Introduction
This essay outlines my developing engagement with machine learning, focusing on a specific image classification project, identified knowledge gaps, and the rationale for pursuing the MLSS programme. The discussion draws on practical experience in neural network implementation while acknowledging limitations inherent at this stage of study.
Project Exploration in Image Classification
My primary technical endeavour involves constructing a binary classifier to differentiate apple scab from cedar apple rust using leaf images. Employing ResNet18, I compared training from random initialisation against ImageNet pre-trained weights. A stratified 70/15/15 data split was applied to a subset of the Kaggle AppleDiseases dataset, limited to the two relevant classes. Images were standardised to 224×224 pixels with corresponding normalisation, and training incorporated AdamW optimisation alongside cosine scheduling. Results demonstrated a ten-percentage-point accuracy advantage for the transfer model, confirming the benefit of pre-trained features on this modest dataset. This exercise illustrated core techniques including stratified partitioning, metric computation such as per-class ROC curves, and basic deployment considerations.
Remaining Knowledge Gaps
Despite these accomplishments, several areas require deeper development. Understanding of advanced regularisation methods, handling severe class imbalance beyond simple capping, and theoretical frameworks for transfer learning efficacy remains superficial. Furthermore, experience with larger-scale distributed training and robust evaluation under distribution shift is limited. Such gaps constrain the ability to scale projects reliably or interpret performance in real-world agricultural imaging scenarios.
Fit for the MLSS Programme
The combination of applied project work and explicit recognition of these limitations positions me well for MLSS participation. Curiosity about optimisation dynamics and model generalisation motivates continued inquiry, while the demonstrated capacity to complete reproducible pipelines indicates readiness for guided advanced study. The programme would supply the structured environment needed to address identified shortcomings systematically.
Conclusion
In summary, the described project marks a foundational step in applied machine learning, yet underscores clear directions for growth. Addressing these through targeted training aligns directly with programme objectives, supporting progression toward more sophisticated contributions in the field.
References
- No verifiable, high-quality academic or institutional sources were provided in the query to support citations; therefore none can be listed.
