Search

Generative 3D part assembly involves understanding part relationships and predicting their 6-DoF poses for assembling a realistic 3D shape. Prior work often focus on the geometry of individual parts, neglecting part-whole hierarchies of objects. Leveraging two key observations: 1) super-part poses provide strong hints about part poses, and 2) predicting super-part poses is easier due to fewer superparts, we propose a part-whole-hierarchy message passing network for efficient 3D part assembly. We first introduce super-parts by grouping geometrically similar parts without any semantic labels. Then we employ a part-whole hierarchical encoder, wherein a super-part encoder predicts latent super-part poses based on input parts. Subsequently, we transform the point cloud using the latent poses, feeding it to the part encoder for aggregating super-part information and reasoning about part relationships to predict all part poses. In training, only ground-truth part poses are required. During inference, the predicted latent poses of super-parts enhance interpretability. Experimental results on the PartNet dataset show that our method achieves state-of-the-art performance in part and connectivity accuracy and enables an interpretable hierarchical part assembly. Code is available at https://github.com/pkudba/3DHPA.

Numerous mobile apps have leveraged deep learning capabilities. However, on-device models are vulnerable to attacks as they can be easily extracted from their corresponding mobile apps. Existing on-device attacking approaches only generate black-box attacks, which are far less effective and efficient than white-box strategies. This is because mobile deep learning frameworks like TFLite do not support gradient computing, which is necessary for white-box attacking algorithms. Thus, we argue that existing findings may underestimate the harmfulness of on-device attacks. To this end, we conduct a study to answer this research question: Can on-device models be directly attacked via white-box strategies? We first systematically analyze the difficulties of transforming the on-device model to its debuggable version, and propose a Reverse Engineering framework for On-device Models (REOM), which automatically reverses the compiled on-device TFLite model to the debuggable model. Specifically, REOM first transforms compiled on-device models into Open Neural Network Exchange format, then removes the non-debuggable parts, and converts them to the debuggable DL models format that allows attackers to exploit in a white-box setting. Our experimental results show that our approach is effective in achieving automated transformation among 244 TFLite models. Compared with previous attacks using surrogate models, REOM enables attackers to achieve higher attack success rates with a hundred times smaller attack perturbations. In addition, because the ONNX platform has plenty of tools for model format exchanging, the proposed method based on the ONNX platform can be adapted to other model formats. Our findings emphasize the need for developers to carefully consider their model deployment strategies, and use white-box methods to evaluate the vulnerability of on-device models., Comment: Published in The International Conference on Software Engineering 2024 (ICSE'24)

Large language models (LLMs) are becoming increasingly important for machine learning applications. However, it can be challenging to align LLMs with our intent, particularly when we want to generate content that is preferable over others or when we want the LLM to respond in a certain style or tone that is hard to describe. To address this challenge, we propose an approach that uses contrastive examples to better describe our intent. This involves providing positive examples that illustrate the true intent, along with negative examples that show what characteristics we want LLMs to avoid. The negative examples can be retrieved from labeled data, written by a human, or generated by the LLM itself. Before generating an answer, we ask the model to analyze the examples to teach itself what to avoid. This reasoning step provides the model with the appropriate articulation of the user's need and guides it towards generting a better answer. We tested our approach on both synthesized and real-world datasets, including StackExchange and Reddit, and found that it significantly improves performance compared to standard few-shot prompting, Comment: Accepted to appear at AAAI 2024

The three-level system represents the smallest quantum system capable of autonomous cycling in quantum heat engines. This study proposes a method to simulate the steady-state dynamics of a three-level quantum heat engine by designing and implementing superconducting quantum circuits. Following error mitigation, the outcomes from the quantum circuit model designed in this study, when executed on a real quantum device, closely align with theoretical predictions, thereby validating the effectiveness of the circuit model. This study offers a novel approach for investigating three-level quantum heat engines, enabling the verification of theoretical research findings while also reducing the complexity and cost of experimental procedures.

The convenience of 3D sensors has led to an increase in the use of 3D point clouds in various applications. However, the differences in acquisition devices or scenarios lead to divergence in the data distribution of point clouds, which requires good generalization of point cloud representation learning methods. While most previous methods rely on domain adaptation, which involves fine-tuning pre-trained models on target domain data, this may not always be feasible in real-world scenarios where target domain data may be unavailable. To address this issue, we propose InvariantOODG, which learns invariability between point clouds with different distributions using a two-branch network to extract local-to-global features from original and augmented point clouds. Specifically, to enhance local feature learning of point clouds, we define a set of learnable anchor points that locate the most useful local regions and two types of transformations to augment the input point clouds. The experimental results demonstrate the effectiveness of the proposed model on 3D domain generalization benchmarks.

