1. Which mathematical operation 27.

In CNN, connections in the convolution layer is

characterizes Al Machine Learning sparse because one neuron in
compute? Matrix Multiplication the next layer is connected to only n ? neurons in
2. What is the most common thing in the current
CNN/Dnn – Matrix multiplication layer, where ? is x
3. What is the input to a neuron called? 28. In CNN, images and corresponding
dendrite are input while, in traditional computer vision,
4. What is the typical error rate of humans images and algorithm are input. Outputs
classifying ImageNet objects? 5% 29. Which is the most common operations in
5. Consider 8-bit QP (Quarter Precision) float8 CNN/DNN? Matrix multiplication
FP representation, where one bit is 30. What is FFT of [1 1 0 0]? 2:1-1; 0; 1+i
reserved for the sign, three bits for biased 31. What is the precision limitation, i.e., minimum
exponent, and the fixed interval, on half-precision
remaining four bits are used for the floating-point decimals between 2^-5 and 2^-4?
mantissa. Which is +1.125? 0-011-0010 (^ means power) 2^-15
6. Which is the neural network model 32. In order to correctly perform -129 - 145, how
appropriate for unstructured data? MLP many bits are at least required to
7. 5x5 2D convolution on real-time 1Kx1K represent the numbers? 10
color video requires roughly how many 33. The 5-bit binary fixed-point equivalent of -0.5861
OPS (operations per second)? is 10111
Real-time video means 30 FTS (Frames Per 34. Which is the correct Sobel convolution filter
Second). 2 GOPS (Giga) kernel? [-101; -2 0 2; -1 0 1]
8. In CNN, which part requires most 35. The maximum positive representable number for
computation? C o n v o l u t i o n HP floating-point number is (2-2^-10 x2^15
9. In DNN/CNN, recognizing multiple objects 36. The smallest-magnitude number for HP floating-
on the picture, along with a bounding point number is 0 00000 0000000001
box for each object, is called Object 37. What is the fixed interval, or resolution, of
Detection decimals between 64 and 128? 2^-4
10. Very few people /companies have the 38. Theoretically, low-pass filter is based on (), while
luxury of accessing very high-speed GPs high-pass filter is based on (addition, subtraction
and powerful hardware to train a very deep 39. The closer the filter specification is to the ideal
network from scratch in a reasonable filter, the
time. So one often uses a pre-trained the filter length is likely to be. Longer
network for similar application and "fine- 40. Which of the following is the correct IEEE long
tune" floating point representation of
the weights for the target application. What 57.125? 0 1028 11001001000...0
is this type of learning called? Tr a n s f e r 41. Which does not belong to Tesla FSD Chip? H.265
learning Video Decoder
11. What is pooling in CNN equivalent to? 42. Which category does TPU belong to? ASIC
Downsampling 43. Which is the correct decimal equivalent of the
12. In DNN, training determines weights, but IEEE short floating-point number:
applies weights to determine output? 0 10000100 1100 1101 0000....... 0000 55.625
Inference 44. For floating-point adder, which is the correct
13. What is the function (equation) that is used pipeline? align-add-normalize
for Sigmoidal Function? F(x) = 1/(1+e^-NET) 45. Which is the correct exponent field of 110.875 in
14. When is the starting year of Al and Machine IEEE single precision floating
Learning? 2015 point? 10000101

