arXiv subjects

Charles Lin

Publications and source records attributed to Charles Lin.

14 recordsLinked to original sources

arXiv2022

Memory-Based Model Editing at Scale

Even the largest neural networks make errors, and once-correct predictions can become invalid as the world changes. Model editors make local updates to the behavior of base (pre-trained) models to inject updated knowledge or correct undesirable behaviors. Existing model editors have shown promise, but also suffer from insufficient expressiveness: they struggle to accurately model an edit's intended scope (examples affected by the edit), leading to inaccurate predictions for test inputs loosely related to the edit, and they often fail altogether after many edits. As a higher-capacity alternative, we propose Semi-Parametric Editing with a Retrieval-Augmented Counterfactual Model (SERAC), which stores edits in an explicit memory and learns to reason over them to modulate the base model's predictions as needed. To enable more rigorous evaluation of model editors, we introduce three challenging language model editing problems based on question answering, fact-checking, and dialogue generation. We find that only SERAC achieves high performance on all three problems, consistently outperforming existing approaches to model editing by a significant margin. Code, data, and additional project information will be made available at https://sites.google.com/view/serac-editing.

cs.AI↗

arXiv2021

Field canalization using anisotropic 2D plasmonics

Optical devices capable of suppressing diffraction nature of light are of great technological importance to many nanophotonic applications. One important technique to achieve diffractionless optics is to exploit field canalization effect. However, current technological platforms based on metamaterial structures typically suffer from strict loss-confinement tradeoff, or lack dynamic reconfigurability over device operations. Here we report an integrated canalization platform that can alleviate this performance tradeoff. It is found that by leveraging material absorption of anisotropic 2D materials, the dispersion of this class of materials can flatten without increasing propagation losses and compromising confinement. The realization of such plasmon canalization can be considered using black phosphorus (BP), where topological transition from elliptic to hyperbolic curves can be induced by dynamically leveraging material absorption of BP. At the transition point, BP film can support long range, deeply subwavelength, near-diffractionless field propagation, exhibiting diffraction angle of 5.5o, propagation distance of 10{\lambda}spp, and {\lambda}spp < {\lambda}0/300.

physics.optics↗

arXiv2021

Fast Model Editing at Scale

While large pre-trained models have enabled impressive results on a variety of downstream tasks, the largest existing models still make errors, and even accurate predictions may become outdated over time. Because detecting all such failures at training time is impossible, enabling both developers and end users of such models to correct inaccurate outputs while leaving the model otherwise intact is desirable. However, the distributed, black-box nature of the representations learned by large neural networks makes producing such targeted edits difficult. If presented with only a single problematic input and new desired output, fine-tuning approaches tend to overfit; other editing algorithms are either computationally infeasible or simply ineffective when applied to very large models. To enable easy post-hoc editing at scale, we propose Model Editor Networks using Gradient Decomposition (MEND), a collection of small auxiliary editing networks that use a single desired input-output pair to make fast, local edits to a pre-trained model's behavior. MEND learns to transform the gradient obtained by standard fine-tuning, using a low-rank decomposition of the gradient to make the parameterization of this transformation tractable. MEND can be trained on a single GPU in less than a day even for 10 billion+ parameter models; once trained MEND enables rapid application of new edits to the pre-trained model. Our experiments with T5, GPT, BERT, and BART models show that MEND is the only approach to model editing that effectively edits the behavior of models with more than 10 billion parameters. Code and data available at https://sites.google.com/view/mend-editing.

