The question behind the work.
How does a fully connected model compare with a convolutional model when the input contains local sequence structure?
Dataset & environment
Sequence-based latch classification task from academic coursework.
From the idea to the method.
- Define a consistent input representation and hold out validation and test splits.
- Compare a feedforward network with a 1D convolutional architecture under the same data and evaluation protocol.
- Inspect kernel size, receptive field, parameter count, and the gap between training and validation performance.
Different architectural assumptions
A fully connected layer can combine all input positions, while a convolutional layer shares a local kernel across the sequence. Kernel size determines the immediate receptive field; stacking convolutions expands it. Neither inductive bias guarantees a better result without a controlled comparison.
Sequence → Flatten → Dense layers → Classifier
Sequence → Conv1D → Activation → Pooling → Classifier
# Architectural example, not a reproduced experiment
import torch.nn as nn
cnn = nn.Sequential(
nn.Conv1d(1, 16, kernel_size=3, padding=1),
nn.ReLU(),
nn.AdaptiveAvgPool1d(1),
nn.Flatten(),
nn.Linear(16, 2),
)What this study explores.
- Academic focus: feedforward and convolutional neural networks.
- PyTorch model development, architecture comparison, and generalization analysis.
How to assess it carefully.
Use the same test split, compare classification performance and model complexity, and report seed variability when experiment logs are available.
What is documented.
Experimental artifacts and verified numerical results have not been published. The interactive visuals explain the method; they are not benchmark results.
Training curves, accuracy comparisons, and repository links will be added alongside the corresponding artifacts. No unverified benchmark numbers are shown.
A foundation for the next experiment.
- Dense layers learn global combinations; convolution introduces locality and parameter sharing.
- A larger training score alone does not establish better generalization.
Future updates can include code, screenshots, reproducibility instructions, and verified outcomes.