Methane (CH4) is the second most critical greenhouse gas after carbon dioxide, contributing to 16-25% of the observed atmospheric warming. Wetlands are the primary natural source of methane emissions globally. However, wetland methane emission estimates from biogeochemistry models contain considerable uncertainty. One of the main sources of this uncertainty arises from the numerous uncertain model parameters within various physical, biological, and chemical processes that influence methane production, oxidation, and transport. Sensitivity Analysis (SA) can help identify critical parameters for methane emission and achieve reduced biases and uncertainties in future projections. This study performs SA for 19 selected parameters responsible for critical biogeochemical processes in the methane module of the Energy Exascale Earth System Model (E3SM) land model (ELM). The impact of these parameters on various CH4 fluxes is examined at 14 FLUXNET- CH4 sites with diverse vegetation types. Given the extensive number of model simulations needed for global variance-based SA, we employ a machine learning (ML) algorithm to emulate the complex behavior of ELM methane biogeochemistry. ML enables the computational time to be shortened significantly from 6 CPU hours to 0.72 milliseconds, achieving reduced computational costs. We found that parameters linked to CH4 production and diffusion generally present the highest sensitivities despite apparent seasonal variation. Comparing simulated emissions from perturbed parameter sets against FLUXNET-CH4 observations revealed that better performances can be achieved at each site compared to the default parameter values. This presents a scope for further improving simulated emissions using parameter calibration with advanced optimization techniques like Bayesian optimization., Comment: 24 pages, 9 figures and 2 tables

Quantum simulation of a monopole-spin hybrid system is performed on basis of a dipolar ultracold gas in a ladder lattice. The site-occupation states of the dipolar ladder lattice gas can spontaneously emulate both the monopole and spin excitations. The hopping of the atoms induces a particle conversion process between spin and monopole pairs, and the dipole-dipole interaction determines the spin-spin, spin-monopole and monopole-monopole interactions. The anisotropic nature of the dipole-dipole interaction allows hereby for a flexible engineering of the designed hybrid system, and for a significant tunability of the interaction strengths. As a result, we encounter a rich phase diagram, and specifically a self-assembled Coulomb phase arises, in which monopoles and spins coexist and are orderly arranged according to the local Gauss's law. The Coulomb phase hosts a zoo of different types of quasiparticles, and provides the possibility to simulate various phenomena in particle physics, such as a degenerate vacuum, particle decay and conversion processes. Our work provides a significant extension of the scope of quantum simulations based on the anisotropy of dipolar interactions., Comment: 11 pages,8 figures

In recent years, the field of metal-halide perovskite emitters has rapidly emerged as a new community in solid-state lighting. Their exceptional optoelectronic properties have contributed to the rapid rise in external quantum efficiencies (EQEs) in perovskite light-emitting diodes (PeLEDs) from <1% (in 2014) to approaching 30% (in 2023) across a wide range of wavelengths. However, several challenges still hinder their commercialization, including the relatively low EQEs of blue/white devices, limited EQEs in large-area devices, poor device stability, as well as the toxicity of the easily accessible lead components and the solvents used in the synthesis and processing of PeLEDs. This roadmap addresses the current and future challenges in PeLEDs across fundamental and applied research areas, by sharing the community's perspectives. This work will provide the field with practical guidelines to advance PeLED development and facilitate more rapid commercialization., Comment: 103 pages, 29 figures. This is the version of the article before peer review or editing, as submitted by an author to Journal of Physics: Photonics. IOP Publishing Ltd is not responsible for any errors or omissions in this version of the manuscript or any version derived from it

In a recent paper, Kolountzakis and Papageorgiou ask if for every $\epsilon \in (0,1)$, there exists a set $S \subseteq \mathbb{R}$ such that $\vert S \cap I\vert \geq 1 - \epsilon$ for every interval $I \subset \mathbb{R}$ with unit length, but that does not contain any affine copy of a given increasing sequence of exponential growth or faster. This question is an analogue of the well-known Erd\H{o}s similarity problem. In this paper, we show that for each sequence of real numbers whose integer parts form a set of positive upper Banach density, one can explicitly construct such a set $S$ that contains no affine copy of that sequence. Since there exist sequences of arbitrarily rapid growth that satisfy this condition, our result answers Kolountzakis and Papageorgiou's question in the affirmative. A key ingredient of our proof is a generalization of results by Amice, Kahane, and Haight from metric number theory. In addition, we construct a set $S$ with the required property -- but with $\epsilon \in (1/2, 1)$ -- that contains no affine copy of $\{2^n\}$., Comment: 17 pages