cs.LG↗

arXiv2021

Integrated Photonic Functions Using Anisotropic 2D Material Structures

Plasmonic waveguides based on 2D materials, which enable the formations of guided modes confined around few-layered material, are promising plasmonic platforms for the miniaturization of photonic devices. Nonetheless, such waveguides support modes that are evanescent in the waveguide core with the majority of the fields concentrated around waveguide edges, which are different from those supported by 3D dielectric waveguides where the modal fields are of oscillatory nature and peak at the center. As a result, many photonic devices and functionalities that can be achieved within 3D dielectric waveguides based on total-internal-reflation modes cannot be realized using 2D material-based plasmonic structures. In this work, we propose and demonstrate how to leverage anisotropy in 2D materials to tailor of modal fields supported by 2D material waveguide for the first time. By regulating material absorption of the constituent 2D materials, the modal fields of these 2D modes can be tailored to localize around the waveguide center, which in turn can improve the efficiencies of coupling-based photonic functions using 2D materials, from in-plane multimode-interference couplers to out-of-plane optical radiation. Using natural anisotropic 2D materials such as black phosphorus, these pivotal functions can expand existing device capabilities that are typically achieved in 3D dielectrics but using 2D materials, thus allowing for the implementation of 2D plasmonic circuits with no need to relying on 3D layers.

physics.optics↗

arXiv2021

LASER: Learning a Latent Action Space for Efficient Reinforcement Learning

The process of learning a manipulation task depends strongly on the action space used for exploration: posed in the incorrect action space, solving a task with reinforcement learning can be drastically inefficient. Additionally, similar tasks or instances of the same task family impose latent manifold constraints on the most effective action space: the task family can be best solved with actions in a manifold of the entire action space of the robot. Combining these insights we present LASER, a method to learn latent action spaces for efficient reinforcement learning. LASER factorizes the learning problem into two sub-problems, namely action space learning and policy learning in the new action space. It leverages data from similar manipulation task instances, either from an offline expert or online during policy learning, and learns from these trajectories a mapping from the original to a latent action space. LASER is trained as a variational encoder-decoder model to map raw actions into a disentangled latent action space while maintaining action reconstruction and latent space dynamic consistency. We evaluate LASER on two contact-rich robotic tasks in simulation, and analyze the benefit of policy learning in the generated latent action space. We show improved sample efficiency compared to the original action space from better alignment of the action space to the task space, as we observe with visualizations of the learned action space manifold. Additional details: https://www.pair.toronto.edu/laser

cs.RO↗

arXiv2020

Electrically-Reconfigurable Passive and Active Circuits in a Single Plasmonic Architecture

Guided-wave plasmonic circuits are promising platforms for sensing, interconnection, and quantum applications in the sub-diffraction regime. Nonetheless, the loss-confinement trade-off remains a collective bottleneck for plasmonic-enhanced optical processes. Here, we report a unique plasmonic waveguide that can alleviate such trade-off and improve the efficiencies of plasmonic-based emission, light-matter-interaction, and detection simultaneously. Through different bias configurations, record experimental attributes such as normalized Purcell factor approaching 10^4, 10-dB amplitude modulation with <1 dB insertion loss and fJ-level switching energy, and photodetection sensitivity and internal quantum efficiency of -54 dBm and 6.4 % respectively can be realized within the same amorphous-based plasmonic structure. The ability to support multiple optoelectronic phenomena while providing performance gains over existing plasmonic and dielectric counterparts offers a clear path towards reconfigurable, monolithic plasmonic circuits.

physics.optics↗

arXiv2020

Supermode hybridization: a material-independent route towards record Schottky detection sensitivity using <0.05 {\mu}m3 amorphous absorber volume

Schottky photodetectors are attractive for CMOS-compatible photonic integrated circuits, but the trade-off between photon absorption and hot carrier emission often compromises the detection fidelity and sensitivity. Here, we report a hybridization-based waveguiding effect that can improve the sensitivity of guided-wave Schottky detection by >200x. By hybridizing the supermodes guided by asymmetrical coupled plasmonic nanostructures, light-matter-interaction can be significantly enhanced even if the thickness of the light absorbing region is only a few nanometers, thus allowing the absorption and emission efficiencies to be simultaneously optimized for the first time. Using amorphous-based active junction with absorbing volume of only 0.031 {\mu}m3, our detectors demonstrate record experimental sensitivity of -55 dBm, approaching that of crystalline-based Ge counterparts with >36x larger absorption region. The ability to maximize field overlap within small device volume and improve photodetection fidelity without using crystalline materials pave the way towards improved back-end-of-line interconnection, single photon detection, and 2D optoelectronics.

