arXiv subjects

Kan Chen

Publications and source records attributed to Kan Chen.

At least 37 records · Page 2Linked to original sources

arXiv2024

Mind the kinematics simulation of planet-disk interactions: time evolution and numerical resolution

Planet-disk interactions can produce kinematic signatures in protoplanetary disks. While recent observations have detected non-Keplerian gas motions in disks, their origins are still being debated. To explore this, we conduct 3D hydrodynamic simulations using the code FARGO3D to study non-axisymmetric kinematic perturbations at 2 scale heights induced by Jovian planets in protoplanetary disks, followed by examinations of detectable signals in synthetic CO emission line observations at millimeter wavelengths. We advocate for using residual velocity or channel maps, generated by subtracting an azimuthally averaged background of the disk, to identify planet-induced kinematic perturbations. We investigate the effects of two basic simulation parameters, simulation duration and numerical resolution, on the simulation results. Our findings suggest that a short simulation (e.g., 100 orbits) is insufficient to establish a steady velocity pattern given our chosen viscosity ($α=10^{-3}$), and displays plenty of fluctuations on orbital timescale. Such transient features could be detected in observations. By contrast, a long simulation (e.g., 1,000 orbits) is required to reach steady state in kinematic structures. At 1,000 orbits, the strongest and detectable velocity structures are found in the spiral wakes close to the planet. Through numerical convergence tests, we find hydrodynamics results converge in spiral regions at a resolution of 14 cells per disk scale height (CPH) or higher. Meanwhile, synthetic observations produced from hydrodynamic simulations at different resolutions are indistinguishable with 0.1$^{\prime\prime}$ angular resolution and 10 hours of integration time on ALMA.

astro-ph.EP↗

arXiv2024

Real-Time Interactions Between Human Controllers and Remote Devices in Metaverse

Supporting real-time interactions between human controllers and remote devices remains a challenging goal in the Metaverse due to the stringent requirements on computing workload, communication throughput, and round-trip latency. In this paper, we establish a novel framework for real-time interactions through the virtual models in the Metaverse. Specifically, we jointly predict the motion of the human controller for 1) proactive rendering in the Metaverse and 2) generating control commands to the real-world remote device in advance. The virtual model is decoupled into two components for rendering and control, respectively. To dynamically adjust the prediction horizons for rendering and control, we develop a two-step human-in-the-loop continuous reinforcement learning approach and use an expert policy to improve the training efficiency. An experimental prototype is built to verify our algorithm with different communication latencies. Compared with the baseline policy without prediction, our proposed method can reduce 1) the Motion-To-Photon (MTP) latency between human motion and rendering feedback and 2) the root mean squared error (RMSE) between human motion and real-world remote devices significantly.

cs.RO↗

arXiv2024

From the $P^{N}_ψ$/$P^Λ_{ψs}$ to $\bar{T}^f_{cc}$: symmetry analysis to the interactions of the $(\bar{c}q)(\bar{c}q)$/$(ccq)(\bar{c}q)$/$(ccq)(ccq)$ di-hadron systems

We investigate the interactions of the $(\bar{c}q)(\bar{c}q)$/$(ccq)(\bar{c}q)$/$(ccq)(ccq)$ di-hadron systems based on a contact lagrangian possessing the SU(3) flavor and SU(2) spin symmetries. Under the assumptions of two scenarios for the $J^P$ quantum numbers of the $P_ψ^N(4440)$ and $P_ψ^N(4457)$ states, we obtain the parameters ($\tilde{g}_s$, $\tilde{g}_a$) introduced from this contact lagrangian. Then we include the SU(3) breaking effect by introducing a factor $g_x$, this quantity can be further constrained by the experimental mass of the $P_{ψs}^Λ(4338)$ state. We can reproduce the mass of the $T^f_{cc}(3875)$ state with the parameters extracted from the observed $P_ψ^N$ states, this consistency indicates a unified description of the di-hadron molecular states composed of two heavy-light hadrons. With the same parameters, we discuss the possible mass spectra of the $\bar{T}_{cc}^f$/$P_{ψc}^Λ$/$H_{Ω_{ccc}c}^Λ$ systems. Then we proceed to discuss the existences of the $\bar{T}_{cc\bar{s}}^θ$/$P_{ψcs}^N$/$H_{Ω_{ccc}cs}^N$ states by investigating the SU(3) breaking effects. Our results show that the states in the $\bar{T}_{cc\bar{s}}^θ$/$P_{ψcs}^N$ systems can hardly form bound states, while the states in the $H_{Ω_{ccc}cs}^N$ system can form bound states due to their larger reduced masses.

