## Type of Work

Article (322) Book (0) Theses (7) Multimedia (0)

## Peer Review

Peer-reviewed only (299)

## Supplemental Material

Video (0) Audio (0) Images (0) Zip (0) Other files (1)

## Publication Year

## Campus

UC Berkeley (27) UC Davis (125) UC Irvine (26) UCLA (87) UC Merced (44) UC Riverside (200) UC San Diego (214) UCSF (29) UC Santa Barbara (93) UC Santa Cruz (11) UC Office of the President (17) Lawrence Berkeley National Laboratory (221) UC Agriculture & Natural Resources (1)

## Department

Physical Sciences (211) Department of Physics (198) School of Medicine (191) Research Grants Program Office (17) Computing Sciences (11) Department of Computer Science & Engineering (7)

## Journal

Proceedings of the Vertebrate Pest Conference (1) UC Merced Undergraduate Research Journal (1)

## Discipline

Physical Sciences and Mathematics (5) Engineering (1) Life Sciences (1) Medicine and Health Sciences (1) Social and Behavioral Sciences (1)

## Reuse License

BY - Attribution required (24) BY-NC - Attribution; NonCommercial use only (1)

## Scholarly Works (330 results)

The increasing computational demands of Machine Learning (ML) models, coupled with a slowdown in hardware advancements, have led to a significant compute supply-demand gap. This gap is evident in the rising costs and limited availability of resources needed for training complex ML models like GPT-4. These challenges hinder the progress and accessibility of ML.

This dissertation aims to bridge the compute supply-demand gap by improving resource efficiency of ML. We introduce Ekya, Cilantro, and ESCHER, three new systems and methods for improving resource efficiency at different layer in the ML stack. Ekya, at the ML application layer, implements a Thief Scheduling algorithm and a Microprofiler to intelligently redistribute resources between inference and retraining tasks, thereby making continuous learning four times more resource-efficient. Cilantro, in the cluster management layer, utilizes online learning to develop dynamic resource-performance models, enabling performance-aware resource allocation in multi-tenant environments. At the orchestration layer, ESCHER introduces ephemeral resources, allowing ML applications to specify custom scheduling requirements without overhauling the underlying cluster manager. This unique approach provides applications with the flexibility to adapt to evolving needs while maintaining simplicity in system design. Together, these systems represent a comprehensive approach to mitigating the compute supply-demand gap, contributing sustainable and efficient resource management techniques.

Binary Conic Quadratic Knapsack set is the lower level set of the conic quadratic set functions. They are natural generalizations of linear knapsack sets, and have several applications in several areas, such as, combinatorial optimization, finance, and optimal control. In particular, conic quadratic knapsacks can be used to model the 0-1 linear knapsack sets with uncertain coefficients. In addition to being of theoretical interest, these problems are practically relevant as they can be used to mathematically formulate probabilistic and robust equivalents of the deterministic combinatorial and decision problems.

Although the non-linear binary sets, specifically the quadratic knapsack sets have been studied in the literature, the specific combinatorial structure associated with these sets remains to be explored. Non-linear binary sets involving bilinear terms are at present solved using a combination of lift and project relaxations and a branch-and-bound scheme that solves continuous non-linear relaxations at the nodes of a branch-and-bound search tree. The branch-and-cut methods developed for general integer conic quadratic sets make use of the problem's geometrical structure for removing fractional solutions of conic relaxations can be employed to solve the particular case of 0-1 conic quadratic knapsacks, however these approaches do not utilize the additional combinatorial structure specific to these sets. Motivated by the performance improvement observed by exploring geometrical structure for conic mixed integer programs, and the fact that exploring combinatorial structure has proven extremely useful in addressing the linear $0-1$ knapsack sets, we expect this to work well in the case of conic quadratic knapsacks as well.

In this dissertation, we study pure-binary programs with conic quadratic constraints and develop branch-and-cut algorithms to solve them with applications to robust network design problem. First, we study the combinatorial structure embedded into these problems with assumptions of monotonicity and develop valid inequalities for these problems. In Chapter 2, we consider a more general version of the problem without any monotonicity assumptions on the conic constraint, and derive valid inequalities linear in the space of the original variables. These cuts generalize a well-known class of linear cuts for binary knapsacks, and turn out to be very effective in reducing the computational effort involved in solving some practical problems.

In Chapter 3, we propose a further generalization of the problem without any structural assumptions of the constraint in context and study the binary quadratically constrained set. We show that our results generalize several known results for the 0-1 non-linear constrained sets. We take a detour from combinatorial discussion and develop a geometrical understanding for 0-1 quadratic problems in Chapter 4. We develop convexifications for the non-convex quadratic sets defined over a hypercube and provide strengthening procedures for the same.

Finally, in Chapter 5, we study the problem of robust network design with uncertainties on the arc capacities. We formulate the robust network design problem as a 0-1 conic quadratic program without particular assumptions on the characteristics of uncertainties. We consider the scenarios when the uncertainties on the arc capacities are independent and correlated. We show that the inequalities derived prove to be useful in our computational experiments.

Dual-polarized weather radars are gaining popularity due to their promise of accurate and faster weather prediction. This work presents the design of a dual-polarized, patch antenna element operating in the band 2.7 GHz - 3.0 GHz, with 30 dB isolation between the ports, which can be utilized for a dual-pol weather radar array. To characterize current design for weather radar, recently published parameters called W-parameters, have been evaluated for the demonstrated antenna hardware. Certain other properties of these W-parameters have also been studied. In the process of reaching a low cross-pol design, basic mechanism of cross-polar radiation in rectangular patch antennas has also been analyzed using a novel strategy of near-field analysis. This near field analysis has been further applied on slotted antennas to understand their radiation properties. New strategy of understanding the radiation properties based on the near-field provides visualization based understanding of the radiation mechanism in small antennas.

Magnetothermal Electrical devices are the prominent systems to generate electrical energy from heat source having small temperature gradient. The spin-Seebeck effect mediated thermoelectric energy conversion can provide an efficient alternative to traditional thermoelectric for waste heat recovery. To achieve this goal, efficient spin to charge conversion using earth-abundant materials is essential. Also, the thermal spin current from the spin-Seebeck effect has been reported to be more energy efficient than the electrical spin injection methods. However, spin detection has been the one of the bottlenecks since metals with large spin-orbit coupling is an essential requirement. Silicon is widely used in semiconductor electronics due to its abundance and versatility but having a centrosymmetric crystal structure, has insignificant intrinsic spin-orbit coupling, leading to negligible spin-charge conversion. However, in silicon, strain engineering mediated Rashba spin orbit coupling can induce interfacial spin to charge conversion arises due to an electric potential perpendicular to the interface.

The electric potential can be artificially induced, for example, using ferroelectric and piezoelectric thin films at the interface. An alternate way to induce the electric potential could be flexoelectric field. The flexoelectricity can be observed in all the material that either have or lack inversion symmetry, additionally no large gate bias is needed. Hence, the interfacial asymmetry in conjunction with strain engineering can provide an alternate pathway to achieving efficient and controllable spin-to-charge conversion for spintronics applications. In this experimental study, we report large spin to charge conversion (spin-Hall angle- 0.578) at Ni80Fe20/amorphous-Si interfaces attributed to flexoelectricity mediated Rashba spin-orbit coupling. The flexoelectricity at the interface also gave rise to flexoelectric mediated interlayer coupling. In addition to spin-charge conversion, the strained interfaces also led to almost three-fold increase in anomalous Nernst effect. This strain engineering for spin dependent thermoelectric behavior at room temperature opens a new window to the realization of spintronics and spin-caloritronics devices.