physics.optics↗

arXiv2019

The Stanford Acuity Test: A Precise Vision Test Using Bayesian Techniques and a Discovery in Human Visual Response

Chart-based visual acuity measurements are used by billions of people to diagnose and guide treatment of vision impairment. However, the ubiquitous eye exam has no mechanism for reasoning about uncertainty and as such, suffers from a well-documented reproducibility problem. In this paper we make two core contributions. First, we uncover a new parametric probabilistic model of visual acuity response based on detailed measurements of patients with eye disease. Then, we present an adaptive, digital eye exam using modern artificial intelligence techniques which substantially reduces acuity exam error over existing approaches, while also introducing the novel ability to model its own uncertainty and incorporate prior beliefs. Using standard evaluation metrics, we estimate a 74% reduction in prediction error compared to the ubiquitous chart-based eye exam and up to 67% reduction compared to the previous best digital exam. For patients with eye disease, the novel ability to finely measure acuity from home could be a crucial part in early diagnosis. We provide a web implementation of our algorithm for anyone in the world to use. The insights in this paper also provide interesting implications for the field of psychometric Item Response Theory.

cs.AI↗

arXiv2019

Ultra-high Purcell factor using long-range modes in asymmetric plasmonic waveguides

For integrated optical devices, realistic utilization of the superior wave-matter interaction offered by plasmonics is typically impeded by optical losses, which increase rapidly with mode volume reduction. Although coupled-mode plasmonic structures has demonstrated effective alleviation of the loss-confinement trade-off, stringent symmetry requirements must be enforced for such reduction to prevail. In this work, we report an asymmetric plasmonic waveguide that is not only capable of guiding subwavelength optical mode with long-range propagation, but is also unrestricted by structural, material, or modal symmetry. In these composite hybrid plasmonic waveguides (CHPWs), the versatility afforded by coupling dissimilar plasmonic structures, within the same waveguide, allow better fabrication tolerance and provide more degrees of design freedom to simultaneously optimize various device attributes. The CHPWs used to demonstrated the concept in this work, exhibit propagation loss and mode area of only 0.03 dB/{\mu}m and 0.002 {\mu}m^2 respectively, corresponding to the smallest combination amongst experimentally demonstrated long-range plasmonic structures to-date. Moreover, CHPW micro-rings were realized with record in/out coupling excitation efficiency (71%), extinction ratio (29 dB), as well as Purcell factor (1.5\times10^4).

physics.optics↗

arXiv2019

Fully Integrated On-FPGA Molecular Dynamics Simulations

The implementation of Molecular Dynamics (MD) on FPGAs has received substantial attention. Previous work, however, has consisted of either proof-of-concept implementations of components, usually the range-limited force; full systems, but with much of the work shared by the host CPU; or prototype demonstrations, e.g., using OpenCL, that neither implement a whole system nor have competitive performance. In this paper, we present what we believe to be the first full-scale FPGA-based simulation engine, and show that its performance is competitive with a GPU (running Amber in an industrial production environment). The system features on-chip particle data storage and management, short- and long-range force evaluation, as well as bonded forces, motion update, and particle migration. Other contributions of this work include exploring numerous architectural trade-offs and analysis on various mappings schemes among particles/cells and the various on-chip compute units. The potential impact is that this system promises to be the basis for long timescale Molecular Dynamics with a commodity cluster.