hep-ph↗

arXiv2024

Spectrum of the molecular hexaquarks

We investigate the mass spectra of molecular-type hexaquark states in the dibaryon systems. These systems are composed of the charmed baryons $[Σ_c^{(\ast)}$, $Ξ_c^{(\prime,\ast)}]$, doubly charmed baryons $[Ξ_{cc}^{(\ast)}]$, and hyperons $[Σ^{(\ast)}$, $Ξ^{(\ast)}]$. We consider all possible combinations of particle-particle and particle-antiparticle pairs, including the S-wave spin multiplets in each combination. We establish the underlying connections among the molecular tetraquarks, pentaquarks, and hexaquarks with the effective quark-level interactions. We find that the existence of molecular states in $DD^\ast$, $D\bar{D}^\ast$, and $Σ_c\bar{D}^{(\ast)}$ systems leads to the emergence of a large number of deuteron-like hexaquarks in the heavy flavor sectors. Currently, there have been several experimental candidates for molecular tetraquarks and pentaquarks. The experimental search for near-threshold hexaquarks will further advance the establishment of the underlying dynamical picture of hadronic molecules and deepen our understanding of the role of spin-flavor symmetry in near-threshold residual strong interactions.

hep-ph↗

arXiv2024

DISCRET: Synthesizing Faithful Explanations For Treatment Effect Estimation

Designing faithful yet accurate AI models is challenging, particularly in the field of individual treatment effect estimation (ITE). ITE prediction models deployed in critical settings such as healthcare should ideally be (i) accurate, and (ii) provide faithful explanations. However, current solutions are inadequate: state-of-the-art black-box models do not supply explanations, post-hoc explainers for black-box models lack faithfulness guarantees, and self-interpretable models greatly compromise accuracy. To address these issues, we propose DISCRET, a self-interpretable ITE framework that synthesizes faithful, rule-based explanations for each sample. A key insight behind DISCRET is that explanations can serve dually as database queries to identify similar subgroups of samples. We provide a novel RL algorithm to efficiently synthesize these explanations from a large search space. We evaluate DISCRET on diverse tasks involving tabular, image, and text data. DISCRET outperforms the best self-interpretable models and has accuracy comparable to the best black-box models while providing faithful explanations. DISCRET is available at https://github.com/wuyinjun-1993/DISCRET-ICML2024.

cs.LG↗

arXiv2024

MoST: Multi-modality Scene Tokenization for Motion Prediction

Many existing motion prediction approaches rely on symbolic perception outputs to generate agent trajectories, such as bounding boxes, road graph information and traffic lights. This symbolic representation is a high-level abstraction of the real world, which may render the motion prediction model vulnerable to perception errors (e.g., failures in detecting open-vocabulary obstacles) while missing salient information from the scene context (e.g., poor road conditions). An alternative paradigm is end-to-end learning from raw sensors. However, this approach suffers from the lack of interpretability and requires significantly more training resources. In this work, we propose tokenizing the visual world into a compact set of scene elements and then leveraging pre-trained image foundation models and LiDAR neural networks to encode all the scene elements in an open-vocabulary manner. The image foundation model enables our scene tokens to encode the general knowledge of the open world while the LiDAR neural network encodes geometry information. Our proposed representation can efficiently encode the multi-frame multi-modality observations with a few hundred tokens and is compatible with most transformer-based architectures. To evaluate our method, we have augmented Waymo Open Motion Dataset with camera embeddings. Experiments over Waymo Open Motion Dataset show that our approach leads to significant performance improvements over the state-of-the-art.

cs.CV↗

arXiv2024

Darkverse -- A New DarkWeb?

