Research Focus

AutoML & Few-Shot Learning

Automating machine learning to build AI systems that adapt rapidly with minimal data

What is AutoML?

AutoML (Automated Machine Learning) aims to automate the end-to-end process of applying machine learning to real-world problems. From data preprocessing to model selection and hyperparameter tuning, AutoML reduces the need for manual intervention, making machine learning accessible to non-experts and accelerating experimentation for researchers.

Few-Shot Learning complements this by enabling models to generalize from very few labeled examples, mimicking human-like rapid adaptation. Together, they form a powerful toolkit for building versatile AI systems that learn quickly and deploy efficiently.

AutoML Pipeline Phases

Stages where automation accelerates the machine learning lifecycle

Data

Automatic Preparation

Cleaning, imputation, normalization, and feature engineering handled automatically to ensure models receive high-quality inputs without manual preprocessing.

Selection

Architecture Search

Automated exploration of model architectures and algorithms to find the best fit for the task, reducing the trial-and-error burden on practitioners.

Optimization

Hyperparameter Tuning

Intelligent search through hyperparameter spaces to maximize performance without exhaustive manual grid searches.

Deployment

Rapid Adaptation

Few-shot techniques allow models to adapt to new tasks or domains with minimal additional training, enabling faster deployment in changing environments.

Few-Shot Learning Foundations

Approaches to learning from minimal examples

Meta-Learning

Learning to learn: models acquire generalizable strategies across tasks so they can solve new problems from just a handful of examples.

Transfer Learning

Leveraging pre-trained representations from large datasets to bootstrap learning on smaller, domain-specific tasks with limited annotations.

Prototypical Networks

Metric-learning approaches that compare query examples to class prototypes learned from few samples, enabling efficient classification without retraining.

Challenges in AutoML & Few-Shot Learning

Open problems limiting widespread adoption

Search Space Explosion

The number of possible architectures, hyperparameters, and pipelines grows exponentially. Efficient search strategies remain an active research frontier.

Domain Transfer

Solutions found on one dataset rarely transfer seamlessly to others. Generalization across domains is still a major hurdle for AutoML systems.

Computational Cost

Automated search is computationally expensive. Balancing exploration depth with time and resource constraints is a critical practical challenge.

Practical Applications

Domains where AutoML and few-shot learning accelerate AI adoption

Healthcare Diagnostics: Rapid deployment of medical imaging classifiers at new hospitals with minimal annotated data, adapting models across different scanners and patient populations.
Precision Agriculture: Automated crop-disease detection systems that generalize to new regions and seasons with few local examples, reducing the need for extensive field data collection.
Industrial Quality Control: AutoML pipelines that configure defect-detection models for new manufacturing lines without expert tuning, cutting setup time from weeks to hours.
Natural Language Interfaces: Few-shot adaptation of large language models to new domains (legal, medical, technical) using only a handful of labeled examples per intent.
Scientific Discovery: Accelerating hypothesis testing by automatically selecting and tuning models for novel datasets in materials science, genomics, and drug discovery.
Edge and Mobile AI: Lightweight AutoML-generated models deployed on smartphones and IoT devices for real-time speech recognition, gesture detection, and sensor analytics.

Our Research

In the AutoML & Few-Shot Learning line of the GPAIS group at IAFER, we focus on making machine learning faster, more accessible, and capable of adapting to new tasks with minimal supervision. Our goal is to lower the barrier to building high-performing AI systems.

🚀 Active Research Lines

Neural Architecture Search (NAS): Automated discovery of efficient architectures tailored to specific tasks and hardware constraints.
Meta-learning for rapid adaptation: Training models that can quickly specialize to new domains with few examples.
AutoML for structured data: Automation pipelines optimized for tabular and time-series data common in industry.
Lightweight deployment: Compression and optimization techniques to make AutoML-generated models feasible for edge devices.

📚 Publications

Loading publications...