With the widespread success of deep learning technologies, many trained deep neural network (DNN) models are now publicly available. However, directly reusing the public DNN models for new tasks often fails due to mismatching functionality or performance. Inspired by the notion of modularization and composition in software reuse, we investigate the possibility of improving the reusability of DNN models in a more fine-grained manner. Specifically, we propose two modularization approaches named CNNSplitter and GradSplitter, which can decompose a trained convolutional neural network (CNN) model for $N$-class classification into $N$ small reusable modules. Each module recognizes one of the $N$ classes and contains a part of the convolution kernels of the trained CNN model. Then, the resulting modules can be reused to patch existing CNN models or build new CNN models through composition. The main difference between CNNSplitter and GradSplitter lies in their search methods: the former relies on a genetic algorithm to explore search space, while the latter utilizes a gradient-based search method. Our experiments with three representative CNNs on three widely-used public datasets demonstrate the effectiveness of the proposed approaches. Compared with CNNSplitter, GradSplitter incurs less accuracy loss, produces much smaller modules (19.88% fewer kernels), and achieves better results on patching weak models. In particular, experiments on GradSplitter show that (1) by patching weak models, the average improvement in terms of precision, recall, and F1-score is 17.13%, 4.95%, and 11.47%, respectively, and (2) for a new task, compared with the models trained from scratch, reusing modules achieves similar accuracy (the average loss of accuracy is only 2.46%) without a costly training process. Our approaches provide a viable solution to the rapid development and improvement of CNN models., Comment: Accepted by ACM Transactions on Software Engineering and Methodology (TOSEM). arXiv admin note: substantial text overlap with arXiv:2209.06116

We investigate the scattering processes of two photons in a one-dimensional waveguide coupled to two giant atoms. By adjusting the accumulated phase shifts between the coupling points, we are able to effectively manipulate the characteristics of these scattering photons. Utilizing the Lippmann-Schwinger formalism, we derive analytical expressions for the wave functions describing two-photon interaction in separate, braided, and nested configurations. Based on these wave functions, we also obtain analytical expressions for the incoherent power spectra and second-order correlation functions. In contrast to small atoms, the incoherent spectrum, which is defined by the correlation of the bound state, can exhibit more tunability due to the phase shifts. Additionally, the second-order correlation functions in the transmission and reflection fields could be tuned to exhibit either bunching or antibunching upon resonant driving. These unique features offered by the giant atoms in waveguide QED could benefit the generation of nonclassical itinerant photons in quantum networks.

Injecting structure into neural networks enables learning functions that satisfy invariances with respect to subsets of inputs. For instance, when learning generative models using neural networks, it is advantageous to encode the conditional independence structure of observed variables, often in the form of Bayesian networks. We propose the Structured Neural Network (StrNN), which injects structure through masking pathways in a neural network. The masks are designed via a novel relationship we explore between neural network architectures and binary matrix factorization, to ensure that the desired independencies are respected. We devise and study practical algorithms for this otherwise NP-hard design problem based on novel objectives that control the model architecture. We demonstrate the utility of StrNN in three applications: (1) binary and Gaussian density estimation with StrNN, (2) real-valued density estimation with Structured Autoregressive Flows (StrAFs) and Structured Continuous Normalizing Flows (StrCNF), and (3) interventional and counterfactual analysis with StrAFs for causal inference. Our work opens up new avenues for learning neural networks that enable data-efficient generative modeling and the use of normalizing flows for causal effect estimation., Comment: 10 pages with 5 figures, to be published in Neural Information Processing Systems 2023

Mobile software engineering has been a hot research topic for decades. Our fellow researchers have proposed various approaches (with over 7,000 publications for Android alone) in this field that essentially contributed to the great success of the current mobile ecosystem. Existing research efforts mainly focus on popular mobile platforms, namely Android and iOS. OpenHarmony, a newly open-sourced mobile platform, has rarely been considered, although it is the one requiring the most attention as OpenHarmony is expected to occupy one-third of the market in China (if not in the world). To fill the gap, we present to the mobile software engineering community a research roadmap for encouraging our fellow researchers to contribute promising approaches to OpenHarmony. Specifically, we start by presenting a literature review of mobile software engineering, attempting to understand what problems have been targeted by the mobile community and how they have been resolved. We then summarize the existing (limited) achievements of OpenHarmony and subsequently highlight the research gap between Android/iOS and OpenHarmony. This research gap eventually helps in forming the roadmap for conducting software engineering research for OpenHarmony.