distribution is nonlincar. Gaussian is nonlinear

 33. Which is the correct mantissa field of 11.75 1. Which statement is not correct about gamma
in lEEE single precision floating correction? Gamma greater than 1 changes the
point? 011110.....0 output pixel value greater than the input value.
34. What is the 5-bit product result of signed 2. Which is the basis function of 2x2 DCT? All o f
multiplication 1101×1011 with rounding. 00100 these
35. If a neural network accelerator (NNA) has 64K 3. Which is the correct definition of PSNR? -
(256x256) 8-bit MACs and 10*log10(MSE/255^2)
operates on 1GHz clock, what is its rough 4. Which is not the correct filter mask of the
performance in TOPS? 60 Laplacian filter? [-1-1-1: 0 4 0: 1 1 1]
Quiz: 5. In Eigenface-based face recognition, the average
15. Which is the correct mantissa field of 11.75 face corresponds to. DC
in IEEE single precision floating point? 6. What is not one of the major factors DNN and
011110...0 CNN are popular now? Autonomous vehicle
16. In order to correctly perform - 29 - 145, at technology
least how many bits are required to 7. What is the input to a neuron call?
represent the numbers? 9 8. Which DNN has first beaten humans in correctly
17. What is the 4-bit minimum negative recognizing visual objects? ResNet
number in fraction? -0.875
18. What is the correct -bit fixed-point Matlab
code after executing p=h*х? p =
floor(p*2^(N-1))/2^(N-1)
19. You are supposed to design a HD audio.
processing system using a FPGA chip.
A real-time stereo audio is sampled at
40KHz, 16 bits per sample. A 21-tap
bandpass audio filter is used to process the
audio input. How many MAC operations
per second are required for this system?
80.000x21
20. Which 7-series Xilinx FPGA family has the
best SP performance? Virtex
21. How is the processing power of a neural
network processor measured?
Interconnected updates per second
22. What is not one of the major factors DNN
and CNN are popular now? Autonomous
vehicle technology
23. Which is not the major metrics or features
of CNN? None of the three
24. Which DNN has first beaten humans in
correctly recognizing visual objects? ResNet
25. Which statement is not correct about rods
and cones? Human eyes have 4 times more
rods than cones
26. Which statement is not correct about
Gaussian filter? As the Gaussian
• The Xilinx LUT is equivalent to ? MUX • What is the correct decimal equivalent of IEEE
short FP number? 57.625
• In signed mult of 1111 and 1111 what should be
the bits in a through e?
1 1 1 1
x 1 1 1 1
1 1 1 1
• e 1 1 1 1
11 1 1 1 1
• What is the main difference between Xilinx zynq 7000 and a b c D
zynq ultrascale+ SOC FPGA? All of these ? ? ? ? ? ? ? ?
• In CNN which part requires most computations? (a)-9e) = 00011
Convolution. • Suppose 2 fp numbers a1=143 and a2=-9 are
• What is not a feature of FPGA? NOT added on the pipelined half precision FP adder
• Which FPGA board is python based ? PYNQ-Z1
• What is FFT of |1100|? 2:1-i:0:1+i
• The maximum positive representable number of HP
floating point number is ? (2-2^-10)x2^15

Return 0;
-----------------------------------------------------------------

• Below is the code shown listing the FIR filter in mini-proj 2 Suppose x1=0.65 and x2=0.6 for a simple neural
revise the code to implement LMS adaptive FIR filter network net connection shown below. Show op y
Assume that you given d(n) any other assumptions you in binary form from for “int5” low bit precision(5-
may make should be written bit signed fixed point ) repersentation

#include “fir.h”
Void fir(data_t*y.coef_tmu,data_tx,data_td)/
//d(n) is passed on the function fir
{ static data_t_shift_reg[N];
Acc_t_acc;
Data_t_data;
Int I,j
 //declaring extra variables required
Double error;
Double w[N] ={0.0};
Acc=0;
// code changed from here
LMS Accum Loop: for (j=0;j<N;j++) {
For(i=j;i>j-N;i++) {
If(i>0)
Shift reg[N+(i-j)-1] – x[i] // LMS FIR filter
Else break;
Y=0; //filter o/p set to 0
For(i=0;i<N;i++)
Acc++(w[i]*shift_reg[i]); //calculate filter output
Error = d-acc;// calculate error

For(i=0;i<N;i++)
W[i] = w[i] +(mu*E*shift_reg[i]);//adaptive error
adjustment
Y=acc;
------------------------------------
• In DNN training determines weight but __ applies weight
to determine output? Independence or impedence
• Which DNN has feed back connections? RNN
• Which is correct ReLU activation function / y=max(0,x)
• LPF is analogus to Wine
• Which is not correct convolution of filter mask of the
laplician convolution
1 1110 4 0111
• What is the fixed interval or resolution of decimal between
64 and 128 is 2*-4 or 4
• Which is not correct statement on the figure below
It is called sobel operation-- answer
-1 0 1 111
-1 0 1 000
-1 0 1 sx -1 -1 -1 sy

