Şekilsel Matematiksel ve Deneysel Boyutları Birleştiren Derin Öğrenme Atlası: Blok Tabanlı CNN Mimari Tasarımı ve Kıyaslama
Göktaş Ö. F. (Yürütücü), Doğan Ş., Tuncer T.
Yükseköğretim Kurumları Destekli Proje, BAP Araştırma Projesi, 2026 - 2028
- Proje Türü: Yükseköğretim Kurumları Destekli Proje
- Destek Programı: BAP Araştırma Projesi
- Başlama Tarihi: Ağustos 2026
- Bitiş Tarihi: Şubat 2028
Proje Özeti
This project aims to transform the largely intuition-driven and trial-and-error-based architectural design approach of convolutional neural networks (CNNs) into a systematic engineering process that is governed by well-defined rules, mathematically traceable, and reproducibly generative. The project’s originality lies in addressing CNN architectures not through the conventional approach of proposing a single model, but through a block-based design language that enables bidirectional translation between diagrammatic representations and mathematical formulations. In this way, existing architectures will be analytically decomposed, while new architectures will be derived from mathematical definitions in a justified and controlled manner. This framework will integrate CNN design into a “design–generation–benchmarking” cycle based on the principles of explainability, reproducibility, and measurability.
The proposed method will standardize CNN architectures using four fundamental building blocks: stem, main, downsampling, and outcome. In the first stage, these blocks will be developed as reusable components with clearly defined layer types, dimensional transformations, parameter counts, and computational costs. In the second stage, each block will be modeled as a mathematical operator based on the “input → transformation → output” principle. Sequential connections will be represented through function composition, while parallel pathways and residual connections will be modeled using appropriate aggregation operations. In the third stage, the established formalism will be used to derive novel blocks, with the rationale, expected benefits, and conditions of use for each derivation documented through a standardized template. In the fourth stage, the block repository will be transformed into modular and reconfigurable CNN families at different scales—lightweight, medium, and large. The objective will not be limited to improving accuracy but will also encompass the systematic exploration of accuracy–cost trade-offs. In the fifth stage, all architectures will be compared using a standardized training and evaluation protocol. Performance results will be reported not only in terms of accuracy metrics but also alongside cost indicators such as parameter count, FLOPs, memory consumption, and training time.
Project management will follow a stage-gated plan designed to generate verifiable outputs at the end of each stage: (i) a standardized block library, (ii) mathematical representations and transformation rules, (iii) a repository of derived blocks, (iv) scalable architecture families, and (v) benchmarking reports. Control points will be established to address computational cost, the practical implications of the derived structures, and scalability risks. Through a progressive selection strategy, computational resources will be directed toward the candidates with the highest potential impact.
The project’s broader impact will be delivered through its flagship output, the Interactive CNN Atlas. The Atlas will present each architecture comprehensively on a single platform through an integrated combination of diagrammatic schematics, mathematical definitions, and experimental results. It will accelerate architectural comparisons and facilitate the design of new CNN architectures. Its open and extensible structure will strengthen reproducibility, while its educational component will contribute to the training of future researchers. Ultimately, the project will transform CNN architecture design from a practice dependent on individual expertise into a shared and sustainable knowledge infrastructure guided by measurable rules, thereby generating lasting scientific value at both national and international levels.