The "Darkverse" could be the negative harmful area of the Metaverse; a new virtual immersive environment for the facilitation of illicit activity such as misinformation, fraud, harassment, and illegal marketplaces. This paper explores the potential for inappropriate activities within the Metaverse, and the similarities between the Darkverse and the Dark Web. Challenges and future directions for investigation are also discussed, including user identification, creation of privacy-preserving frameworks and other data monitoring methods.

cs.HC↗

arXiv2024

Flavor-spin symmetry of the $P^N_ψ/H_{Ω_{ccc}}^N$ and $P^Λ_{ψs}/H^Λ_{Ω_{ccc}s}$ molecular states

Based on a contact lagrangian that incorporates the SU(3) flavor and SU(2) spin symmetries, we discuss the symmetry properties of the interactions among the heavy flavor meson-baryon $P_ψ^N$, $P_{ψs}^Λ$ (with quark components [$n\bar{c}$][$nnc$], [$s\bar{c}$][$nnc$], or [$n\bar{c}$][$nsc$]) systems and di-baryon $H_{Ω_{ccc}}^N$, $H^Λ_{Ω_{ccc}s}$ (with quark components [$nnc$][$ncc$], [$nnc$][$scc$] or [$nsc$][$ncc$]) systems ($n=u$, $d$). The light quark components of the $P_ψ^N$ ($P_{ψs}^Λ$) and $H_{Ω_{ccc}}^N$ ($H_{Ω_{ccc}s}^Λ$) systems have identical flavors, the interactions generated from the exchanges of light mesons in the $P_ψ^N$ ($P^Λ_{ψs}$) systems should be very similar to that of the $H_{Ω_{ccc}}^N$ ($H^Λ_{Ω_{ccc}s}$) systems. We perform the single-channel and multi-channel calculations on the $P_ψ^N/P^Λ_{ψs}/H_{Ω_{ccc}}^N/H^Λ_{Ω_{ccc}s}$ systems and introduce the SU(3) breaking effect to identify the different mass spectra among the $P_ψ^N$ ($H_{Ω_{ccc}}^N$) and $P^Λ_{ψs}$ ($H^Λ_{Ω_{ccc}s}$) systems. We suggest two kinds of evidences for the existence of the flavor-spin symmetry among the heavy flavor $P_ψ^N/H_{Ω_{ccc}}^N/P^Λ_{ψs}/H^Λ_{Ω_{ccc}s}$ molecule community, i.e., the mass arrangements of the $P_ψ^N/H_{Ω_{ccc}}^N/P^Λ_{ψs}/H^Λ_{Ω_{ccc}s}$ mass spectra and the binding energies of the heavy flavor meson-baryon (di-baryon) systems attributed to the same contact potentials.

hep-ph↗

arXiv2024

Spectrum of the molecular pentaquarks

We investigate the mass spectrum of the molecular pentaquarks composed of a baryon and a meson. We establish the underlying relations among the near-threshold interactions of the molecular tetraquark and pentaquark systems. We find the existence of the molecule candidates in the $Σ_c\bar{D}^{(\ast)}$, $DD^\ast$, and $D\bar{D}^\ast$ systems indicates a substantial presence of the hadronic molecules in the {\it heavy} baryon plus {\it heavy} meson systems ({\it heavy} refers to the hadrons with the $c$ and/or $s$ quarks). We make an exhaustive prediction of the possible bound/virtual molecular states in the systems: $Σ_{c}^{(\ast)}\bar{D}^{(\ast)}$, $Σ_{c}^{(\ast)}D^{(\ast)}$, $Ξ_{c}^{(\prime,\ast)}\bar{D}^{(\ast)}$, $Ξ_{c}^{(\prime,\ast)}D^{(\ast)}$, $Σ_{c}^{(\ast)}K^{\ast}$, $Σ_{c}^{(\ast)}\bar{K}^{\ast}$, $Ξ_{c}^{(\prime,\ast)}K^{\ast}$, $Ξ_{c}^{(\prime,\ast)}\bar{K}^{\ast}$, $Ξ_{cc}^{(\ast)}\bar{D}^{(\ast)}$, $Ξ_{cc}^{(\ast)}D^{(\ast)}$, $Ξ_{cc}^{(\ast)}K^{\ast}$, $Ξ_{cc}^{(\ast)}\bar{K}^{\ast}$, $Σ^{(\ast)}\bar{D}^{(\ast)}$, $Σ^{(\ast)}D^{(\ast)}$, $Ξ^{(\ast)}\bar{D}^{(\ast)}$, $Ξ^{(\ast)}D^{(\ast)}$, $Σ^{(\ast)} K^\ast$, $Σ^{(\ast)} \bar{K}^\ast$, $Ξ^{(\ast)} K^\ast$, $Ξ^{(\ast)} \bar{K}^\ast$. Hunting for the predicted states in experiments will significantly deepen our understanding of the formation mechanism of the hadronic molecules, and shed light on the manifestation of flavor symmetry in the low-energy residual strong interactions.

