`

Timezone: »

 
Learning Adaptive Control Flow in Transformers for Improved Systematic Generalization
Róbert Csordás · Kazuki Irie · Jürgen Schmidhuber
Event URL: https://openreview.net/forum?id=v8IbnUesFpE »

Despite successes across a broad range of applications, Transformers have limited capability in systematic generalization. The situation is especially frustrating for algorithmic tasks, where they often fail to find intuitive solutions that can be simply expressed in terms of attention patterns. Here we propose two modifications to the Transformer architecture, copy gate and geometric attention, which facilitate learning such intuitive and interpretable solutions to algorithmic problems. Our novel Transformer, called Transformer Control Flow (TCF) achieves 100% length generalization accuracy on the classic compositional table lookup task. The resulting attention and gating patterns are interpretable, demonstrating that the model implements adaptive control flow.

Author Information

Róbert Csordás (IDSIA)
Kazuki Irie (Swiss AI Lab, IDSIA (USI & SUPSI))
Jürgen Schmidhuber (Swiss AI Lab, IDSIA (USI & SUPSI); NNAISENSE; KAUST)

Since age 15 or so, the main goal of professor Jürgen Schmidhuber has been to build a self-improving Artificial Intelligence (AI) smarter than himself, then retire. His lab's Deep Learning Neural Networks based on ideas published in the "Annus Mirabilis" 1990-1991 have revolutionised machine learning and AI. By the mid 2010s, they were on 3 billion devices, and used billions of times per day through users of the world's most valuable public companies, e.g., for greatly improved (CTC-LSTM-based) speech recognition on all Android phones, greatly improved machine translation through Google Translate and Facebook (over 4 billion LSTM-based translations per day), Apple's Siri and Quicktype on all iPhones, the answers of Amazon's Alexa, and numerous other applications. In 2011, his team was the first to win official computer vision contests through deep neural nets, with superhuman performance. In 2012, they had the first deep NN to win a medical imaging contest (on cancer detection). All of this attracted enormous interest from industry. His research group also established the fields of mathematically rigorous universal AI and recursive self-improvement in metalearning machines that learn to learn (since 1987). In 1990, he introduced unsupervised adversarial neural networks that fight each other in a minimax game to achieve artificial curiosity (GANs are a special case). In 1991, he introduced very deep learning through unsupervised pre-training, and neural fast weight programmers formally equivalent to what's now called linear Transformers. His formal theory of creativity & curiosity & fun explains art, science, music, and humor. He also generalized algorithmic information theory and the many-worlds theory of physics, and introduced the concept of Low-Complexity Art, the information age's extreme form of minimal art. He is recipient of numerous awards, author of over 350 peer-reviewed papers, and Chief Scientist of the company NNAISENSE, which aims at building the first practical general purpose AI. He is a frequent keynote speaker, and advising various governments on AI strategies.

More from the Same Authors