arXiv subjects

Liang Tan

Publications and source records attributed to Liang Tan.

28 records · Page 2Linked to original sources

arXiv2023

Learning Easily Updated General Purpose Text Representations with Adaptable Task-Specific Prefixes

Many real-world applications require making multiple predictions from the same text. Fine-tuning a large pre-trained language model for each downstream task causes computational burdens in the inference time due to several times of forward passes. To amortize the computational cost, freezing the language model and building lightweight models for downstream tasks based on fixed text representations are common solutions. Accordingly, how to learn fixed but general text representations that can generalize well to unseen downstream tasks becomes a challenge. Previous works have shown that the generalizability of representations can be improved by fine-tuning the pre-trained language model with some source tasks in a multi-tasking way. In this work, we propose a prefix-based method to learn the fixed text representations with source tasks. We learn a task-specific prefix for each source task independently and combine them to get the final representations. Our experimental results show that prefix-based training performs better than multi-tasking training and can update the text representations at a smaller computational cost than multi-tasking training.

cs.CL↗

arXiv2023

Defending Against Patch-based Backdoor Attacks on Self-Supervised Learning

Recently, self-supervised learning (SSL) was shown to be vulnerable to patch-based data poisoning backdoor attacks. It was shown that an adversary can poison a small part of the unlabeled data so that when a victim trains an SSL model on it, the final model will have a backdoor that the adversary can exploit. This work aims to defend self-supervised learning against such attacks. We use a three-step defense pipeline, where we first train a model on the poisoned data. In the second step, our proposed defense algorithm (PatchSearch) uses the trained model to search the training data for poisoned samples and removes them from the training set. In the third step, a final model is trained on the cleaned-up training set. Our results show that PatchSearch is an effective defense. As an example, it improves a model's accuracy on images containing the trigger from 38.2% to 63.7% which is very close to the clean model's accuracy, 64.6%. Moreover, we show that PatchSearch outperforms baselines and state-of-the-art defense approaches including those using additional clean, trusted data. Our code is available at https://github.com/UCDvision/PatchSearch

cs.CV↗

arXiv2023

Database of semiconductor point-defect properties for applications in quantum technologies

Solid-state point defects are attracting increasing attention in the field of quantum information science, because their localized states can act as a spin-photon interface in devices that store and transfer quantum information, which have been used for applications in quantum computing, sensing, and networking. In this work we have performed high-throughput calculations of over 50,000 point defects in various semiconductors including diamond, silicon carbide, and silicon. Focusing on quantum applications, we characterize the relevant optical and electronic properties of these defects, including formation energies, spin characteristics, transition dipole moments, zero-phonon lines. We find 2331 composite defects which are stable in intrinsic silicon, which are then filtered to identify many new optically bright telecom spin qubit candidates and single-photon sources. All computed results and relaxed defect structures are made publicly available online at quantumdefects.com, a living database of defect characteristics which will be continually expanded with new defects and properties, and will enable researchers to select defects tailored to their applications.

quant-ph↗

arXiv2023

UNIREX: A Unified Learning Framework for Language Model Rationale Extraction

An extractive rationale explains a language model's (LM's) prediction on a given task instance by highlighting the text inputs that most influenced the prediction. Ideally, rationale extraction should be faithful (reflective of LM's actual behavior) and plausible (convincing to humans), without compromising the LM's (i.e., task model's) task performance. Although attribution algorithms and select-predict pipelines are commonly used in rationale extraction, they both rely on certain heuristics that hinder them from satisfying all three desiderata. In light of this, we propose UNIREX, a flexible learning framework that generalizes rationale extractor optimization as follows: (1) specify architecture for a learned rationale extractor; (2) select explainability objectives (i.e., faithfulness and plausibility criteria); and (3) jointly the train task model and rationale extractor on the task using the selected objectives. UNIREX enables replacing prior works' heuristic design choices with a generic learned rationale extractor in (1) and optimizing it for all three desiderata in (2)-(3). To facilitate comparison between methods with respect to multiple desiderata, we introduce the Normalized Relative Gain (NRG) metric. Across five text classification datasets, our best UNIREX configuration outperforms baselines by an average of 32.9% NRG. Plus, we find that UNIREX-trained rationale extractors can even generalize to unseen datasets and tasks.

cs.CL↗

arXiv2023

All-silicon quantum light source by embedding an atomic emissive center in a nanophotonic cavity

Silicon is the most scalable optoelectronic material, and it has revolutionized our lives in many ways. The prospect of quantum optics in silicon is an exciting avenue because it has the potential to address the scaling and integration challenges, the most pressing questions facing quantum science and technology. We report the first all-silicon quantum light source based on a single atomic emissive center embedded in a silicon-based nanophotonic cavity. We observe a more than 30-fold enhancement of luminescence, a near unity atom-cavity coupling efficiency, and an 8-fold acceleration of the emission from the quantum center. Our work opens avenues for large-scale integrated all-silicon cavity quantum electrodynamics and quantum photon interfaces with applications in quantum communication, sensing, imaging, and computing.

