Skip to Content
Στείλτε μας email με τα σχόλια και τις ερωτήσεις σας

Bridging the gap between academia and industry

Research Framework

Research Areas

Previous Next

Learning with minimal supervision

While supervised learning requires large annotated datasets for model training, machine learning with minimal supervision uses unlabeled data and requires minimal human intervention. Minimal supervision approaches, such as semi-supervised, self-supervised, active learning, and uncertainty modeling, use other learning frameworks to improve accuracy. These methods are relevant where manual labeling is time-intensive or costly.

Few-shot learning

We develop few-shot learning approaches that work when limited data is available for model training. Our approaches include using multimodal data that leverages both images and text, as well as cross-modal hallucination and meta-learning methods. Few-shot learning approaches are relevant in all business scenarios with limited training data, such as the classification of new products in online product catalogs.

Visual question answering

Our integrative visual question answering (VQA) models improve the detection of fine-grained information and VQA evaluation metrics, enabling computers to answer questions about an image using natural language. The application of VQA models spans various industries and ranges from integration in intelligent chatbots to ticketing and invoice processing systems, as well as intelligent information retrieval for disease diagnoses.

Efficient deep learning

State-of-the-art deep learning models involve extensive computational costs and expensive hardware. We develop novel approaches to resource-efficient deep learning, including assessing model complexity, resource-efficient networks, quantization, pruning, and knowledge distillation. Efficient deep learning approaches in industrial applications help minimize monetary cost, power consumption, inference time, and environmental impact.

Privacy and fairness

We develop approaches to machine learning that impose privacy and fairness constraints, such as differential privacy and federated learning, as well as multitask learning. This enables institutions to use generalized diagnostic or prediction models without risking individuals’ privacy. Developing algorithms with fairness constraints also helps mitigate unconscious bias in applications such as credit scoring, bank loans, or résumé matching. 

Lifelong learning

We investigate lifelong learning approaches that enable machine learning models to learn the way humans do, incrementally and by utilizing previous knowledge to learn new tasks. Our methods include class-incremental approaches, continual domain adaptation, neuroplasticity, and adaptive capacity expansion. Machine learning models that learn continuously are relevant in all situations where it is not possible to keep historical training data.

Sentiment analysis

Understanding the sentiment and opinions expressed in natural language is a key challenge in natural language processing. We work on novel approaches for sentiment analysis that focus on neural word embedding and attention-based methods. Sentiment analysis is relevant for business sectors, such as telecommunications, banking, insurance, and e-commerce, where opinions about products and services must be analyzed and acted on.

Interpretable machine learning

Machine learning models and algorithms have reached an outstanding level of sophistication, making it challenging to explain predictions. Interpretable machine learning approaches, such as self-supervised and meta-supervised learning and curiosity-driven models, enable the discovery of patterns in data. This is crucial in business applications as it provides transparency and helps explain the underlying reasons for proposed outputs. 

Information extraction

Extracting information from structured and unstructured documents is challenging for natural language processing (NLP), and finding labeled data to train machine learning models is difficult. Our research focuses on sequential and two-dimensional approaches for structured documents, combines elements from NLP with computer vision, and incorporates unsupervised and weakly-supervised models using transfer learning into information extraction pipelines.


Back to top