• In which type of learning output is assessed via reward and

punishment? Reinforcement
• In CNN before showing images to the computer for training
objects in the image have to be isedtified to be manu –
Annotation
• Neural network is deep if it has ____ of non linear features
transformation – Relatively large number of stages
1. The mathematical operation which characterizes AI/Machine Learning compute is Matrix Multiplication.
2. The input neuron is called dendrite.
3. The typical error rate of humans classifying ImageNet objects is 5%.
4. Consider 8-bit QP (Quarter Precision) float8 FP representation, where one bit is reserved for the sign, three bits for biased exponent, and the remaining four
bits are used for the mantissa. +1.125 represents 0-011-0010.
5. The neural network model appropriate for unstructured data is MLP.
6. 5x5 2D Convolution on real-time 1Kx1K color video requires roughly 2 GOPS(Giga).
7. In CNN, Convolution part requires most computation.
8. In DNN/CNN, recognizing multiple objects on the picture, along with a bounding box for each object is called Object Detection.
9. Very few people/companies have the luxury of accessing very high-speed GPUs and powerful hardware to train a very deep network from scratch in a
reasonable time. So, one often uses a pre-trained network for similar application and ‘fine-tune’ the weights for the target application. Such type of learning is
called Transfer Learning.
10. Pooling in CNN is equivalent to Downsampling.
11. In DNN, training determines weights, but Inference applies weights to determine the output.
12. The function(equation) that is used for Sigmoidal Function is F(x) = 1/(1+e^-NET).
13. The starting year of AI and Machine Learning is 2015.
14. In CNN, connections in the convolutional layer is sparse because one neuron in the next layer is connected to only nXn neurons in the current layer.
15. In CNN, images and corresponding weights are input while in traditional computer vision, images and algorithm are input.
16. The most common operations in CNN/DNN are Matrix Multiplication.
17. The FFT of [1 1 0 0] is 2; 1-i; 0; 1+i.
18. (3) The precision limitation, i.e., minimum fixed interval, on half-precision floating-point decimals between 2^-5 and 2^-4 (^means power) is 2^-15.
19. In order to correctly perform -12910 – 14510, the least number of bits required to represent the numbers is 10.
20. The 5-bit binary fixed-point equivalent of -0.5861 is 10111.
21. The correct Sobel convolution filter kernel is [-1 0 1; -2 0 2; -1 0 1].
22. The maximum positive representable number for HP floating-point number is (2-2^-10) x 2^15 = 65504.
23. The smallest magnitude number for HP floating-point number is 0 00000 0000000001.
24. The fixed interval, or resolution, of decimals between 64 and 128 is 2^-4.
25. Theoretically, low-pass filter is based on addition, while high-pass filter is based on subraction.
26. The closer the filter specification is to the ideal filter, the Longer the filter length is likely to be.
27. The correct IEEE long floating-point representation of 57.125 is 0 1028 11001001000……0.
28. H.265 Video Decoder does not belong to Tesla FSD Chip.
29. TPU belongs to ASIC category.
30. The correct decimal equivalent of the IEEE short floating-point number: 0 10000100 1100 1101 0000…….0000 is 57.625.
31. For floating-point adder, the correct pipeline is align-add-normalize.
32. The correct exponent field of 110.875 in IEEE single precision floating-point is 10000101.
33. The correct mantissa field of 11.75 in IEEE single precision floating-point is 011110…...0.
34. The 5-bit product result of signed multiplication 1101x1011 with rounding is 00100.
35. If a neural network accelerator (NNA) has 64K (256x256) 8-bit MACs and operates on 1GHz clock, its rough performance in TOPS is 60.
36. 1111 and 1111 in 2’s complement number are multiplied as 0001
37. The correct N-bit fixed-point MALTAB Code after executing p=h*x is p = floor(p*2^(N-1))/2^(N-1).
38. The 4-bit minimum negative number in fraction is -0.875.
39. The quantization error when 0.74 is represented as signed int5 is 0.01.
40. -0.74 x 0.6 in signed int5 representation with rounding is 11100.