hep-ph↗

arXiv2024

Structure of the $\mathbf{Λ(1670)}$ resonance

We examine the internal structure of the $Λ(1670)$ through an analysis of lattice QCD simulations and experimental data within Hamiltonian effective field theory. Two scenarios are presented. The first describes the $Λ(1670)$ as a bare three-quark basis state, which mixes with the $πΣ$, $\bar{K}N$, $ηΛ$ and $KΞ$ meson-baryon channels. In the second scenario, the $Λ(1670)$ is dynamically generated from these isospin-0 coupled channels. The $K^-p$ scattering data and the pole structures of the $Λ(1405)$ and the $Λ(1670)$ can be simultaneously described well in both scenarios. However, a comparison of the finite-volume spectra to lattice QCD calculations reveals significant differences between these scenarios, with a clear preference for the first case. Thus the lattice QCD results play a crucial role in allowing us to distinguish between these two scenarios for the internal structure of the $Λ(1670)$.

hep-ph↗

arXiv2024

WOMD-LiDAR: Raw Sensor Dataset Benchmark for Motion Forecasting

Widely adopted motion forecasting datasets substitute the observed sensory inputs with higher-level abstractions such as 3D boxes and polylines. These sparse shapes are inferred through annotating the original scenes with perception systems' predictions. Such intermediate representations tie the quality of the motion forecasting models to the performance of computer vision models. Moreover, the human-designed explicit interfaces between perception and motion forecasting typically pass only a subset of the semantic information present in the original sensory input. To study the effect of these modular approaches, design new paradigms that mitigate these limitations, and accelerate the development of end-to-end motion forecasting models, we augment the Waymo Open Motion Dataset (WOMD) with large-scale, high-quality, diverse LiDAR data for the motion forecasting task. The new augmented dataset WOMD-LiDAR consists of over 100,000 scenes that each spans 20 seconds, consisting of well-synchronized and calibrated high quality LiDAR point clouds captured across a range of urban and suburban geographies (https://waymo.com/open/data/motion/). Compared to Waymo Open Dataset (WOD), WOMD-LiDAR dataset contains 100x more scenes. Furthermore, we integrate the LiDAR data into the motion forecasting model training and provide a strong baseline. Experiments show that the LiDAR data brings improvement in the motion forecasting task. We hope that WOMD-LiDAR will provide new opportunities for boosting end-to-end motion forecasting models.

cs.CV↗

arXiv2023

Spectrum of the molecular tetraquarks: Unraveling the $T_{cs0}(2900)$ and $T_{c\bar{s}0}^a(2900)$

We relate the interactions of the $\bar{D}^{(\ast)} K^\ast$ and $D^{(\ast)} K^\ast$ systems to those of $D^{(\ast)}D^{(\ast)}$ and $D^{(\ast)}\bar{D}^{(\ast)}$ respectively, considering the residual strong interactions at the near-threshold energy is too weak to excite the strange quarks inside the hadrons. We propose an effective model to describe the low-energy S-wave interactions that are undertaken by the light $u$, $d$ quarks between two separated heavy hadrons. We find that the existence of molecules in the heavy-(anti)heavy sectors will naturally lead to the emergence of molecular states in $\bar{D}^{(\ast)} K^\ast$ and $D^{(\ast)} K^\ast$ systems. The recently observed $T_{cs0}(2900)$ and $T_{c\bar{s}0}^a(2900)$ can be well identified as the $0(0^+)$ and $1(0^+)$ partners of $T_{cc}(3875)$ and $Z_c(3900)$ in the charmed strange sector, respectively. We also predict their members under the {\it heavy} ($c$ and $s$) quark symmetry and SU(2) flavor symmetry. Most of them are very good molecule candidates, for example, (i) the $0(1^+)$ states in $D^\ast D^\ast$, $\bar{D}K^\ast$, $\bar{D}^\ast K^\ast$; (ii) the $0^{(+)}(2^{+(+)})$ states in $D^\ast \bar{D}^\ast$, $\bar{D}^\ast K^\ast$, $D^\ast K^\ast$; (iii) the $1^-(0^{++})$ state in $D^\ast\bar{D}^\ast$ and $1(1^+)$ state in $D^\ast K^\ast$. The $0^+(0^{++})$ state in $D\bar{D}$ and the $0(1^+)$ state in $DK^\ast$ might also exist as virtual states, and the $0(1^+)$ $DK^\ast$ can serve as a key to infer the existence of $0^+(0^{++})$ $D\bar{D}$. The $D_sπ$ invariant mass spectrum of $T_{c\bar{s}0}^a(2900)$ is also studied within the coupled-channel approach, and the molecular interpretation of $T_{c\bar{s}0}^a(2900)$ is consistent with the experimental data. Searching for the predicted states in experiments is crucial to discriminate the different pictures for interpreting these near-threshold exotica.

