True robopunk!Physicists used loudspeakers to build a neural network described today in Nature

2022-05-12 0 By

Xiao check number from the sunken the temple qubits | public QbitAI in speaker recognition of handwritten Numbers?It sounds metaphysical, but this is actually a serious Nature paper.The image below, which looks like a modified speaker, was used to recognize handwritten numbers almost 90 percent of the time.That’s what physicists from Cornell University have done.They have built acoustic, electrical and optical versions of physical neural networks (PNN) using loudspeakers, electronics and lasers.These neural networks can also be trained by backpropagation algorithms.The reason physicists came up with PNN is that Moore’s Law is dead, and we need to save machine learning with physical systems.According to the article, PNN promises to increase the energy efficiency and speed of machine learning by orders of magnitude compared to software-implemented neural networks.Scientists can build neural networks with physical devices because the essence of physical experiments and machine learning is the same — tuning and optimizing.There are many nonlinear systems in physics (acoustic, electrical, optical) that can be used to approximate arbitrary functions in the same way as artificial neural networks.That’s what acoustic neural networks do.The two postdocs removed the diaphragm above the speaker and attached a square titanium plate to the speaker’s moving coil.A feedback loop is created by receiving the control signal from the computer and the input signal generated by the vibration of the metal plate, which is then output to the speaker.As for how to do back-propagation, the authors propose a hybrid physical world and computer algorithm called “Physical Awareness Training” (PAT) that can back-propagate a general framework for directly training any physical system to perform deep neural networks.In the acoustic neural network system, the vibration plate receives the sound input sample (red) modified by MNIST image. After driving the vibration plate, the signal is recorded by the microphone (gray) and converted into the output signal (blue) in time.The process of the physical system is as follows: Digital signals are converted to analog signals and input to the physical system. Then, the output is compared with the real result. After reverse propagation, parameters of the physical system are adjusted.Through repeated debugging of the speaker parameters, they achieved 87% accuracy on the MNIST data set.You may ask, what’s the advantage of using a computer during training?Indeed, PNN may not be advantageous in training, but the operation of PNN depends on the laws of physics. Once the network training is completed, there is no need for computer intervention, and PNN has advantages in reasoning delay and power consumption.Moreover, PNN is much simpler in structure than the software version of neural network.There are electrical and optical versions as well as acoustic versions, researchers have also built electrical and optical versions of neural networks.Electrical edition uses four electronic components resistance, capacitance, inductance and three tubes, just like the middle school physics experiment, the circuit is extremely simple.This analog circuit PNN can perform MNIST image classification with 93% accuracy.The optical version is the most complicated, with the near-infrared laser converted to blue light through a frequency-doubling crystal, but this system has the highest accuracy of 97 percent.In addition, the optical system can easily classify speech.PAT, the physical system training algorithm used above, can be applied to any system. You can even use it to build fluid or mechanical punk neural networks.Reference links:[1][2][3]