41. 0.64 x 0.4 in signed int5 representation with rounding is 00011.
42. -0.9 x 0.8 – 0.7 x 0.6 in signed int5 representation with rounding and saturation is 10001.
43. The correct mantissa field of -11.75 in IEEE single precision floating-point is 011110……0.
44. Incorrect Statement: Mythic mixed -signal computing is done in the memory. Mythic's main strength is order-of-magnitude power reduction. Mythic chip's
CNN processing performance is far higher than that of GPU. Embedded flash transistors may be modeled as variable resistors.
45. 192 different ordinary number are represented in 8-bit QP float8.
46. <diag> EMS Normalizing and Rounding: 2
47. Output value from shifter in Normalizing and Rounding is 00010101110…
48. Output value from Mantissa Adder in Normalizing and Rounding is 00111101010…
49. Incorrect statement: High-pass filter is basically based on arithmetic subtraction. High-pass filter is used to extract edges from the image. High-pass filter is
like wine in its effect. High-pass filter may be used to emphasize the difference.
50. Incorrect statement: 32-bit DRAM read requires 200 times more energy than 32-bit fixed-point multiplication. Conversion of SP FP to HP FP arithmetic
results in typically 400% energy saving. 32-bit fixed-point multiplication requires 100 times more energy than 8-bit fixed-point addition.16-bit HP FP add
requires 10 times more energy than16-bit fixed-point addition.
51. The deep neural network is most expensive in terms of memory and computation is VGG-16.
52. Suppose a 2x2 convolution filter is to be applied to 3×3 input feature map. The element (pixel) of the input feature map which is reused most in the
corresponding Toeplitz matrix is a Center element.
53. On the "bit width scaling" diagram below, if neons and synapses are signed number, the output at the end of the second cycle will be 001.
54. In Tesla FSD NNA, convolution is 99.7%, and ReLU/Pooling accounts for 0.3%. The next generation NNA-2 has a super-fast ReLU/Pooling unit running
10,000 times faster. The overall speedup is 1.003.
55. The kind of activation which enables pruning based on magnitude of weights and is widely used in industry is ReLU.
56. The performance metric of DNN hardware architecture: Cost. Throughput. Power. All of these.
57. Not part of the DNN Heterogeneous System: GPU. FPGA. CPU. None of the three.
58. One 32-bit DRAM read requires roughly 700 times the energy of one IEEE single-precision floating-point addition.
59. Not present in Xilinx Al Engine (AIE): SIMD Vector Unit. None of the three. Load/Store Units. Scalar Registers.
60. Increase in computational requirements is not the recent trends in CNN.
61. <diag> Not a Correct statement on Sx= [-1 0 1; -1 0 1; -1 0 1] and Sy= [1 0 1; 1 0 1; 1 0 1] is It’s called Sobel operator.
62. The main difference between the Xilinx Zynq 7000 and Zynq Ultrascale+ SoC FPGA: ARM Cortex A9 VS A53. 28 nm vs 16 m. No DPU vs DPU (Deep
Learning Processing Unit). All of these.
63. Run length coding is the technique called that compresses sparse matrix.
64. Briefly summarize your results, preferably quantitatively, from your Mini project#4 as described in the conclusion section. (Our miniproject)
OP:
a. From the miniproject 4 we can summarize the result and conclusion as below
b. The sigmoid gradient evaluated at [ -1 -0.5 0 -0.5 1]: 0.196612 0.235004 0.250000 0.235004 0.196612
c. Followed by the backpropagation implementation, the relative difference is 2.31694e-11 to 2.18975e-11
d. The cost at debugging parameters at lambda = 3.00000 is 0.576051
e. The Training accuracy is 94.88%
f. Theta1 and Theta2 values obtained from the Matlab run are used as inputs for the neural network circuit in Verilog. The entire neural network
circuit is broken down into individual functional components, i.e., FPMAC instantiations for input, hidden and output layers.
g. The total number of neurons are decided with the combination of input and hidden layer.
h. The input layer consists of 5000*400 neurons, the hidden layer consists of 25*401 neurons and the output layer consists of 10*26 neurons.
i. Thereby creating an output matrix of 401*10 neurons.