We present a lattice QCD calculation of the transversity isovector- and isoscalar-quark parton distribution functions (PDFs) of the proton utilizing a perturbative matching at next-to-leading-order (NLO) accuracy. Additionally, we determine the isovector and isoscalar tensor charges for the proton. In both calculations, the disconnected contributions to the isoscalar matrix elements have been ignored. The calculations are performed using a single ensemble of $N_f = 2 +1$ highly-improved staggered quarks simulated with physical-mass quarks and a lattice spacing of $a = 0.076$ fm. The Wilson-clover action, with physical quark masses and smeared gauge links obtained from one iteration of hypercubic smearing, is used in the valence sector. Using the NLO operator product expansion, we extract the lowest four to six Mellin moments and the PDFs via a neural network from the matrix elements in the pseudo-PDF approach. In addition, we calculate the PDFs in the quasi-PDF approach with hybrid-scheme renormalization and the recently developed leading-renormalon resummation technique, at NLO with the resummation of leading small-$x$ logarithms., Comment: 23 pages, 22 figures, and 2 tables; version accepted for publication in PRD

Recently, we made significant advancements in improving the computational efficiency of lattice QCD calculations for Generalized Parton Distributions (GPDs). This progress was achieved by adopting calculations of matrix elements in asymmetric frames, deviating from the computationally-expensive symmetric frame typically used, and allowing freedom in the choice for the distribution of the momentum transfer between the initial and final states. A crucial aspect of this approach involves the adoption of a Lorentz covariant parameterization for the matrix elements, introducing Lorentz-invariant amplitudes. This approach also allows us to propose an alternative definition of quasi-GPDs, ensuring frame independence and potentially reduce power corrections in matching to light-cone GPDs. In our previous work, we presented lattice QCD results for twist-2 unpolarized GPDs ($H$ and $E$) of quarks obtained from calculations performed in asymmetric frames at zero skewness. Building upon this work, we now introduce a novel Lorentz covariant parameterization for the axial-vector matrix elements. We employ this parameterization to compute the axial-vector GPD $\widetilde{H}$ at zero skewness, using an $N_f=2+1+1$ ensemble of twisted mass fermions with clover improvement. The light-quark masses employed in our calculations correspond to a pion mass of approximately 260 MeV., Comment: 32 pages, 20 figures. Version accepted for publication in Physical Review D

The integration of individual microgrids (MGs) into Microgrid Clusters (MGCs) significantly improves the reliability and flexibility of energy supply, through resource sharing and ensuring backup during outages. The dispatch of MGCs is the key challenge to be tackled to ensure their secure and economic operation. Currently, there is a lack of optimization method that can achieve a trade-off among top-priority requirements of MGCs' dispatch, including fast computation speed, optimality, multiple objectives, and risk mitigation against uncertainty. In this paper, a novel Multi-Objective, Risk-Sensitive, and Online Trust Region Policy Optimization (RS-TRPO) Algorithm is proposed to tackle this problem. First, a dispatch paradigm for autonomous MGs in the MGC is proposed, enabling them sequentially implement their self-dispatch to mitigate potential conflicts. This dispatch paradigm is then formulated as a Markov Game model, which is finally solved by the RS-TRPO algorithm. This online algorithm enables MGs to spontaneously search for the Pareto Frontier considering multiple objectives and risk mitigation. The outstanding computational performance of this algorithm is demonstrated in comparison with mathematical programming methods and heuristic algorithms in a modified IEEE 30-Bus Test System integrated with four autonomous MGs.

