A Transparent Baseline CNN for CIFAR-100: Design and Pilot Evaluation in a Reproducible Learning Framework

Madhan Raj Gopi Akila, Charan Tharak Varma, Sai Charan Reddy, A. Abinay, Prasanna Ranjith Christodoss, Andino Maseleno Department of Computer Science, University of Stuttgart, Stuttgart, Baden-Württemberg, Germany. Department of Computer Science and Engineering, SRM Institute of Science and Technology, Ramapuram,  Chennai, Tamil Nadu, India. Department of Computing, Mathematics and Physics, Messiah University, Mechanicsburg, Pennsylvania,  United States of America. Department of Information Systems, Institut Bakti Nusantara, Pringsewu, Lampung, Indonesia.

Abstract: CIFAR-100 remains a useful benchmark for image classification because it is small in image size yet challenging in class diversity. In this paper, researchers document the design and pilot evaluation of a compact CNN built for an academic Neural Networks and Machine Learning study. A central aim of the draft is transparency: background ideas from the literature are clearly cited, the model design is described in the authors’ own words, and the reported results are limited to what was actually run. The proposed baseline uses stacked 3x3 convolutions, batch normalisation, ReLU, max pooling, dropout, and a fully connected classifier with 690,308 trainable parameters. Training uses normalisation, random cropping, horizontal flipping, cross-entropy with label smoothing, and AdamW optimisation. In a five-epoch pilot run on a 5,000-image training subset and a 1,000-image test subset, the model achieved 13.1% top-1 accuracy and 36.9% top-5 accuracy. These numbers are intentionally presented as early benchmark evidence for a reproducible student pipeline, not as competitive final performance. 

---

Keywords: Pilot Evaluation; Max Pooling; Reproducible Learning Framework; Machine Learning (ML); Random Cropping; Horizontal Flipping; Visual Patterns; Image Classification.

---

Received on: 08/07/2025Revised on: 15/09/2025Accepted on: 20/10/2025Published on: 03/03/2026

---

---