hep-ph↗

arXiv2023

A General Framework for Portfolio Construction Based on Generative Models of Asset Returns

In this paper, we present an integrated approach to portfolio construction and optimization, leveraging high-performance computing capabilities. We first explore diverse pairings of generative model forecasts and objective functions used for portfolio optimization, which are evaluated using performance-attribution models based on LASSO. We illustrate our approach using extensive simulations of crypto-currency portfolios, and we show that the portfolios constructed using the vine-copula generative model and the Sharpe-ratio objective function consistently outperform. To accommodate a wide array of investment strategies, we further investigate portfolio blending and propose a general framework for evaluating and combining investment strategies. We employ an extension of the multi-armed bandit framework and use value models and policy models to construct eclectic blended portfolios based on past performance. We consider similarity and optimality measures for value models and employ probability-matching ("blending") and a greedy algorithm ("switching") for policy models. The eclectic portfolios are also evaluated using LASSO models. We show that the value model utilizing cosine similarity and logit optimality consistently delivers robust superior performances. The extent of outperformance by eclectic portfolios over their benchmarks significantly surpasses that achieved by individual generative model-based portfolios over their respective benchmarks.

q-fin.PM↗

arXiv2023

SU(3) breaking effect in the $Z_c$ and $Z_{cs}$ states

Based on the hadronic resonance picture, we propose a possible framework to simultaneously describe the resonance parameters of the observed $Z_{cs}$ and $Z_{c}$ states that are close to the thresholds of the $D^{(*)}\bar{D}^{(*)}_s$ and $D^{(*)}\bar{D}^{(*)}$ systems, respectively. We construct the effective potentials of the $Z_{cs}$ and $Z_{c}$ states by analogy with the effective potentials of the leading order (LO) and next-to-leading-order (NLO) $N-\bar{N}$ interactions. Then we introduce an SU(3) breaking factor $g_x$ to identify the differences between the effective potentials in the $Z_{cs}$ and $Z_{c}$ states. We perform two calculations to discuss the differences and similarities of the $Z_{cs}$ and $Z_{c}$ states. In the first calculation, we adopt the LECs extracted from the experimental $Z_{cs}$ states to calculate the $Z_c$ states, and show that if the $Z_{cs}(4000)$ is the SU(3) partner of the $Z_c(3900)$, then our framework can reproduce the large width difference between the $Z_{cs}(4000)$ and $Z_{c}(3900)$ by adjusting the SU(3) breaking factor in a reasonable region. Besides, this SU(3) breaking effect also accounts for the absence of a $Z_{c}$ state with $J^{PC}=1^{++}$, which should be the SU(3) partner of the $Z_{cs}(3985)$. In the second calculation, we separately fit the LECs from the experimental $Z_{cs}$ and $Z_{c}$ states. We show that these two sets of LECs are very similar to each other, indicating a unified set of LECs that could describe the effective potentials of the $Z_{cs}$ and $Z_{c}$ states simultaneously. Then we proceed to systematically predict the other possible $Z_{cs}$ and $Z_{c}$ states that are close to the $D^{(*)}\bar{D}^{(*)}_s$ and $D^{(*)}\bar{D}^{(*)}$ systems, respectively. Further explorations on the $Z_{cs}$ states would be crucial to test our theory.

hep-ph↗

arXiv2023

Planet Gap-opening Feedback on Disk Thermal Structure and Composition

(Exo-)planets inherit their budget of chemical elements from a protoplanetary disk. The disk temperature determines the phase of each chemical species, which sets the composition of solids and gas available for planet formation. We investigate how gap structures, which are widely seen by recent disk observations, alter the thermal and chemical structure of a disk. Planet-disk interaction is a leading hypothesis of gap formation and so such changes could present a feedback that planets have on planet-forming material. Both the planet gap-opening process and the disk thermal structure are well studied individually, but how the gap-opening process affects disk thermal structure evolution remains an open question. We develop a new modelling method by iterating hydrodynamical and radiative transfer simulations to explore the gap-opening feedback on disk thermal structure. We carry out parameter studies by considering different planet locations rp and planet masses Mp. We find that for the same rp and Mp, our iteration method predicts a wider and deeper gap than the non-iteration method. We also find that the inner disk and gap temperature from the iteration method can vary strongly from the non-iteration or disk without planets, which can further influence dust-trap conditions, iceline locations, and distribution of various ices, such as H2O, CO2, and CO on large dust grains ("pebbles"). Through that, a gap-opening planet can complicate the canonical picture of the non-planet disk C/O ratio and influence the composition of the next generation of planetesimals and planets.

