Here, we present approaches leading to high-fidelity quantum control in silicon quantum dot processors by fabricating devices resilient to charge noise and developing numerically optimized control protocols. Demonstrating single and 2-qubit gates exceeding fault-tolerant fidelities (>99 %) is necessary for general-purpose quantum computation and is currently an active area of research. Two effects that lead to infidelity are noise and decoherence, this work will address the former.  We present a systematic approach to identify and suppress noise sources using optimized microwave pulse shapes and magnetic fields. Significant work over the past decade has been dedicated to fabricating high-quality silicon devices in a scalable manner and developing proof-of-concept protocols for single and 2-qubit gates, but their control fidelities remain significantly below the fault-tolerant thresholds. This work bridges this gap to achieve universal quantum computing in silicon above the fault-tolerant threshold by developing a set of techniques for high-fidelity quantum operations using new device architectures and optimized control pulses.

FPGA-based accelerators are demonstrating significant absolute performance and energy efficiency compared with general-purpose CPUs. While FPGA computations can now be described in the standard, programming languages, like C, development for FPGA accelerators remains tedious and inaccessible to modern software engineers. Slow compiles (potentially taking tens of hours) inhibit the rapid, incremental refinement of designs that is the hallmark of modern software engineering. To address this issue, we introduce separate compilation and linkage into the FPGA design flow, providing faster design turns more familiar to software development. To realize this flow, we provide abstractions, compiler options, and compiler flow that allow the same C source code to be compiled to processor cores in seconds and to FPGA regions in minutes, providing the missing -O0 and -O1 options familiar in software development. This raises the FPGA programming level and standardizes the programming experience, bringing FPGA-based accelerators into a more familiar software platform ecosystem for software engineers.

In recent years, the intersection between proof systems and algorithmic design has received a lot of interest within the computer science community. One of these proof systems, the sums-of-square hierarchy, has been applied with great success to many long standing problems in theoretical computer science and machine learning such as learning and clustering a mixture of Gaussians, matrix and tensor completion, etc. At the core lies a relatively straightforward but powerful idea: if there exists a sums-of-squares proof of a polynomial inequality, there exists a polynomial-time algorithm to find such a proof. In this talk, we will explore the basic foundations of the sums-of-squares proof system from a user’s point of view.

Nanoparticles are emerging platforms for quantum sensing and targeted nanomedicine. Advances in integrated quantum sensors and functionalized nanoparticles have catalyzed a demand for colloidal nanoparticle devices and systems. Extraction of particle information, such as size, presence of agglomeration, and concentration, is an essential step in developing new protocols and processes for novel colloidal devices. However, existing characterization methods either require large amounts of source material, have material restrictions, or are destructive against the process being assessed. Here, we present an analytical method for evaluating change in total mass using dynamic light scattering and UV-vis measurements. We demonstrate the effectiveness of this method for dielectric and metallic nanoparticles and verify the results using ion-coupled plasma mass spectroscopy.

Many robotic tasks, like manipulation and locomotion, fundamentally include making and breaking contact with the environment. However, state-of-the-art control policies struggle to deal with the hybrid nature of multi-contact motion. Such controllers often rely heavily upon heuristics or, due to the combinatorial structure in the dynamics, are unsuitable for real-time control. A subset of controllers, such as neural network controllers, can achieve satisfactory performance but they lack guarantees. In this talk, I will present techniques for overcoming these challenges in order to design controllers and verify stability of multi-contact systems.

Breaking the reciprocity of electromagnetic interactions is of paramount importance in photonic and microwave technologies, as it enables unidirectional power flow and other unique electromagnetic phenomena. We have explored a novel method to break the reciprocity of electromagnetic guided waves utilizing an electron beam with a constant velocity. We theoretically explore the break of electromagnetic reciprocity in the presence of swift electron beams in different scenarios, including the interaction of the electron beams with guided and radiated waves, providing the possibility for strong nonreciprocal behavior.