State-of-the-art techniques for monocular camera reconstruction predominantly rely on the Structure from Motion (SfM) pipeline. However, such methods often yield reconstruction outcomes that lack crucial scale information, and over time, accumulation of images leads to inevitable drift issues. In contrast, mapping methods based on LiDAR scans are popular in large-scale urban scene reconstruction due to their precise distance measurements, a capability fundamentally absent in visual-based approaches. Researchers have made attempts to utilize concurrent LiDAR and camera measurements in pursuit of precise scaling and color details within mapping outcomes. However, the outcomes are subject to extrinsic calibration and time synchronization precision. In this paper, we propose a novel cost-effective reconstruction pipeline that utilizes a pre-established LiDAR map as a fixed constraint to effectively address the inherent scale challenges present in monocular camera reconstruction. To our knowledge, our method is the first to register images onto the point cloud map without requiring synchronous capture of camera and LiDAR data, granting us the flexibility to manage reconstruction detail levels across various areas of interest. To facilitate further research in this domain, we have released Colmap-PCD${^{3}}$, an open-source tool leveraging the Colmap algorithm, that enables precise fine-scale registration of images to the point cloud map.

In order to further advance the accuracy and robustness of the incremental parameter estimation-based rotation averaging methods, in this paper, a new member of the Incremental Rotation Averaging (IRA) family is introduced, which is termed as IRAv4. As the most significant feature of the IRAv4, a task-specific connected dominating set is extracted to serve as a more reliable and accurate reference for rotation global alignment. In addition, to further address the limitations of the existing rotation averaging benchmark of relying on the slightly outdated Bundler camera calibration results as ground truths and focusing solely on rotation estimation accuracy, this paper presents a new COLMAP-based rotation averaging benchmark that incorporates a cross check between COLMAP and Bundler, and employ the accuracy of both rotation and downstream location estimation as evaluation metrics, which is desired to provide a more reliable and comprehensive evaluation tool for the rotation averaging research. Comprehensive comparisons between the proposed IRAv4 and other mainstream rotation averaging methods on this new benchmark demonstrate the effectiveness of our proposed approach., Comment: Submitted to IEEE Transactions

In order to use the Dual Simplex Method, one needs to prove a certain bijection between the dictionaries associated with the primal problem and those associated with its dual. We give a short conceptual proof of why this bijection exists.

We first classify all Nijenhuis operators on 2-dimensional complex pre-Lie algebras and 3-dimensional complex associative algebras. Moreover, applying the obtained classification, we provide solutions of classical Yang-Baxter equation of the semi-direct product sub-adjacent Lie algebras.

Quantum thermodynamic relationships in emerging nanodevices are significant but often complex to deal with. The application of machine learning in quantum thermodynamics has provided a new perspective. This study employs reinforcement learning to output the optimal cycle of quantum heat engine. Specifically, the soft actor-critic algorithm is adopted to optimize the cycle of three-level coherent quantum heat engine with the aim of maximal average power. The results show that the optimal average output power of the coherent three-level heat engine is 1.28 times greater than the original cycle (steady limit). Meanwhile, the efficiency of the optimal cycle is greater than the Curzon-Ahlborn efficiency as well as reporting by other researchers. Notably, this optimal cycle can be fitted as an Otto-like cycle by applying the Boltzmann function during the compression and expansion processes, which illustrates the effectiveness of the method.

Static friction induced by moir\'e superstructure in twisted incommensurate finite layered material interfaces reveals unique double periodicity and lack of scaling with contact size. The underlying mechanism involves compensation of incomplete moir\'e tiles at the rim of rigid polygonal graphene flakes sliding atop fixed graphene or h-BN substrates. The scaling of friction (or lack thereof) with contact size is found to strongly depend on the shape of the slider and the relative orientation between its edges and the emerging superstructure, partially rationalizing scattered experimental data. With careful consideration of the flake edge orientation, twist angle, and sliding direction along the substrate, one should therefore be able to achieve large-scale superlubricity via shape tailoring.

This paper proposes a grant-free massive access scheme based on the millimeter wave (mmWave) extra-large-scale multiple-input multiple-output (XL-MIMO) to support massive Internet-of-Things (IoT) devices with low latency, high data rate, and high localization accuracy in the upcoming sixth-generation (6G) networks. The XL-MIMO consists of multiple antenna subarrays that are widely spaced over the service area to ensure line-of-sight (LoS) transmissions. First, we establish the XL-MIMO-based massive access model considering the near-field spatial non-stationary (SNS) property. Then, by exploiting the block sparsity of subarrays and the SNS property, we propose a structured block orthogonal matching pursuit algorithm for efficient active user detection (AUD) and channel estimation (CE). Furthermore, different sensing matrices are applied in different pilot subcarriers for exploiting the diversity gains. Additionally, a multi-subarray collaborative localization algorithm is designed for localization. In particular, the angle of arrival (AoA) and time difference of arrival (TDoA) of the LoS links between active users and related subarrays are extracted from the estimated XL-MIMO channels, and then the coordinates of active users are acquired by jointly utilizing the AoAs and TDoAs. Simulation results show that the proposed algorithms outperform existing algorithms in terms of AUD and CE performance and can achieve centimeter-level localization accuracy., Comment: To appear in IEEE Transactions on Communications. Codes will be open to all on https://gaozhen16.github.io/ soon