cs.DC↗

arXiv2017

Highly Sensitive Wavelength-scale Amorphous Hybrid Plasmonic Detectors

Hybrid integration of plasmonics and Si photonics is a promising architecture for global microprocessor interconnects. To this end, practical plasmonic devices not only should provide athermal, broadband operation over wavelength-scale footprint, but also support non-intrusive integration with low-loss Si waveguides as well as CMOS back-end-of-line processes. Here, we demonstrate a hybrid plasmonic photodetector with a single active junction fabricated via back-end deposited amorphous materials coupled to Si nanowires with only 1.5dB loss. Utilizing internal photoemission, our detectors measured sensitivity of -35dBm in a 620nm by 5{\mu}m footprint at 7V bias. Moreover, responsivity up to 0.4mA/W and dark current down to 0.2nA were obtained. The high process tolerance is demonstrated between {\lambda}=1.2-1.8{\mu}m and up to 100{\deg}C. The results suggest the potential towards plasmonic-photonic optoelectronic integration on top of Si chips without costly process modifications.

physics.app-ph↗

arXiv2017

Efficient integrated graphene photonics in the visible and near-IR

Graphene photonics has emerged as a promising platform for providing desirable optical functionality. However, graphene's monolayer-scale thickness fundamentally restricts the available light matter interaction, posing a critical design challenge for integrated devices, particularly in wavelength regimes where graphene plasmonics is untenable. While several plasmonic designs have been proposed to enhance graphene light interaction in these regimes, they suffer from substantial insertion loss due to metal absorption. Here we report a non-resonant metamaterial-based waveguide platform to overcome the design bottleneck associated with graphene device. Such metamaterial structure enables low insertion loss even though metal is being utilized. By examining waveguide dispersion characteristics via closed-form analysis, it is demonstrated that the metamaterial approach can provide optimized optical field that overlaps with the graphene monolayer. This enables graphene-based integrated components with superior optical performance. Specifically, the metamaterial-assisted graphene modulator can provide 5-fold improvement in extinction ratio compared to Si nanowire, while reducing insertion loss by one order magnitude compared to plasmonic structures. Such a waveguide configuration thus allows one to maximize the optical potential that graphene holds in the telecom and visible regimes.

physics.optics↗

arXiv2015

Polarization Engineering in Nano-Scale Waveguides Using Lossless Media

A device that achieves controllable rotation of the state of polarization by rotating the orientation of the eigenmodes of a waveguide by 45$^{\circ}$ is introduced and analyzed. The device can be implemented using lossless materials on a nanoscale and helps circumvent the inherent polarization dependence of photonic devices realized within the silicon on insulator platform. We propose and evaluate two novel polarization rotator-based schemes to achieve polarization engineering functions: (1) A multi-purpose device, with dimensions on the order of a few wavelengths which can function as a polarization splitter or an arbitrary linear polarization state generator. (2) An energy efficient optical modulator that utilizes eigenmode rotation and epsilon near zero (ENZ) effects to achieve high extinction ratio, polarization insensitive amplitude modulation without the need to sweep the device geometry to match the TE and TM mode attributes. By using indium tin oxide (ITO) as an example for a tunable material, the proposed modulator provides polarization insensitive operation and can be realized with a modulation bandwidth of 112 GHz, a length of 1800 nm an energy per bit of 7.5 $fJ$ and an optical bandwidth of 210 nm.

physics.optics↗

arXiv2015

Multilayer black phosphorus as a versatile mid-infrared electro-optic material

We investigate the electro-optic properties of black phosphorus (BP) thin films for optical modulation in the mid-infrared frequencies. Our calculation indicates that an applied out-of-plane electric field may lead to red-, blue-, or bidirectional shift in BP's absorption edge. This is due to the interplay between the field-induced quantum-confined Franz-Keldysh effect and the Pauli-blocked Burstein-Moss shift. The relative contribution of the two electro-absorption mechanisms depends on doping range, operating wavelength, and BP film thickness. For proof-of concept, simple modulator configuration with BP overlaid over a silicon nanowire is studied. Simulation result shows that operating BP in the quantum-confined Franz-Keldysh regime can enable improved maximal attainable absorption as well as power efficiency compared to its graphene counterpart.

cond-mat.mes-hall↗