Graph processes consist of sequences of graph signals that vary in time on top of a static graph. They can be used to model data such as climate variables on weather station networks and seismic wave readings on a network of seismographs. Our objective is to introduce a permutation equivariant Graph Recurrent Neural Network (GRNN). GRNNs learn representations of graph processes by taking both their sequential structure and the underlying graph topology into account, while also keeping the number of parameters independent of the length of the sequence and of the size of the graph. The stability of GRNNs to graph perturbations is analyzed, and their architecture is extended to include three gating strategies, which address the problem of vanishing/exploding gradients over time and across paths of the graph. The advantages of the GRNN parametrization are demonstrated in a regression and a classification experiment, where GRNNs outperform both GNNs and RNNs, and time, node and edge gating yield gains in performance in different time and spatial correlation scenarios.

Reinforcement learning, mathematically described by Markov Decision Problems, may be approached either through dynamic programming or policy search. Actor-critic algorithms combine the merits of both approaches by alternating between steps to estimate the value function and policy gradient updates. Due to the fact that the updates exhibit correlated noise and biased gradient updates, only the asymptotic behavior of actor-critic is known by connecting its behavior to dynamical systems. This work puts forth a new variant of actor-critic that employs Monte Carlo rollouts during the policy search updates, which results in controllable bias that depends on the number of critic evaluations. As a result, we are able to provide for the first time the convergence rate of actor-critic algorithms when the policy search step employs policy gradient, agnostic to the choice of policy evaluation technique. In particular, we establish conditions under which the sample complexity is comparable to stochastic gradient method for non-convex problems or slower as a result of the critic estimation error, which is the main complexity bottleneck. These results hold for in continuous state and action spaces with linear function approximation for the value function. We then specialize these conceptual results to the case where the critic is estimated by Temporal Difference, Gradient Temporal Difference, and Accelerated Gradient Temporal Difference. These learning rates are then corroborated on a navigation problem involving an obstacle, which suggests that learning more slowly may lead to improved limit points, providing insight into the interplay between optimization and generalization in reinforcement learning.

Abstract: Information processing and autonomous systems have become ubiquitous in modern life and as their societal impact increases, so does the need to curtail their behavior. Recent failures of learning-based solutions have shown that, left untethered, they are susceptible to tampering and prone to prejudiced and unsafe actions. Currently, this issue is tackled by leveraging domain expert knowledge to either construct models that embed the desired properties or tune the learning objective so as to promote them. However, the growing scale and complexity of modern information processing and autonomous systems renders this manual behavior tuning infeasible. Already today, explainability, interpretability, and transparency combined with human judgment are no longer enough to design systems that perform according to specifications. This talk therefore proposes to explicitly impose learning constraints instead. It discusses preliminary results and perspectives on the theory of constrained statistical learning that addresses the challenge of solving these statistical and often non-convex constrained problems. Infinite dimensionality and rich finite dimensional representations will be the key ingredients to tackle this issue in practical settings. This general theory can be applied to solve problems involving sparsity, nonlinear modeling, fairness in neural networks, and safe reinforcement learning.

Quantum technologies have the potential to revolutionize a wide range of fields, from computation to communication and sensing. The physical platforms for developing such technologies, however, remain in their infancy. The nitrogen-vacancy (NV) center in a nanodiamond functions as an optically addressable spin qubits with high levels of environmental sensitivity and room-temperature quantum coherence. Existing methods for synthesizing nanodiamonds yield particles that vary significantly in size and shape. The heterogeneity in nanodiamond morphology affects the quantum properties of the embedded spin qubits in ways that remain poorly understood. Accordingly, these material issues have stymied efforts to incorporate nanodiamonds in photonic structures or multiparticle assemblies. Here, we present work to assembly large area uniform arrays of a nanodiamonds. These templated assemblies provide not only a method for systematically studying nanodiamonds as a quantum material but also offer a method for the fabrication of more complex devices.

We consider the robust model predictive control (MPC) of a linear time-varying (LTV) system with norm bounded disturbances and model uncertainty, wherein a series of constrained optimal control problems (OCPs) are solved. Guaranteeing robust feasibility of these OCPs is challenging, due to both disturbances perturbing the predicted states, and model uncertainty which can render the closed-loop system unstable. As such, a trade-off between the numerical tractability and conservativeness of the solutions is often required. We use the System Level Synthesis (SLS) framework to reformulate these constrained OCPs over closed-loop system responses, and show that this allows us to transparently account for norm bounded additive disturbances and LTV model uncertainty by computing robust state feedback policies. We further show that by exploiting the underlying linear-fractional structure of the resulting robust OCPs, we can significantly reduce the conservatism of existing SLS-based robust control methods.