An anisotropic interlayer force field that describes the interlayer interactions in homogeneous and heterogeneous interfaces of group-VI transition metal dichalcogenides (MX2 where M = Mo, W and X = S, Se) is presented. The force field is benchmarked against density functional theory calculations for bilayer systems within the Heyd-Scuseria-Ernzerhof hybrid density functional approximation, augmented by a nonlocal many-body dispersion treatment of long-range correlation. The parametrization yields good agreement with reference calculations of binding energy curves and sliding potential energy surfaces. It is found to be transferable to TMD junctions outside the training set that contain the same atom types. Calculated bulk moduli agree with most previous dispersion corrected DFT predictions, which underestimate available experimental values. Calculated phonon spectra of the various junctions under consideration demonstrate the importance of appropriately treating the anisotropic nature of layered interfaces. Considering our previous parameterization for MoS2, the interlayer potential enables accurate and efficient large-scale simulations of the dynamical, tribological, and thermal transport properties of a large set of homogeneous and heterogeneous TMD interfaces.

The extended Kalman filter (EKF) is a common state estimation method for discrete nonlinear systems. It recursively executes the propagation step as time goes by and the update step when a set of measurements arrives. In the update step, the EKF linearizes the measurement function only once. In contrast, the iterated EKF (IEKF) refines the state in the update step by iteratively solving a least squares problem. The IEKF has been extended to work with state variables on manifolds which have differentiable $\boxplus$ and $\boxminus$ operators, including Lie groups. However, existing descriptions are often long, deep, and even with errors. This note provides a quick reference for the IEKF on manifolds, using freshman-level matrix calculus. Besides the bare-bone equations, we highlight the key steps in deriving them., Comment: 2 pages excluding references

We propose a new method to calculate parton distribution functions (PDFs) from lattice correlations of boosted quarks and gluons in the Coulomb gauge. Compared to the widely used gauge-invariant Wilson-line operators, these correlations greatly simplify the renormalization thanks to the absence of linear power divergence. Besides, they enable access to larger off-axis momenta under preserved 3D rotational symmetry, as well as enhanced long-range precision that facilitates the Fourier transform. We verify the factorization formula that relates this new observable to the quark PDF at one-loop order in perturbation theory. Moreover, through a lattice calculation of the pion valence quark PDF, we demonstrate the aforementioned advantage and features of the Coulomb gauge correlation and show that it yields consistent result with the gauge-invariant method. This opens the door to a more efficient way to calculate parton physics on the lattice., Comment: 5 pages, 4 figures, and the appendix

Deep neural network (DNN) models have become increasingly crucial components in intelligent software systems. However, training a DNN model is typically expensive in terms of both time and money. To address this issue, researchers have recently focused on reusing existing DNN models - borrowing the idea of code reuse in software engineering. However, reusing an entire model could cause extra overhead or inherits the weakness from the undesired functionalities. Hence, existing work proposes to decompose an already trained model into modules, i.e., modularizing-after-training, and enable module reuse. Since trained models are not built for modularization, modularizing-after-training incurs huge overhead and model accuracy loss. In this paper, we propose a novel approach that incorporates modularization into the model training process, i.e., modularizing-while-training (MwT). We train a model to be structurally modular through two loss functions that optimize intra-module cohesion and inter-module coupling. We have implemented the proposed approach for modularizing Convolutional Neural Network (CNN) models in this work. The evaluation results on representative models demonstrate that MwT outperforms the state-of-the-art approach. Specifically, the accuracy loss caused by MwT is only 1.13 percentage points, which is 1.76 percentage points less than that of the baseline. The kernel retention rate of the modules generated by MwT is only 14.58%, with a reduction of 74.31% over the state-of-the-art approach. Furthermore, the total time cost required for training and modularizing is only 108 minutes, half of the baseline., Comment: Accepted at ICSE'24