Efficient CNN Model Compression & Confidence-Aware Risk Optimization

This research explores the trade-off between computational efficiency and predictive reliability in compressed Convolutional Neural Networks (CNNs). By integrating model compression techniques with a Genetic Algorithm–based risk weighting engine, the system identifies deployment-ready models optimized for real-world resource-constrained environments.

Project Overview

Modern CNNs deliver state-of-the-art accuracy but demand significant memory and computation. This project evaluates structured and unstructured compression strategies to reduce model size and latency while preserving reliability.

Pruning — Remove redundant weights
Quantization — Reduce numerical precision
Low-Rank Factorization — Decompose weight matrices
Knowledge Distillation — Transfer knowledge from teacher model
Weight Sharing — Cluster and reuse parameters

System Specifications

DatasetMNIST (28x28 grayscale digits)

ArchitectureStandard CNN (Conv → ReLU → Pool → FC)

Compression Methods5 Techniques

Optimization EngineGenetic Algorithm

Fitness ObjectiveNormalized FAR / FRR Log-Risk

Confidence-Aware Risk Weight Evolution

The system optimizes the trade-off between False Acceptance Rate (FAR) and False Rejection Rate (FRR) using a Genetic Algorithm. A single weighting parameter α is evolved to minimize total normalized risk across all compressed models.

Risk(α) = α · log(FAR_norm) + (1 − α) · log(FRR_norm)

Log-scaling stabilizes optimization, preventing dominance from skewed distributions. Interior regularization and soft boundary penalties prevent collapse toward extreme α values.

Max Generations: 200

Population Size: 20

Elite Retention: 4

Tournament Selection

Crossover Blending

Adaptive Mutation Decay

Diversity Injection

Early Convergence Detection

Evolution Strategy Design

The Genetic Algorithm evolves α using controlled stochastic search:

• Tournament-based parent selection
• Elitism to preserve top solutions
• Blended crossover for smooth interpolation
• Adaptive mutation with sigma decay
• Diversity reinjection when population collapses
• Stagnation-based recovery mechanism

Convergence occurs when improvement falls below tolerance and stagnation persists across generations.

Evaluation Criteria

Accuracy on Clean Inputs

Robustness to Noise

Model Size Reduction

Inference Latency

FAR / FRR Trade-off

Entropy Stability

Confidence Calibration

Deployment Feasibility

Research Objective

The objective is to identify compression strategies that minimize computational cost while preserving predictive reliability. The risk-weight evolution mechanism ensures deployment-aware decision calibration for financial-grade validation systems.

Research Team

Sai Vikas

CB.EN.U4CSE22363

Albert

CB.EN.U4CSE22505

Rathna

CB.EN.U4CSE22526

Mukesh

CB.EN.U4CSE22531