机器学习相关论文写作整理
This is an excellent collection of articles for analysis, as they are from a cohesive research area (and often the same research group), showcasing a clear and consistent style for writing high-impact scientific papers in computational science and engineering.
Based on an analysis of the provided textbook chapter and research articles, here is a summary of their writing characteristics and a guide on the train of thought for composing a similar article.
Overall Train of Thought
The core logic follows a “funnel” structure, moving from a general problem to a specific solution and then back to general implications.
- Problem: What is the broad, important problem we are trying to solve? (e.g., Inverse Scattering Problems).
- Challenge: Why is this problem hard? What specific aspect makes it difficult? (e.g., It’s ill-posed, nonlinear, and becomes even harder with phaseless data).
- State of the Art: How have others tried to solve this? What are the existing methods and what are their limitations? (e.g., Traditional iterative methods are slow; early deep learning methods are “black boxes”).
- Our Idea (The Gap): We have a novel idea to overcome a specific limitation. (e.g., “We can inspire the neural network design with physics” or “We can quantify the uncertainty of the deep learning prediction”).
- Our Method (The Solution): This is exactly how our idea is implemented, mathematically and algorithmically.
- Proof: How do we prove our method works and is better than the state of the art? (e.g., Through carefully designed numerical experiments and quantitative metrics).
- Implications: What is the significance of our work? Why does it matter and what’s next? (e.g., Our method is faster, more accurate for challenging cases, and opens up new research directions).
Breakdown by Section
Here is how this train of thought is structured in each part of the paper.
1. Abstract
The abstract is a highly condensed version of the entire paper, following a strict formula.
- Starts with: A single sentence establishing the context and the primary challenge.
- Example: “Reconstructing the exact electromagnetic property of unknown targets from the measured scattered field is challenging due to the intrinsic nonlinearity and ill-posedness.”
- The “Gap” and “Our Proposal”: Immediately state what is missing and what you are proposing to fix it.
- Example: “In this article, a new scheme, named the modified contrast scheme (MCS), is proposed to tackle nonlinear inverse scattering problems (ISPs).”
- Core Technical Idea: Briefly state the key mechanism of your method.
- Example: “A local-wave amplifier coefficient is used to form the modified contrast, which is able to alleviate the global nonlinearity…”
- Key Result & Validation: Summarize the main finding from your experiments.
- Example: “The numerical results show that MCS with the modified contrast input performs well in both 2-D and 3-D testing examples…”
- Concluding Impact: A final sentence on why this is a significant improvement.
- Example: “…a significant improvement is achieved in reconstructing high-contrast scatterers.”
2. Introduction
The introduction expands the abstract into a full narrative that guides the reader from the general field to your specific contribution.
- Starts with (The Broad Funnel Opening): A broad statement about the importance and application of the field.
- Example: “Electromagnetic inverse scattering problems (ISPs) are aimed at determining the nature of unknown scatterer… and have wide applications in the fields of nondestructive evaluation, geophysics, biomedical imaging…”
- Defining the Challenge: Explain in more detail why the problem is difficult (nonlinearity, ill-posedness, computational cost, challenges of phaseless data). This educates the reader and sets the stage for why new methods are needed.
- Literature Review (State of the Art): Systematically review existing categories of solutions.
- Start with traditional methods (e.g., iterative solvers like DBIM, CSI, SOM) and state their pros and cons (e.g., physically grounded but slow).
- Move to more recent approaches (e.g., early machine learning, deep learning) and state their pros and cons (e.g., fast but often “black boxes” lacking generalization).
- Identifying the Specific Gap: Explicitly state the limitation in the state-of-the-art that your work will address. This is the most crucial part of the introduction.
- Example: “However, little attention has been paid to the uncertainty quantification of these deep-learning methods…” or “In [29], the similarities between the iterative method… and the architecture of ANN have also been examined, which inspires to propose…”
- Stating Your Contribution (The “Here We Show” Moment): Clearly and concisely list your paper’s contributions. Use a bulleted or numbered list for maximum clarity.
- Example: “The main contributions of the proposed method are threefold. First… Second… Third…”
- Paper Roadmap: End with a brief outline of the rest of the paper.
