Situation
You have data such as images, signals, texts, or documents and want to use machine learning or deep learning to extract meaningful trends or develop a predictive model, but you don’t know where to start or which resources to use at each step of the process.
Here are guidelines to properly define, prepare, and execute your machine learning project.
→ Define your research question within the ML/DL paradigm.
Translate your scientific question into a learning task: classification, regression, segmentation, clustering, or generation. Identify your outcome variable (what you are trying to model/predict), the type of label or its absence (the label is the expected value), and specify whether your task is supervised, self-supervised, or unsupervised. Also evaluate what has been done in the literature for your specific case.
Need help?
- IID: https://iid.ulaval.ca/
- PREDIS: https://www.predis.ca/
- Université Laval: https://www.ulaval.ca/
→ Collect, organize, and store your data.
Organize your raw data (images, signals, texts, documents) into a reproducible folder structure. Apply appropriate anonymization measures and obtain necessary ethical approvals before storage. Choose a platform that meets your needs for storage volume, access, and regulatory requirements (HIPAA/GDPR, if applicable).
→ Prepare and process your dataset.
Clean, normalize, and transform your data into a format suitable for learning according to the task and data type. Split the data into training, validation, and test sets from the start and before any model-related decisions to avoid data leakage.
→ Select and design your model architecture.
Choose an architecture suited to your data type: convolutional neural networks (CNN) or Vision Transformers for images, recurrent neural networks (RNN) or Transformers for sequential signals, large language models (LLM) or BERT-like encoders for text. Prefer pre-trained models and fine-tuning when labeled data is scarce. Document all architectural choices and hyperparameters to ensure reproducibility. Beware of common pitfalls: overfitting, class imbalance, data leakage.
Need help?
- IID: https://iid.ulaval.ca/
- PREDIS: https://www.predis.ca/
→ Train your model on appropriate computing infrastructure.
Transfer training to GPU-equipped infrastructure when your dataset or model becomes too large for a local machine.
See use case: Using CPU/GPU resources.
→ Evaluate, interpret, and validate your results.
Report metrics appropriate to your task (AUC-ROC, F1, Dice, MAE, etc.) only on the held-out test set. Use explainability tools (Grad-CAM, SHAP, attention maps) to verify the consistency of what the model has learned. Validate results against clinical or scientific ground truth.
