AI chip that embodies hidden neural...

  1. Home
  2. Article

13
10

AI chip that embodies hidden neural network theory

The implementation of image recognition technology that applies deep learning (deep learning), which is one of the artificial intelligence (AI) technologies, is expanding, and security such as the jumping out detection function of autonomous cars, work control of robots, face authentication, etc.It is expected to be used in fields.In order to instantly process images and images captured by the camera, it is necessary to install a computer on an edge device such as automobiles and robots, and to calculate and judge by AI on the spot, but explosive calculation.Increasing and accompanying power consumption is an issue.In particular, it is difficult to supply power from the outside, especially for mobile devices such as automobiles and drones, so it is necessary to drive a computer with as little power as it maintains inference accuracy.

隠れニューラルネットワーク理論を具現化したAIチップを世界

Deep learning determines the situation from information such as images and images by a information processing model that imitates a human brain called a deep neural network (DNN).The fact that the structure of this network becomes complicated and huge is a huge factor in the amount of calculation, especially when reading a calculation parameter such as "weight" of the DNN model from an external memory.

As a revolutionary technology that realizes a lightweight DNN model, in 2020, "Hidden Neural Network) discovers a partial network that does not deteriorate the reasoning accuracy even if only a part of the DNN is used. The theory was newly announced (Fig. 1). Unlike the conventional deep learning method, this theory is fixed at the initial value of the random number without learning the weight. Instead, learn the "score" that represents the importance of each bond of the network, and use a partial network with only the top K%(K is an arbitrary number), so that the overall size of K%is lightweight DNN. Build a model. Specifically, the value 1 is corresponded to the binding of the top score to be selected, and the other "super mask" is generated with value 0, and the weight of the random number initial value and the supermark of the super mask. By removing, the top -ranked partial network can be excavated.

In this study, focusing on the latest hidden neural network theory, (1) not learning the weight, (2) discovering a partial network with a supermask, the existing DNN has not had the two new features.A DNN inference accelerator that can be used is realized.

Tags How to build drones from zero

Category

Hot Articles

Tags

Software cyber IT Computing Consulting Business