- Example: “The structure of this article is as follows. Section II presents the formulation… Section III shows the numerical comparison…”
3. Methods (Formulation / Methodology)
This section is the technical core. The goal is precision and clarity, allowing another expert to understand and replicate your work.
- Starts with the Foundation: Begin with the established physics and mathematical model (e.g., the Lippmann–Schwinger equation, the state and data equations). Define all terms, variables, and operators.
- Example: “In ISPs, the forward model can be formulated by the following two discretized equations…”
- Connecting to Your Idea: Bridge the foundational theory to your novel concept. Explain the high-level idea before diving into the mathematical details.
- Example: “Inspired by contrast-source type iterative inversion solvers… we turn our attention from directly regressing contrast to the problem of estimating induced current…”
- Detailing the Novelty: Provide a step-by-step derivation of your method. Use equations, algorithms (in pseudo-code), and diagrams (especially for neural network architectures). Justify your design choices.
- Example: When describing a new input for a neural network, explain why that input is physically meaningful and more suitable than previous approaches (e.g., “The new constant value β₀ for the input generation provides us a freedom to control the performance of the U-Net.”).
- Final Output Formulation: Clearly state how the final result (e.g., the reconstructed image) is obtained from your method’s output.
4. Numerical Results / Experiments
This section provides the evidence. The structure is methodical and focused on comparing your method against relevant benchmarks.
- Starts with the Setup: Describe the simulation or experiment configuration in detail. This is critical for reproducibility.
- Example: “In the numerical tests, we consider a DOI D with the size of 2 × 2 m² and discretize the domain into 64 × 64 pixels. The operating frequency is 400 MHz…”
- Specify training/testing data, noise levels, and the quantitative metrics you will use (e.g., Relative Error, SSIM).
- Present Results by Case Study: Organize the results into distinct examples or tests. Each test should be designed to prove a specific point (e.g., performance on weak scatterers, performance on strong scatterers, generalization ability, robustness to noise).
- Show, Don’t Just Tell: Use figures (reconstructed images) and tables (quantitative scores) extensively. The figures should always include the “Ground Truth” for visual comparison.
- Interpret the Results: For each figure and table, add text that guides the reader’s interpretation.
- Example: “As demonstrated in Fig. 5(c) and (d), some artifacts appear in the reconstruction… while MCSM still offers a satisfactory reconstruction result. The average Re… is listed in Table II.”
5. Discussion and Conclusion
This section zooms back out. It summarizes the work and explains its broader significance.
- Starts with a Restatement: Begin by restating the problem and your proposed solution in a single paragraph.
- Example: “In this article, an improved scheme, named MCS, is proposed to tackle nonlinear ISPs. Under the MCS, both 2-D and 3-D real-time reconstruction results are provided…”
- Summarize Key Findings: Briefly reiterate the most important results and what they prove.
- Example: “It is found that, although the network is trained with MNIST data set, it is able to solve general ISPs and outperforms SOM and DCS in reconstructing some challenging profiles.”
- Discuss the “Why”: Explain why your method performed well. Connect the successful results back to the design features you described in the Methods section. This provides deeper insight.
- Example: “One reason for the better performance of MCSM is that the range of the modified contrast is always between 0 and 1, which makes it more suitable to be the input of the U-Net.”
- Acknowledge Limitations and Propose Future Work: Show that you have a critical understanding of your own work by mentioning its limitations and suggesting how the research could be extended in the future.
- Example: “Although the proposed method quantitatively achieves significantly better results… there is still room for improvements considering the advantages that have been exhibited by some iterative algorithms that are specially designed to tackle highly nonlinear ISPs…”
By following this structured train of thought, you can create a clear, logical, and persuasive scientific article that effectively communicates your contribution to the field.
