The common feature between forest fire and nerve networks shows the universal framework

Researchers at the University of Tokyo, in collaboration with Isin Corporation, have proven that Universal Scaleting rules, which describe how the system features change with size and scale, apply to deep nervous networks that display emotional phase transfer behavior, which is commonly seen in physical system. This discovery not only provides a framework that describes deep nervous networks, but also helps them predict their training or general capacity. These results were published in the journal Physical review research.

In recent years, it does not matter where we look, we come to artificial intelligence in some form. The current version of the technology is powered by deep nerve networks: several layers of digital “neurons” are with the weight of the weight between them. The network learns by modifying the weight between “neurons” unless it produces the right production. However, a united theory that describes how the signal spreads between layers of neurons in the system has so far eliminated scientists.

“Our research was encouraged by two drivers,” says Kachi Tamai, the first author. “Partially, according to industrial needs, because the brutal force of these large -scale models affects the tuning environment. But there was another, deep pursuit: the scientific understanding of the intelligence physics itself.”

This is the place where Tamai’s background in the statistical physics of Phase Transition gave him the first hint. The absorbent phase transitions indicate an emotional phase at a rapid speed at an active location, where the system cannot escape without external help. An example of such a physical system will be flared.

Significantly, these systems display universal behavior near the Tipping Point and can be described using globally scaling laws if some features are preserved. If the deep neurological network absorbing phase transfer, then Universal Scaleing Rules may apply, providing a united framework to describe how they work. As a result, researchers will be able to predict whether the signal will “burn” in a particular deep learning.

Investigation, the researchers linked the theory with imitation. They derived the evacuation, which are universal in the systems, and scaling factors, which are different in the system, when possible when possible and in more complex cases can be used to verify the scaling laws.

“What is coincidental, I thought,” says Temi, when he first remembered the links between the deep nervous networks and the transfer of the absorbing phase. “I never thought I would do research on deep learning, as a doctorate student in physics, let me find effective use of this concept.”

This detection also draws us closer to understanding the physics of intelligence, as it reinforces the assumption of brain criticism, which says that some biological networks operate near the phase transfer. The Tamai research line is excited about the possibility.

“Alan Touring indicated in this regard in the early 1950s, but the tools were not ready at that time. With the rapid accumulation of evidence in neuro science and the rise of nearby human level AI, I am sure we are at a great moment to revise and deepen our understanding of this basic relationship.”