请求推荐适用于大数据场景的深度神经网络分类器相关研究论文

Hey there! I totally get your frustration—when working on DNN classifiers for large datasets, it’s way too easy to get flooded with CNN-focused literature while missing the core work that explains how to design and train DNNs thoughtfully, not just by stacking layers. Here’s a curated list of papers that dive into the "why" and "how" behind effective DNN classification, perfect for someone familiar with neural nets but looking to go deeper:

1. Architecture Design: Moving Beyond Trivial Layer Stacking #

• Network In Network
  While it’s often cited in CNN contexts, the modular "mlpconv" concept translates directly to fully-connected DNN classifiers. This paper breaks down the myth that more layers = better performance; instead, it emphasizes designing task-aligned feature transformation modules rather than just adding fully-connected layers. You’ll learn how to structure sub-networks to capture complex patterns relevant to your classification task, with clear comparisons of module performance to guide your design choices.
• Wide & Deep Learning for Recommender Systems
  Even though it’s focused on recommenders, the Wide&Deep framework is a goldmine for DNN classifier design. It explains why balancing a "wide" layer (for memorizing known patterns) and a "deep" network (for generalizing to new data) is critical for large-scale tasks. For your big data use case, it also covers strategies for handling high-dimensional input features efficiently—no more guessing how to structure your input layer for massive datasets.

2. Training Workflows: Optimizing for Large Data & Stable Convergence #

• Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
  If you’re dealing with large datasets, training stability is make-or-break. This paper demystifies why deep networks often struggle to converge, and introduces batch normalization as a solution. You’ll get a clear breakdown of how normalizing layer inputs fixes internal covariate shift, plus practical tips for integrating it into your DNN classifier training pipeline. No more tweaking learning rates endlessly just to get your model to converge.
• Adam: A Method for Stochastic Optimization
  Adam is everywhere, but do you know why it works so well for large-scale DNN classification? This paper walks through the design logic behind combining momentum and adaptive learning rates, with side-by-side comparisons of Adam against other optimizers on classification tasks. You’ll learn how to tune Adam’s parameters for your specific dataset, rather than just using default settings.

3. Advanced: Scaling DNN Classifiers Without Over-Stacking #

• Deep Residual Learning for Image Recognition
  Yes, it’s a CNN paper—but the residual connection idea is game-changing for fully-connected DNNs too. It solves the "degradation problem" where adding more layers actually hurts performance, by introducing residual paths that let the network learn identity mappings easily. You’ll understand how to build deeper, more expressive DNN classifiers without falling into the "stack more layers" trap, which is perfect for complex classification tasks on big data.
• Squeeze-and-Excitation Networks
  The SE module’s attention mechanism can be adapted to fully-connected DNNs to boost classification accuracy. This paper shows how to add a lightweight attention layer that lets your model automatically prioritize features relevant to your classification task. It’s a great way to improve performance without drastically increasing network size or training time.

These papers all focus on the fundamental logic behind DNN design and training, not just superficial layer stacking. They’ll give you the insights to build a DNN classifier tailored to your large-scale classification task.

内容的提问来源于stack exchange，提问作者DieDen