Abstract
Far-Field Approximation Learning Method for Millimeter-Wave Short-Range Imaging
一种毫米波短距离成像的远场近似学习方法
1 | Abstract — Based on the far-field approximations, a deep learning-based method is proposed for millimeter-wave shortrange imaging. By using convolutional neural networks, the distortions caused by the far-field approximations and limitedaperture measurements could be corrected. Dissimilar to traditional algorithms, the proposed method has no restrictions on the placements of the antenna arrays and single-frequency illuminations are sufficient for the generations of 3-D highresolution reflectivity maps. In addition, it is fast to generate the input of neural network since the algorithm is based on inverse Fourier transform, which is ideal for generating training dataset. The performance of the proposed method is verified using both synthetic and experimental data. It is also demonstrated that enlarging the k-space coverage, which can be accomplished by increasing the dimensions of the antenna arrays, can improve the resolution of the proposed method. |
An Improved Deep Learning Scheme for Solving 2-D and 3-D Inverse Scattering Problems
一种用于求解二维和三维逆散射问题的改进深度学习方案
1 | Abstract— Reconstructing the exact electromagnetic property of unknown targets from the measured scattered field is challenging due to the intrinsic nonlinearity and ill-posedness. In this article, a new scheme, named the modified contrast scheme (MCS), is proposed to tackle nonlinear inverse scattering problems (ISPs). A local-wave amplifier coefficient is used to form the modified contrast, which is able to alleviate the global nonlinearity in original ISPs without decreasing the accuracy of the problem formulation. Moreover, the modified contrast is more suitable to be the input of the deep learning scheme, due to the unity bound of the modified contrast. The numerical results show that MCS with the modified contrast input performs well in both 2-D and 3-D testing examples in real time after offline training process, even in high-relative-permittivity cases. Compared with the dominant current scheme, a significant improvement is achieved in reconstructing high-contrast scatterers. |
MUSIC Imaging and Electromagnetic Inverse Scattering of Multiple-Scattering Small Anisotropic Spheres
多散射小各向异性球体的MUSIC成像和电磁逆散射
1 | Abstract—The Foldy-Lax equation is used to derive a multiple scattering model for the multiple-scattering small anisotropic spheres. By this model, if the number of the non-zero singular values of the multistatic response (MSR) matrix is smaller than the number of the antennas, the range space of the MSR matrix is found to be spanned by the background Green’s function vectors corresponding to the and components of the electric and magnetic dipoles induced in each scatterer, which indicates that the multiple signal classification (MUSIC) method could be implemented to obtain the locations of the scatterers. After estimating the positions of the scatterers, a non-iterative analytical method is proposed for retrieving the polarization strength tensors as well as the orientations of the principle axes of each scatterer. Two numerical simulations show that, the MUSIC method and the non-iterative method are efficacious for the nonlinear inverse scattering problem of determining the locations and polarization strength tensors of multiple-scattering small anisotropic spheres. Such methods could also be applied to the inversion of small isotropic spheres or extended to the inversion of small bianisotropic spheres. |
Deep Learning-Based Inversion Methods for Solving Inverse Scattering Problems With Phaseless Data
基于深度学习的求解无相位数据逆散射问题的反演方法
1 | Abstract— Without phase information of the measured field data, the phaseless data inverse scattering problems (PD-ISPs) counter more serious nonlinearity and ill-posedness compared with full data ISPs (FD-ISPs). In this article, we propose a learning-based inversion approach in the frame of the U-net convolutional neural network (CNN) to quantitatively image unknown scatterers located in homogeneous background from the amplitude-only measured total field (also denoted PD). Three training schemes with different inputs to the U-net CNN are proposed and compared, i.e., the direct inversion scheme (DIS) with phaseless total field data, retrieval dominant induced currents by the Levenberg–Marquardt (LM) method (PD-DICs), and PD with contrast source inversion (PD-CSI) scheme. We also demonstrate the setup of training data and compare the performance of the three schemes using both numerical and experimental tests. It is found that the proposed PD-CSI and PD-DICs perform better in terms of accuracy, generalization ability, and robustness compared with DIS. PD-CSI has the strongest capability to tackle with PD-ISPs, which outperforms the PD-DICs and DIS. |
Uncertainty Quantification in Inverse Scattering Problems With Bayesian Convolutional Neural Networks
贝叶斯卷积神经网络在逆散射问题中的不确定性量化
