Exploring Layer-specific Quantization in CNN-based Selective Sweep Detection

This study explores the complex field of population genetics known as "selective sweep detection," in which the fast spread of particular genetic variations within a population creates patterns in the genome. The ASDEC framework and the novel SweepNet architecture have made noteworthy advancements in genomic scan sensitivity, success rates, and detection accuracy. However running these CNNs requires a lot of resources, which presents a significant obstacle. Extensive experiments revealed that quantizing the layers separately produces encouraging outcomes. The majority of layers show a notable decrease in the required precision; some layers only require 2 bits, and in other circumstances, 1 bit is sufficient. By optimising resource utilisation and decreasing the memory requirement of these CNNs' operations, the layer-by-layer quantization technique makes these models run faster with fewer available resources.