Deep Learning

Question 1

This is a traditional linear regression problem.

Your goal here is to find the hidden relationship between input (i_1,2,3) and output (o).

For example o = i₁ x a₁ + i₂ x a₂ + i₃ x a₃ find out what a_1,2,3 is?

Endpoint

Provide a POST endpoint /machine-learning/question-1 that given existing input, guess the expected output.

Input

The HTTP POST request will come with a body of Content-Type: application/json containing:

{
  "input": [
    [i11, i12, i13],
    [i21, i22, i23],
    ...
  ],
  "output": [
    o1, 
    o2,
    ...
   ],
  "question": [q1, q2, q3]
}

• input: 2d array (5 x 3), the sample input
• output: 1d array, the sample output, where:
  • o₁ = i₁₁ x a₁ + i₁₂ x a₂ + i₁₃ x a₃
  • o₂ = i₂₁ x a₁ + i₂₂ x a₂ + i₂₃ x a₃
• data: the input for you to find the corresponding answer

Output

Your endpoint should then return with body of Content-Type: application/json contains:

{
  "answer": a
}

• answer: your answer to the given question, where:
  • a = q₁ x a₁ + q₂ x a₂ + q₃ x a₃

Limits

• Your answer need to be correct within threshold: 99%
• Time limit: 20s
• Carry 30% of marks

Sample

For example, in this case the a₁ = 1, a₂ = 1, a₃ = 1:

Input:

{
  "input": [
    [1, 2, 3],
    [2, 3, 4],
    [2, 1, 4],
    [5, 3, 2],
    [2, 1, 2]
  ],
  "output": [
    6, 
    9,
    7,
    10,
    5
   ],
  "question": [3, 4, 5]
}

Output:

{
  "answer": 12
}

Question 2

Moving on, we goes into the area of a classic deep learning question.

Given 30 hand-written image (represented by pixel value), find what they represent.

The hand-written images will be send to you as JSON, each image is of dimension of 28 x 28. the payload will have the pixel value in gray scale (0 to 255), and flatten into array of size 784

Example (a hand-written 2): data (as png)

You can use whatever library or famework you can found online. Sample training data can be found from: http://yann.lecun.com/exdb/mnist

Endpoint

Provide a POST endpoint /machine-learning/question-2 that given a hand-written number image represented by pixel values, return the value of the number.

Input

The HTTP POST request will come with a body of Content-Type: application/json containing:

{
  "question": [
    [p1, p2, p3, p4 ..., p784],
    [q1, q2, q3, q4 ..., q784],
    ...
  ] 
}

• question: 2d array (30 x 784), where:
  • p_1..784 is pixel value of an 28 x 28 images
  • 0 <= p_1..784 <= 255

That contains 30 set of numeric representation of the image

Output

Your endpoint should then return with body of Content-Type: application/json contains:

{
  "answer": [
    a1,
    a2,
    ...
  ]
}

• answer: 1d array (30) your answer to the given input, where:
  • a₁ is number represented by p_1...784
  • a₂ is number represented by q_1...784

Limits

• Your answer need to be correct within threshold: 90%, i.e. number of correct answer >=27
• Time limit: 10s
• Carry 70% of marks

Sample

Input:

{
  "question": [
    [0, 0, 0 , 123, 34, 0 ... ,0],
    [0, 0, 0 , 0, 134, 0 ... ,0],
    ...
  ] 
}

Output:

{
  "answer": [
    2,
    3,
    ...
  ]
}