1 | Abstract— Recently, tremendous progress has been achieved in applying deep learning schemes (DLSs) to solve inverse scattering problems (ISPs), where state-of-the-art performance has been attained. However, little attention has been paid to the uncertainty quantification of these deep-learning methods in solving ISPs. In other words, the error of the prediction is not known since the ground truth is not available in practice. In this article, a Bayesian convolutional neural network (BCNN) is used to quantify the uncertainties in solving ISPs. With Monte Carlo dropout, the proposed BCNN is able to directly predict the pixel-based uncertainties of the widely used DLSs in ISPs. To quantitatively evaluate the performance of uncertainty predictions, both correlation coefficient and nonlinear correlation distribution between the predicted uncertainty and true absolute error are calculated. The tests on both synthetic and experimental data show that the predictive uncertainty is highly correlated with true absolute error calculated from the ground truth. Besides, when DLSs fail to solve an ISP, the predicted uncertainty increases significantly, which offers a pixel-based “confidence level” in solving ISPs. |
Physics-Inspired Convolutional Neural Network for Solving Full-Wave Inverse Scattering Problems
求解全波逆散射问题的物理启发卷积神经网络
1 | Abstract— In this paper, to bridge the gap between physical knowledge and learning approaches, we propose an induced current learning method (ICLM) by incorporating merits in traditional iterative algorithms into the architecture of convolutional neural network (CNN). The main contributions of the proposed method are threefold. First, to the best of our knowledge, it is the first time that the contrast source is learned to solve full-wave inverse scattering problems (ISPs). Second, inspired by the basis-expansion strategy in the traditional iterative approach for solving ISPs, a combined loss function with multiple labels is defined in a cascaded end-to-end CNN (CEE-CNN) architecture to decrease the nonlinearity of objective function, where no additional computational cost is introduced in generating extra labels. Third, to accelerate the convergence speed and decrease the difficulties of the learning process, the proposed CEE-CNN is designed to focus on learning the minor part of the induced current by introducing several skip connections and to bypass the major part of induced current to the output layers. The proposed method is compared with the state-of-the-art of deep learning scheme and a well-known iterative ISP solver, where numerical and experimental tests are conducted to verify the proposed ICLM. |
Precise Near-Range 3-D Image Reconstruction Based on MIMO Circular Synthetic Aperture Radar
基于MIMO圆形合成孔径雷达的精确近距离三维图像重建
1 | Abstract— This article studies the near-range 3-D image reconstruction based on millimeter-wave multiple-input-multipleoutput circular synthetic aperture radar (MIMO-CSAR). We first derive the exact forward wave model of MIMO-CSAR and then propose a range migration algorithm for near-range high-precision 3-D image reconstruction. Our formulation is strictly based on wave theory. The propagation attenuation, which is an inherent factor in wave equation and is significant for near-range wide-angle sensing case, is considered and compensated efficiently with a Fourier-domain matched filtering. Besides, the phase decoupling is achieved with a 2-D Stolt mapping. The computational complexity of the algorithm is at the same level as the state-of-the-art method. Compared with the state of the art, by considering the amplitude variation, the imaging quality can be improved for near-range measurement with large beamwidth. The results from both synthetic and practical measurement data demonstrate that a wider dynamic range (less sidelobes appeared at the same dynamic level) and a better recovery for widespread targets compared with the state-of-the-art imaging method. |
Learning-Based Fast Electromagnetic Scattering Solver Through Generative Adversarial Network
基于生成对抗网络的快速电磁散射求解器
1 | Abstract— This article proposes a learning-based noniterative method to solve electromagnetic (EM) scattering problems utilizing pix2pix, a popular generative adversarial network (GAN). Instead of calculating induced currents directly from a matrix inversion, a forward-induced current learning method (FICLM) is introduced to calculate the induced current through a neural-network mapping. The scattered fields can be further calculated through a multiplication of the Green’s function with the predicted induced currents. Inspired by wave physics of scattering problems, we have designed three kinds of input schemes, covering different combinations of the given incident field and permittivity contrast, to evaluate the performance of the FICLM model under both single-incidence and multi-incidence cases. Numerical results prove that the proposed FICLM outperforms the method of moments (MoM) in terms of both computational speed and accuracy by use of reference data with a higher precision. The FICLM with the direct sum of permittivity contrast and a so-called Born-type induced current, achieves the best calculation accuracy and generalization capability compared to the other two inputs. The comparison with other types of neural networks, such as U-net, also demonstrates the superior performance of FICLM for dealing with complex scatterers due to the use of adversarial framework in pix2pix. The proposed method paves a new way for the fast solution of EM-scattering problems through deep learning techniques. |