astro-ph.EP↗

arXiv2023

Task-Oriented Cross-System Design for Timely and Accurate Modeling in the Metaverse

In this paper, we establish a task-oriented cross-system design framework to minimize the required packet rate for timely and accurate modeling of a real-world robotic arm in the Metaverse, where sensing, communication, prediction, control, and rendering are considered. To optimize a scheduling policy and prediction horizons, we design a Constraint Proximal Policy Optimization(C-PPO) algorithm by integrating domain knowledge from relevant systems into the advanced reinforcement learning algorithm, Proximal Policy Optimization(PPO). Specifically, the Jacobian matrix for analyzing the motion of the robotic arm is included in the state of the C-PPO algorithm, and the Conditional Value-at-Risk(CVaR) of the state-value function characterizing the long-term modeling error is adopted in the constraint. Besides, the policy is represented by a two-branch neural network determining the scheduling policy and the prediction horizons, respectively. To evaluate our algorithm, we build a prototype including a real-world robotic arm and its digital model in the Metaverse. The experimental results indicate that domain knowledge helps to reduce the convergence time and the required packet rate by up to 50%, and the cross-system design framework outperforms a baseline framework in terms of the required packet rate and the tail distribution of the modeling error.

cs.RO↗

arXiv2023

Combining Broad and Narrow Case Definitions in Matched Case-Control Studies: Firearms in the Home and Suicide Risk

Does having firearms in the home increase suicide risk? To test this hypothesis, a matched case-control study can be performed, in which suicide case subjects are compared to living controls who are similar in observed covariates in terms of their retrospective exposure to firearms at home. In this application, cases can be defined using a broad case definition (suicide) or a narrow case definition (suicide occurred at home). The broad case definition offers a larger number of cases but the narrow case definition may offer a larger effect size. Moreover, restricting to the narrow case definition may introduce selection bias (i.e., bias due to selecting samples based on characteristics affected by the treatment) because exposure to firearms in the home may affect the location of suicide and thus the type of a case a subject is. We propose a new sensitivity analysis framework for combining broad and narrow case definitions in matched case-control studies, that considers the unmeasured confounding bias and selection bias simultaneously. We develop a valid randomization-based testing procedure using only the narrow case matched sets when the effect of the unmeasured confounder on receiving treatment and the effect of the treatment on case definition among the always-cases are controlled by sensitivity parameters. We then use the Bonferroni method to combine the testing procedures using the broad and narrow case definitions. With the proposed methods, we find robust evidence that having firearms at home increases suicide risk.

stat.ME↗

arXiv2023

Differentially Private Bayesian Neural Networks on Accuracy, Privacy and Reliability

Bayesian neural network (BNN) allows for uncertainty quantification in prediction, offering an advantage over regular neural networks that has not been explored in the differential privacy (DP) framework. We fill this important gap by leveraging recent development in Bayesian deep learning and privacy accounting to offer a more precise analysis of the trade-off between privacy and accuracy in BNN. We propose three DP-BNNs that characterize the weight uncertainty for the same network architecture in distinct ways, namely DP-SGLD (via the noisy gradient method), DP-BBP (via changing the parameters of interest) and DP-MC Dropout (via the model architecture). Interestingly, we show a new equivalence between DP-SGD and DP-SGLD, implying that some non-Bayesian DP training naturally allows for uncertainty quantification. However, the hyperparameters such as learning rate and batch size, can have different or even opposite effects in DP-SGD and DP-SGLD. Extensive experiments are conducted to compare DP-BNNs, in terms of privacy guarantee, prediction accuracy, uncertainty quantification, calibration, computation speed, and generalizability to network architecture. As a result, we observe a new tradeoff between the privacy and the reliability. When compared to non-DP and non-Bayesian approaches, DP-SGLD is remarkably accurate under strong privacy guarantee, demonstrating the great potential of DP-BNN in real-world tasks.

cs.LG↗