
The fonts from the first and last sites look the same. Probably the same person, just different domain names. 
Arduino is
These boards are not made by Arduino so you run the slight risk of incompatibility if the board vendor does not update the compiler for some reason.
The default Arduino boards contains AVR ATmega328 microprocessor e.g. Arduino Uno / Arduino Nano. To use the default microcontroller, simply install Arduino IDE, choose the board from the list, enter a program, compile and download. Many microcontroller boards are Arduino IDE but use the ARM processor etc. You must install a replacement compiler into the Arduino IDE. Example boards
To learn more: 
The STM32 is a family of 32bit Flash microcontrollers based on the Arm® Cortex®M processor produced by STMicroelectronics. The official toolchain is Keil, now owned by ARM. For no cost, one can use the GCC toolchain. Some website to find info on the chip:


This is the a good board for learning ARM. Cheap. Very powerful. Arduino compatible. The board has 2 parts: 1. Debugger (STLINK) 2. Actual MCU The form factor is the same for all Arduinocompatible Nucleo boards. The actual board you have depends on the MCU itself. And the MCU choice is largely determined by the Cortex version you need. In our case, we want to used the CortexM4 because it has DSP and SIMD. The STM32F446 MCU is the cheapest M4 at the time of writing. 
Notes from Chapter 2: Linear Algebra for Deep Learning of the "Deep Learning Book by Goodfellow, Bengio & Courville. and all that. 
Try replacing the circuit with a single gate.

How to check if two 4bit numbers are equal? First, check if every bit in the same position have the same value.B = 0011 Let A = {a_{3}a_{2}a_{1}a_{0}} and B = {b_{3}b_{2}b_{1}b_{0}}. Each column is treated separately. So a_{3} = b_{3} = 0, a_{2} = b_{2} = 0, a_{1} = b_{1} = 1 and a_{0} = b_{0} = 1 If we define x_{i} a signal that is true when a bit pair has equal value then x_{i} = (a_{i }= b_{i}) = a_{i} xnor b_{i} = (a_{i} · b_{i}) or (not a_{i} · not b_{i}). Then A is equal B because x_{3} · x_{2} · x_{1} · x_{0} = 1 · 1 · 1 · 1 = 1. Let's look A ≠ B in the following case A = 0001 B = 0011 Because a_{3} = b_{3}, a_{2} = b_{2}, a_{1} ≠ b_{1} and a_{0} = b_{0}, x_{3} · x_{2} · x_{1} · x_{0} becomes 1 · 1 · 0 · 1 which gives a false result. The following circuit does all that in an instant! Due to de Morgan's theorem, you an also built the circuit the following way. E = x_{3} · x_{2} · x_{1} · x_{0} = not (x_{3} + x_{2} + x_{1} + x_{0}) The NOR is more commonly used to detect all zero conditions. 
Collection (collecting? as if work in progress) of the simplest explanations possible: 
Random collection of mathless info on DNN.

It is shipping now. It is called Centriq.Intel has close to 100% of the server market. Competitors like AMD, IBM and Oracle are just rounding errors. Centriq may be able to eat into Intel's lunch if it can prove its claims that it can match Xeon's performance for less power.
References: 
Whether you're learning computer science, any field of engineering or just curious, you should get to know about the ARM processor. Not too long ago, learning assembly language means learning the x86 assembler. Even Tutorialspoint only has x86 in its assembly language tutorial. Well, not all computers are based on the x86 architecture. The following infographic from AMD (who makes x86 processors) show that ARM has grown while all other architectures combined has shrunk. Intel and ARM combined has 80% of the market share. Why is ARM so strong? Find out in my introductory chapter is the slides folder here. Enjoy the ride! 
An absolute value finder circuit always give a positive output. For an nbit input word, bit x_{n1} is the sign bit. This bit control the (n1) units of 2:1 multiplexers. If x_{n1} = 0, the input bits x_{n2:0} are passed unchanged to the adder. If x_{n1} = 1, the input bits x_{n2:0} are inverted before going to the adder. Effectively, x_{n1} selects either the original x or the one's complement of x. The carry in to the adder get x_{n1} itself. Therefore, if x_{n1} = 0, the adder performs X + 0, and if x_{n1} = 1, the adder performs X'+1. Since X' or the ones' complement plus 1 is the two's complement of X, the output of the adder is X if the original input X is negative. The sign bit of the output is always 0. So y_{n1} is permanently connected to GND. 
MFCC Procedure
From https://ccrma.stanford.edu/~unjung/mylec/mfcc.html The Mel filterbank


The hyperbolic tangent function is important in neural networks. The first solution comes from S. Ghaffari and S. Sharifian, "FPGAbased convolutional neural network accelerator design using high level synthesize," 2016 2nd International Conference of Signal Processing and Intelligent Systems (ICSPIS), Tehran, 2016, pp. 16. The second solution: R. A. CallejasMolina, V. M. JimenezFernandez and H. VazquezLeal, "Digital architecture to implement a piecewiselinear approximation for the hyperbolic tangent function," 2015 International Conference on Computing Systems and Telematics (ICCSAT), Xalapa, 2015, pp. 14. A. Tisan, S. Oniga, D. Mic, and A. Buchman, “Digital Implementation Of The Sigmoid Function For FPGA Circuits,” ACTA Tech. NAPOCENSIS, vol. 50, no. 2, 2009. 
Google product applications of deep learning #machinelearning 
Random collection of mathless links

Random mathless info for mere mortals.

This is my version, for my Windows 10 machine, with GPU.
http://us.download.nvidia.com/Windows/390.77/390.77desktopwin1064bitinternationalwhql.exe
C:\ProgramData\NVIDIA Corporation\CUDA Samples\v8.0\bin\win64\Release
To cuDNN, joing the NVIDiA developer network at https://developer.nvidia.com/cudnn. Then choose cuDNN v6.0 from the list The Tensorflow documentation says "the cuDNN version
must match exactly: TensorFlow will not load if it cannot find cuDNN64_6.dll ."

We describe a programmble timer which produces a pulse with a length determined by the binary value at input x. The comparator outputs a reset pulse to the SR latch when counter value equals X.A 50 MHz system clock is assumed. The Trig signal initiates timer operation. It is connected to 3 points in the module. At the timing reference, Trig loads 50 million into the 26bit accumulator. At the counter, Trig clears the counter to 0. At the output SR latch, Trig sets Out to high. The timing reference produces a pulse one clock cycle wide every second. The pulse, labeled Z, enables the counter so that the counter increases by 1 every second. 