Station B boss used my world to create convolutional neural network, Lecun forwarding! Burst the liver for 6 months, playing more than one million

Abundant color From the Aofei temple qubits | official account QbitAI

Not a computer major 、 Neural network Xiaobai , Liver explosion for six months ——

stay 《 My world 》 Build a convolutional neural network , What kind of divine operation is this ？

‍

‍

lately , come from B standing up Lord @ Chen zhanaotou's works are really popular .

He worked with his friends to complete this so-called “ The world's first pure Redstone neural network ”, Neuron 、 Convolution layer 、 Fully connected layer 、 Activation function 、 Multiplier 、 Input 、 Output …… Everything 、 spectacular , And it can really Realize handwritten digit recognition , The accuracy rate has also reached 80%.

This wave of , Totuo is what netizens say ：

It is simply an extraordinary result of strength and patience .

at present , The volume of this video has reached onemillion , On the Internet , It even caused LeCun Attention ： Not only forwarded their works , It also gives “Very meta” The evaluation of .

Time consuming 6 Months , Use red stone to build convolutional neural network

Red stone is a kind of mineral resource that can transmit signals in my world , It can be used to make red stone circuits , And then complete as small as automatic door 、 Optical switch 、 Simple machine of stroboscopic power supply , As big as the elevator 、 Automatic farm 、 Shield machine 、 Small game platform and even complex tools of computer .

The structure of this neural network built by the authors , be based on LeCun On 1998 Classical convolution structure proposed in LeNet-5, It is used to realize handwritten digit recognition .

Compared with the traditional full accuracy calculation （ Multipliers and adders ）, After some thinking and estimation , It is decided to implement this neural network by random calculation , This will make the design and layout simpler .

After all, for multiplication with random computation , A single pole can be represented by an and gate , A homo or can represent bipolar .

Because it is impossible to carry out back-propagation in my world , The weight of the network is first Pytorch Good training , Then move in directly .

To generate weights consisting of random strings , The author uses “ Throwers throw objects randomly ” This principle creates a random number generator .

in general , They use a compressed LeNet-5, First use a weighted window （ Convolution kernel ） Scan the image step by step and extract stroke features , Then these stroke features are fed into the deep neural network （ Fully connected layer ） Classify and identify .

say concretely ：

First, the input device ： A single pulse pressure board, handwriting board and 15×15 Coordinate screen , Generate coordinate signals , And draw the handwriting on the screen .

Then handwritten digits enter the convolution layer , Accumulate the covered part of convolution kernel , And output the results to the next layer .

among ：

（1） In the convolution , The author did not use random calculation , Instead, I use analog signals in my world for addition ;

（2） In order to ensure that the input data can be nonlinear mapped to the high-dimensional feature space and then linear classification , The output passes through the activation function ReLU;

（3） Due to convolution, you can't move at will , Therefore, the direct stacking method is adopted , Then connect it to the tablet input through hardwire .

And then , It's a full connection layer . Each layer is composed of several neural networks , Each neuron is connected to multiple inputs , And produce an output . The neuron weights and accumulates each input , Then bring in an activation function output .

The full connection layer uses random computation ,

The activation function is nonlinear tanh.

The actual neuron circuit is as follows ：

The output of the last layer uses an analog electric counter , Used for statistical 5Hz In a string “1” The number of , The capacity is 1024.

Final , Output part , Counter high 4 Bits are connected to the counter board , Then the circuit selects the maximum value and displays the result on the panel .

An overview of the structure ：

Overview of network architecture ：

The authors introduce , The neural network is used in MNIST The data set is about 80% The accuracy of , As a contrast , The accuracy of the full precision network with the same weight is 88%.

in addition , Its single theoretical recognition time is about 5 minute , But I didn't think Minecraft Our computing power is really Limited —— In the actual test , May want 40 Minutes or more .

From this, the author concludes ,Minecraft Stochastic computational neural networks are not necessarily superior to full precision networks in terms of time overhead . However, no one has made a full precision network yet .

“ The workload and difficulty are great ”

In the comment area of this work , It's all praise and worship from a stream （ Lianda V The seeds are bubbling ）——

After reading the giant's masterpiece , Some netizens even began to doubt themselves and up The masters are not playing a game .（ Manual formation ）

It was also pointed out that , Although the final function is equivalent to that in machine learning “hello world”, But use the Redstone components provided in the game to reproduce , Can be said to be “ The workload and difficulty are great ”.

because “ This requires the author to understand the underlying implementation of the algorithm or the principle of hardware implementation （ similar cuda Programming ） Have a deep understanding , You can also use game mechanisms to optimize the execution process and complete parallel computing ”.

Although the final recognition speed is relatively slow , but “ It is of little significance to discuss efficiency here ”.

Between the lines , It is valuable in itself ,“ We can't take cpu The recognition efficiency of the two threads after countless layers of simulator nesting is compared with that of the graphics card ”.

Last , Others lament ： Okay , Now Redstone neural network has , Hard disk 、CPU、 Monitors have been around for a long time , Isn't Redstone supercomputer not far away ？

“ Maybe we can be in MC Play inside MC 了 ～”

About author

This red stone convolutional neural network has 5 authors ,up Lord @ Chen Zhan is the main contributor , Be responsible for the overall design of the circuit 、 Build and debug .

He and another author @ Learning miscellaneous things is not good （[email protected]） All are Hong Kong University of science and technology Of the students , Now study separately Doctor of theoretical physics and doctor of Electronic Engineering .

other 3 Of the partners , There is a high school graduate （@NKID00）, The other two （@enadixxoOxoxO and @ Little octopus who loves red stone ） Did not disclose his identity .

@ Chen zhanaotou is 2014 I began to contact in 《 My world 》 This game , I have made a display for coding Chinese characters and a 8 Bit CPU.

When asked why a non computer major student wants to build a neural network , He told us , In fact, I have participated in the information competition , Because I have been exposed to random calculation before （stochastic computing）, At first, I wanted to show the advantages of random computing in specific tasks, so I started this project .

Finally, in order to build a complete neural network , He looked at 50+ Page English Literature , Yes 1000 Many lines of code .

The hardest thing to say / The most time-consuming part , He replied ： yes Full connection layer debugging , Because the running speed is very slow and it is difficult to find the problem using random calculation .

In the future ,up The LORD said , He is considering building a Support RISC-V Redstone of instruction set CPU.

Checked the , No one seems to have done it yet , Worth waiting for ～

B Stop video ： https://www.bilibili.com/video/BV1yv4y1u7ZX?vd_source=6eb6d925760348954531a2288dcd74be

Principle introduction ： https://www.bilibili.com/video/BV1wF411F7PU/?spm_id_from=333.788&vd_source=6eb6d925760348954531a2288dcd74be

—  End  —

「 qubits · viewpoint 」 Live registration

What is? “ Intelligent decision making ”？ What is the key technology of intelligent decision ？ How will it build a leading enterprise for secondary growth “ Intelligent gripper ”？

7 month 7 On Thursday , Participate in the live broadcast , Answer for you ～

Focus on me here , Remember to mark the star ～

One key, three links 「 Share 」、「 give the thumbs-up 」 and 「 Looking at 」

The frontier of science and technology meets day by day ~