Abstract:  Epsilon-near-zero (ENZ) hosts doped with non-magnetic dielectric inclusions have been shown to exhibit a highly tunable magnetic permeability. In this talk, we discuss how such a peculiar magnetic response may be exploited to realize a broad range of novel functionalities, such as impedance matching, enhanced magnetic nonlinearity, nonlinear absorbers, optical bistability, tunable electric field enhancement.

Abstract: While many robotic tasks, like manipulation and locomotion, are fundamentally based in making and breaking contact with the environment, state-of-the-art control policies struggle to deal with the hybrid nature of multi-contact motion. Such controllers often rely heavily upon heuristics or, due to the combinatoric structure in the dynamics, are unsuitable for real-time control. Principled deployment of tactile sensors offers a promising mechanism for stable and robust control, but modern approaches often use this data in an ad hoc manner, for instance to guide guarded moves. In this work, by exploiting the complementarity structure of contact dynamics, we propose a control framework which can close the loop on rich, tactile sensors. Critically, this framework is non-combinatoric, enabling optimization algorithms to automatically synthesize provably stable control policies.

Abstract: In this talk, we introduce a flexible notion of safety verification for nonlinear autonomous systems by measuring how much time the system spends in given unsafe regions. We consider this problem in the particular case of nonlinear systems with a polynomial dynamics and unsafe regions described by a collection of polynomial inequalities. In this context, we can quantify the amount of time spent in the unsafe regions as the solution to an infinite-dimensional linear program (LP). We approximate the solution to the infinite-dimensional LP using a hierarchy of finite-dimensional semidefinite programs (SDPs). The solutions to the SDPs in this hierarchy provide monotonically converging upper bounds on the optimal solution to the infinite-dimensional LP.

Abstract: State-of-the-art approaches to legged robots are widely dependent on the use of models like the linear inverted pendulum (LIP) and the spring-loaded inverted pendulum (SLIP), popular because their simplicity enables a wide array of tools for planning, control, and analysis. However, they inevitably limit the ability to execute complex tasks or agile maneuvers. In this work, we aim to automatically synthesize models that remain low-dimensional but retain the capabilities of the high-dimensional system. For example, if one were to restore a small degree of complexity to LIP, SLIP, or a similar model, our approach discovers the form of that additional complexity which optimizes performance. We define a class of reduced-order models and provide an algorithm for optimization within this class. To demonstrate our method, we optimize models for walking at a range of speeds and ground inclines, for both a five-link model and the Cassie bipedal robot.

Abstract: In this talk, we are going to talk about how to realized quick FPGA computing mapping. Today’s FPGA compilation is slow because it compiles and co-optimizes the entire design in one monolithic mapping flow. This achieves high quality results but also means a long edit-compile-debug loop that slows development and limits the scope of design-space exploration. We introduce PRflow that uses partial reconfiguration and an overlay packet-switched network to separate the HLS-to-bitstream compilation problem for individual components of the FPGA design. This separation allows both the incremental compilation, where a single component can be recompiled without recompiling the entire design, and parallel compilation, where all the components are compiled in parallel. Both uses reduce the compilation time. Mapping the Rosetta Benchmarks to a Xilinx XCZU9EG, we show compilation times reduce from 42 minutes to 12 minutes (one case from 160 minutes to 18 minutes) when running on top of commercial tools from Xilinx. Using Symbiflow (Project X-Ray/Yosys/VPR), we show preliminary evidence we can further reduce most compile times under 5 minutes, with some components mapping in less than 2 minutes.

Abstract: In this talk, we introduce a flexible notion of safety verification for nonlinear autonomous systems by measuring how much time the system spends in given unsafe regions. We consider this problem in the particular case of nonlinear systems with a polynomial dynamics and unsafe regions described by a collection of polynomial inequalities. In this context, we can quantify the amount of time spent in the unsafe regions as the solution to an infinite-dimensional linear program (LP). We approximate the solution to the infinite-dimensional LP using a hierarchy of finite-dimensional semidefinite programs (SDPs). The solutions to the SDPs in this hierarchy provide monotonically converging upper bounds on the optimal solution to the infinite-dimensional LP.