Deep Learning

The Problem

Deep learning underpins much of modern AI, but fundamental challenges remain:

• Data efficiency — Current models require enormous amounts of labeled data. Learning from fewer examples — or from unlabeled data — is critical for many real-world applications.
• Architecture design — Finding the right network architecture for a given task is often driven by trial and error rather than principled design.
• Temporal modeling — Many real-world problems involve sequences (video, sensor data, user behavior). Capturing long-range temporal dependencies efficiently remains an open challenge.
• Generalization — Models that perform well on benchmarks may fail on slightly different real-world distributions. Understanding and improving generalization is foundational.

What We're Working On

• Sequence prediction — Bidirectional LSTMs and transformer-based models for continuous quality-of-experience prediction in streaming services and other temporal tasks.
• Convolutional architectures — Novel CNN designs for predictive tasks in multimedia and network traffic analysis.
• Representation learning — Techniques for learning robust feature representations that transfer well across tasks and domains.