(The following are the academic papers I published during my undergraduate studies, which were of a relatively low level. I hope my efforts are not too lacking.)
(以下是我在本科期间发表的学术论文，水平较低，让大家见笑了)
Image classification has become crucial in fields such as computer vision, medical diagnostics, and autonomous driving, with significant advancements driven by convolutional neural networks (CNNs). The performance of CNNs depends not only on their architecture but also on the activation functions used. Activation functions introduce non-linearity, enabling the learning of complex patterns. While ReLU is popular due to its simplicity and effectiveness, it has limitations such as the "dying ReLU" problem. This study systematically evaluates the impact of different activation functions, including Sigmoid, Tanh, Leaky ReLU, and Exponential Linear Unit (ELU), on CNN performance in image classification. To isolate the effects of each activation function, we maintained a consistent network architecture and varied only the activation functions. Experiments were conducted using the MNIST dataset, a benchmark in image classification, to analyze how these functions affect accuracy and learning dynamics. Results indicate significant differences in performance, with each activation function offering unique advantages and drawbacks, providing insights into their specific contributions to CNN effectiveness.
随着卷积神经网络（CNN）的发展，图像分类在计算机视觉、医学诊断和自动驾驶等领域起着至关重要的作用。CNN的性能不仅受到网络架构的影响，还依赖于所使用的激活函数。引入非线性的激活函数使学习复杂模式成为可能。虽然ReLU因其简单有效而广受欢迎，但它也有局限，例如“神经元死亡”问题。本研究系统地评估了不同激活函数（包括Sigmoid、Tanh、Leaky ReLU和ELU）对CNN在图像分类中的性能的影响。为了区分每个激活函数的影响，我们设置了相同的网络架构，只改变了激活函数。实验采用图像分类基准MNIST数据集，以分析这些函数如何影响准确性和学习过程。结果表明，不同激活函数的性能存在显著差异，每种激活函数都有其独特的优缺点，从而深入了解它们对CNN有效性的具体贡献。
It is of great significance for environmental protection to explore the impact of climate and drought on different plant communities. We propose a grassland mechanism model based on ordinary differential equations to analyze the impact of different factors, including three modules and an assessment method for the environment. In order to quantitatively evaluate drought factors, we defined degree of drought as an indicator that plays an important role in the model. We implemented our model with MATLAB and made qualitative and quantitative analysis of analog data. As a conclusion, we propose five measures to give our contribution to long-term viability of a plant community and the larger environment.
探讨气候和干旱对不同植物群落的影响，对环境保护具有重要意义。我们提出了一个基于常微分方程的草原机制模型来分析不同因素的影响，包括三个模块和一种环境评估方法。为了定量评价干旱因素，我们将干旱程度定义为一个在模型中发挥重要作用的指标。我们用MATLAB实现了我们的模型，并对模拟数据进行了定性和定量分析。最后，我们提出了五项措施，为植物群落和更大环境的长期生存做出贡献。
(The following are the academic reports I made during my undergraduate studies, which were of a relatively low level. I hope my efforts are not too lacking.)
(以下是我在本科期间所作的学术报告，水平较低，让大家见笑了)
We first explored the development process of deep learning in the field of image classification. Considering the problem of insufficient computing power, we trained the distilled DeiT-B model based on the CIFAR-10 dataset and compared it with the Resnet18 model. We found that the performance of the large model was much better than that of the traditional model. Meanwhile, we conducted three ablation experiments on the Resnet18 model.
我们首先探究了深度学习在图像分类领域的发展历程，考虑到算力不足的问题，我们基于CIFAR-10数据集用蒸馏出的DeiT-B模型进行训练，并用Resnet18模型进行对比，发现大模型的效果远远好于传统模型。同时，我们对Resnet18模型进行了三项消融实验。
We investigated the current research status of deep learning for animals and identified two goals: creating a wild bird video dataset and developing video analysis tools. We proposed a technical roadmap for individual recognition based on the YOLO model and posture analysis based on the SLEAP model.
我们探究了深度学习用于动物的研究现状，确定了制作野生鸟类视频数据集和开发视频分析工具两个目标，提出了基于yolo模型的个体识别和基于sleap模型的姿态分析的技术路线。
We conducted in-depth research on the network architecture of Resnet18 and conducted main experiments and three ablation experiments based on the MNIST dataset: removing the fourth layer of the network, replacing the SGD optimizer with the Adam optimizer, and changing the learning rate of the optimizer.
我们深入研究了Resnet18的网络架构，基于MNIST数据集进行了主实验和三项消融实验：去除网络的第四层、将SGD优化器更换为Adam优化器和改变优化器的学习率。
We investigated the residual connections and VGG19 based network structure of the milestone Resnet model, and trained it on the garbage classification dataset, achieving good results.
我们探究了具有里程碑意义的Resnet模型的残差连接和基于VGG19的网络结构，并在垃圾分类数据集上进行了训练，取得了优良的结果。
Based on the form of soliciting manuscripts, we propose a business plan to build an E-commerce platform of recreation & entertainment, examine the current market situation, and describe our business model.
基于约稿这一交易形式，我们提出了构建电子娱乐商业平台的商业计划，考察了目前的市场情况，描绘了我们的商业模型。