65. Show the lookup table by filling out the blank spaces In the Verilog code for computing y arithmetic (DA). Coefficient are signed and a1 - 0100, a2 • 1010,
and a3» 0011.
module case8 (table_in, table_out);

input [2:0] table_in; // Three bit

output [3:0] table_out; //

reg [3:0] table_out.

// This is the DA CASE table for

// the 3 coefficients
always @(table_in)
begin
case (table_in)
0: table_ out = ( );
end
endmodule
OP:
table_out = 0 x (4) + 0 x (-6) + 0 x (3) = 0000; //0
table_out = 0 x (4) + 0 x (-6) + 1 x (3) = 0011; //a3
table_out = 0 x (4) + 1 x (-6) + 0 x (3) = 1010; //a2
table_out = 0 x (4) + 1 x (-6) + 1 x (3) = 1101; //a2+a3
table_out = 1 x (4) + 0 x (-6) + 0 x (3) = 0100; //a1
table_out = 1 x (4) + 0 x (-6) + 1 x (3) = 0111; //al+a3
table_out = 1 x (4) + 1 x (-6) + 0 x (3) = 1110;//a1+a2
table_out = 1 x (4) + 1 x (-6) + 1 x (3) = 0001; //alta2+a3

65b. Incorrect statement? Low pass filter maybe used to highlight a portion of an image.
65c. Incorrect statement? SP FP multiplier is 3 times as big as HP FP multiplier.
66. What kind of logical operation can be implemented using a perceptron with decision boundary x1 + x2 = 1.5? Assume x1 and x2 are inputs and take 0 or 1 as
values. Consider the modified perceptron function AND
67. Which of the following methods can be used to improve the performance if 10 different neural networks are trained on the same dataset (w/o training any
other model)? Ensemble of 10 models.
68. Which of the following is FALSE about AlexNet? The input size for AlexNet is 227x227x1.
69. Make comparison for the overall speed of the following algorithms. (Assume that you have no knowledge of the hardware utilized for the algorithms). Each
item should be selected once. Stochastic> Full-batch> Normal equations
Normal equations –
Full-batch Gradient Descent -
Stochastic Gradient Descent –
70. Which of the following statement is FALSE about training RNNs? The cost function for backpropagation algorithm for MLP and BPTT are exactly the
same.
71. The convergence behaviors of the full-bath gradient descent algorithm for a linear regression problem with one input variables is displayed as <curve>.
X-axes denote the weight w; y-axis denotes the cost function. Initial value for the weight is selected as w = 10. And then the gradient descent algorithm steps are
shown by the full circles while the solution converges to w =0. Which of the following statements can be true about the learning rate? Multiple answers are
possible. Learning rate decreases over iteration steps & The learning rate first increase then decreases.
72. Consider the binary classification problem using the datagram using the dataset given in the figure. The squares are positive sample (Y=1) and the circles are
negative sample (Y= -1). There are n= 20 data points. The dataset is balanced. <chart>. Assume a classifier uses the following rule to make a decision. If X2>3,
yhat= +1; otherwise yhat= -1. Which of the following statements is FALSE for this dataset and the classifier? The F1score of the classifier is about 0.345
73. <Diag>. Fig.3. Architecture of our 8-layer convent model. A 224 by 224 crop of an image (with 3 color planes) is presented as the input. This is convoluted
with 96 different 1st layer filters (red), each of the size 7 by 7, using the stride of 2 in both x and y. The resulting feature maps are then: i) passed through a
rectified linear function (not shown), ii) pooled (max within 3x3 regions, using stride 2) and iii) contrast normalized across feature maps to give 96 different 55
by 55 element features maps. Similar operations are repeated in layers 2, 3, 4, 5. The last two layers are fully connected, taking features from the top
convolutional layer as input in vector form (6.6.256 = 9216 dimensions). The final layer is a C-way SoftMax function, C being the number of classes. All filters
and features maps are square in shape.
A CNN architecture is given as above. What is the number of parameters needed in the first convolution layer? (7x7x3+1)x96
74. Put the following in the correct order in the back propagation algorithm: forward pass, backward pass, initialize weights and biases randomly, weight update,
initialize weights and biases to zero. Start with step 0.
Forward pass – Step 1
Backward pass – Step 2
 initialize weights and biases to zero – Not used
initialize weights and biases randomly- Step 0
update weights – Step 3
repeat steps 1 through 3- Step 4
repeat steps 0 through 3 – Not used
repeat steps 2 through 3 – Not used
diagonal pass – Not used
75. For a multi class classification problem, during the back propagation algorithm, the local gradient at a neuron j is calculated using Ցj(n) = phi’j (vj(n))ej(n).
Which of the following statement is FALSE? The neuron j belongs to output layer
76. Which of the following functions can be used as an activation function in the output layer if we wish to predict the probabilities of n classes (n > 2)?
SoftMax and RELU
77. A polynomial regression problem is solved using Normal Equations:
Wopt = (XTX)-1 XTy (Normal Equations)
Where X = 1 12 and y = 2
1 22 6
1 32 8
1 42 7
1 52 11