quant-ph↗

arXiv2022

FRAME: Evaluating Rationale-Label Consistency Metrics for Free-Text Rationales

Following how humans communicate, free-text rationales aim to use natural language to explain neural language model (LM) behavior. However, free-text rationales' unconstrained nature makes them prone to hallucination, so it is important to have metrics for free-text rationale quality. Existing free-text rationale metrics measure how consistent the rationale is with the LM's predicted label, but there is no protocol for assessing such metrics' reliability. Thus, we propose FRAME, a framework for evaluating rationale-label consistency (RLC) metrics for free-text rationales. FRAME is based on three axioms: (1) good metrics should yield highest scores for reference rationales, which maximize RLC by construction; (2) good metrics should be appropriately sensitive to semantic perturbation of rationales; and (3) good metrics should be robust to variation in the LM's task performance. Across three text classification datasets, we show that existing RLC metrics cannot satisfy all three FRAME axioms, since they are implemented via model pretraining which muddles the metric's signal. Then, we introduce a non-pretraining RLC metric that greatly outperforms baselines on (1) and (3), while performing competitively on (2). Finally, we discuss the limitations of using RLC to evaluate free-text rationales.

cs.CL↗

arXiv2021

Description of $^{93}$Nb stellar electron-capture rates by the Projected Shell Model

Capture of electrons by nuclei is an important process in stellar environments where excited nuclear states are thermally populated. However, accurate treatment for excited configurations in electron capture (EC) rates has been an unsolved problem for medium-heavy and heavy nuclei. In this work, we take the $^{93}$Nb $\rightarrow$ $^{93}$Zr EC rates as the example to introduce the Projected-Shell-Model (PSM) in which excited configurations are explicitly included as multi-quasiparticle states. Applying the prevalent assumption that the parent nucleus always stays in its ground state in stellar conditions, we critically compare the obtained PSM results with the recently-measured Gamow-Teller transition data, and with the previous calculations by the conventional shell model and the quasiparticle random-phase approximation. We discuss important ingredients that are required in theoretical models used for stellar EC calculations, and demonstrate effects of the explicit inclusion of excited nuclear states in EC rate calculations, especially when both electron density and environment temperature are high.

nucl-th↗

arXiv2021

MSD: Saliency-aware Knowledge Distillation for Multimodal Understanding

To reduce a model size but retain performance, we often rely on knowledge distillation (KD) which transfers knowledge from a large "teacher" model to a smaller "student" model. However, KD on multimodal datasets such as vision-language tasks is relatively unexplored, and digesting multimodal information is challenging since different modalities present different types of information. In this paper, we perform a large-scale empirical study to investigate the importance and effects of each modality in knowledge distillation. Furthermore, we introduce a multimodal knowledge distillation framework, modality-specific distillation (MSD), to transfer knowledge from a teacher on multimodal tasks by learning the teacher's behavior within each modality. The idea aims at mimicking a teacher's modality-specific predictions by introducing auxiliary loss terms for each modality. Furthermore, because each modality has different saliency for predictions, we define saliency scores for each modality and investigate saliency-based weighting schemes for the auxiliary losses. We further study a weight learning approach to learn the optimal weights on these loss terms. In our empirical analysis, we examine the saliency of each modality in KD, demonstrate the effectiveness of the weighting scheme in MSD, and show that it achieves better performance than KD on four multimodal datasets.

cs.CV↗

arXiv2021

Urca Cooling in Neutron Star Crusts and Oceans: Effects of Nuclear Excitations

The excited-state structure of atomic nuclei can modify nuclear processes in stellar environments. In this work, we study the influence of nuclear excitations on Urca cooling (repeated back-and-forth beta decay and electron capture in a pair of nuclear isotopes) in the crust and ocean of neutron stars. We provide for the first time an expression for Urca process neutrino luminosity which accounts for excited states of both members of an Urca pair. We use our new formula with state-of-the-art nuclear structure inputs to compute neutrino luminosities of candidate Urca cooling pairs. Our nuclear inputs consist of the latest experimental data supplemented with calculations using the projected shell model. We show that, in contrast with previous results that only consider the ground states of both nuclei in the pair, our calculated neutrino luminosities for different Urca pairs vary sensitively with the environment temperature and can be radically different from those obtained in the one transition approximation.

nucl-th↗

arXiv2020

Ground-truth resting-state signal provides data-driven estimation and correction for scanner distortion of fMRI time-series dynamics

The fMRI community has made great strides in decoupling neuronal activity from other physiologically induced T2* changes, using sensors that provide a ground-truth with respect to cardiac, respiratory, and head movement dynamics. However, blood oxygenation level-dependent (BOLD) time-series dynamics are confounded by scanner artifacts, in complex ways that can vary not only between scanners but even, for the same scanner, between sessions. The lack of equivalent ground truth has thus far stymied the development of reliable methods for identification and removal of scanner-induced noise. To address this problem, we first designed and built a phantom capable of providing dynamic signals equivalent to that of the resting-state brain. Using the dynamic phantom, we quantified voxel-wise noise by comparing the ground-truth time-series with its measured fMRI data. We derived the following data-quality metrics: Standardized Signal-to-Noise Ratio (ST-SNR) and Dynamic Fidelity that can be directly compared across scanners. Dynamic phantom data acquired from four scanners showed scanner-instability multiplicative noise contributions of about 6-18% of the total noise. We further measured strong non-linearity in the fMRI response for all scanners, ranging between 8-19% of total voxels. To correct scanner distortion of fMRI time-series dynamics at a single-subject level, we trained a convolutional neural network (CNN) on paired sets of measured vs. ground-truth data. Tests on dynamic phantom time-series showed a 4- to 7-fold increase in ST-SNR and about 40-70% increase in Dynamic Fidelity after denoising. Critically, we observed that the CNN temporal denoising pushes ST-SNR > 1. Denoising human-data with ground-truth-trained CNN showed markedly increased detection sensitivity of resting-state networks.

physics.med-ph↗