In many machine learning applications—such as robotic learning—there is limited data for a given task or platform. A natural solution is to train a shared model to extract generally useful information across tasks. However, standard deep learning optimization algorithms can fail to find good solutions, even in the simplest settings.
The Challenge:
Multi-task transfer learning is a key desideratum in large-scale and safety-critical applications, yet unforeseen issues in their optimization can compromise performance and safety. Thomas Zhang, our 5th-year PhD student (advised by Nikolai Matni), identifies these fundamental failures and introduces a family of algorithms that provably fix them.
Why This Matters:
These failures had previously flown under the radar due to ubiquitous but unrealistic assumptions in prior theoretical work. By establishing a new perspective to design and analyze multi-task learning algorithms, this work takes a crucial step toward making safety-critical applications more reliable and trustworthy.
Beyond Robotics & Autonomous Systems:
Our insights extend beyond self-driving cars and robotics—they apply broadly across deep learning. We identify a bottleneck in neural network optimization that affects many architectures, leading to potential advances in:
• New deep learning optimizers
• Better understanding of normalization layers
• Improved representations for multi-modal vision/language models
Read his full paper here: https://lnkd.in/exB4H5W3