Wopt = 3.5353
0.2968

The value of prediction for x = 2 is 6.

78. The highway which divide Intel and Nvidia headquarters is: 101
79. Triangulation was used to shape rooftop of the new Nvidia buildings.
80. What capability in particular has made the Nvidia GPU the king of AI processors? SIMD
81. AMD has acquired Xilinx.
82. 32-bit floating Multiplication has maximum hardware cost of operation-based area cost.
83. Average Absolute Weight is l1-norm.
84. In non-linear quantization, if we have 128 levels, the number of bits required to represent the precision is 7.
85. What is the performance matric of DNN hardware architecture? All of these.
86. How many TOPS Performance is achieved in Nvidia PASCAL chip set: 47
87. Types of pooling is supported in Tesla FSD chip.: Average and Max pooling
88. FSD SIMD data path has ____ stages of pipeline 3.
89. In FSD, extensions are used for____ Both
90. Multiply 1010 and 0101 using 2’s complement multiplication method. 1000010.
91. 5-bit binary fixed-point representation of -0.3878 1.1010
92. In order to correctly perform -125 10 – 44 10, number of bits at least required to represent these numbers: 9 bits
93. Which is the correct decimal equivalent of the IEEE HP floating point number: 1 10101 1100110101 -115.3
94. Which of the following is the correct IEEE HP floating point representation of 45.678? 0 10100 0110110110
95. The maximum positive representable number for HP floating-point number is: (2-2^-10) *2^15, 65504
96. The correct Prewitt convolution filter kernel: Sx = [-1 -1 -1; 0 0 0; 1 1 1]; Sy = [1 0 -1; 1 0 -1; 1 0 -1]
97. The correct convolution mask of the Laplacian convolution filter. All possible answers: [0 1 0; 1 -4 1; 0 1 0], [1 1 1; 1 -1 1; 1 1 1], [0 -1 0; -1 0 -1; 0 -1 0],
[-1 -1 -1; -1 8 -1; -1 -1 -1].
98. Xilinx LUT is equivalent to Mux.
99. The blank spaces a and b in the 2’s compliment array multiplier for 1111 X 1111. 11
100. The main difference between Xilinx ZYNQ7000 and ZYNQ Ultra scale + Soc FPGA? All of these
101. Not the feature of FPGA: None of these
102. Python based FPGA board is Pynq Z1
103. The signed number multiplication of 1111 X 1111. Then, bits in a through e is a-e = 00011