Scribd
Upload
73 views

Assignment Problem - Gradient Descent

Gradient Descent is one of the most important part of machine learning, this assignment is helpful for understanding it in step wise manner.

Uploaded by

UMESH SUGARA
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
73 views

Assignment Problem - Gradient Descent

Gradient Descent is one of the most important part of machine learning, this assignment is helpful for understanding it in step wise manner.

Uploaded by

UMESH SUGARA
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 6

Assignment Problem: Gradient Descent Algorithm

in Python

Consider the following 3 datapoints.

X1 Y
(Feature) (Target)
1 4.8

3 12.4

5 15.5

• The following plot shows these 3 datapoints in Blue circles. Also shown
is the red-line (with squares), which we are claiming is the “best-fit
line”.

• The claim is that this best fit-line will have the minimum error for
prediction (the predicted values are actually the red-squares, hence the
vertical difference is the error in prediction).

• This total difference (error) across all the datapoints is expressed as the
Mean Squared Error Function, which will be minimized using the
Gradient Descent Algorithm, discussed below.

• Minimizing or maximizing any quantity is mathematically referred as an


Optimization Problem, and hence the solution (the point where the
minima/maxima exists) is referred the “optimal values”.
Here X1 refers to the independent variable (also called as Feature / Attribute in
Machine Learning)
Y is the dependent variable (also known as Target Variable in ML)

The net Objective is to find the Equation of the Best-Fitting Straight Line
(through these 3 data points, mentioned in the above table, also represented by
the blue circles in the above plot).

Ŷ = w 0 + w1X1 --------------- (1)

is the equation of the best-fit line (red-line in the plot)


w1 = slope of the line; w0 = intercept of the line

w0 and w1 are also called model weights.

Y^ is the predicted values of Y, given by the “best-fit line”. These predicted


values are represented by red-squares on the red-line. Of course, the predicted
values are NOT exactly same as the actual values of Y (blue circles), the
vertical difference represents the error in the prediction given by:
ˆ −Y
Errori = Y -------------- (2)
i i

for any ith data point.

1 N 1 N ˆ
MSE = 
N i =1
( Errori ) 2
= 
N i =1
(Yi − Yi )2 --------------- (3)

N = Total no. of data points (in this case, it is 3)

Now the problem is to find the “optimal values” of the slope and intercept of
this best-fit line, such that the “Mean Squared Error” (MSE) is minimum. You
can easily see that the yellow-line (a poor-fit line) which has “non-optimal”
values of slope & intercept fits the data very badly (btw the exact equation of
the yellow line is x+6, so slope is 1 and intercept is 6 units)

How will i get the optimal values of the slope and intercept ????

This is where the Gradient Descent Algorithm comes!

 N ˆ 
w k +1
0 = w −    (Yi − Yi ) 
k
0
 i =1 
 N ˆ 
w k +1
1 = w −    [(Yi − Yi )  X 1i ] 
k
1
 i =1  ---------- (4 & 5)
where w0k & w1k represent the values of the intercept and the slope of the linear-fit
in the kth iteration, whereas w0k +1 & w1k +1 represent the values of the intercept and
the slope of the linear-fit in the (k+1)th iteration (next iteration). w0 & w1 are
also called as model weights or model coefficients.
α represents the Learning Rate.

The derivation of the above equations will be done in the machine learning
lessons.
Gradient Descent Algorithm
1. Initialize the algorithm with random values of α, and weights (w0, w1)
2. Calculate predictions Y^ = w0 + w1 * x as shown in the equation 1
3. Calculate Error terms & MSE Loss Function (L):
N N
Error terms are:  (Yˆi − Yi ) and
i =1
[(Yˆ − Y )  X
i =1
i i 1i ] for the datapoints i=1 to

N (here N is equal to 3)

Loss function as shown in equation 3

4. Update your weights: Using equations 4 and 5


5. Repeat 2-4, until convergence.

Based on the above-mentioned steps, we can calculate the weights. Let the
learning rate (α) be 0.01 and initialize the weights w0 and w1 as 0.
X1 Y Y^ Y^-Y (Y^-Y)*X1

1 4.8 0 -4.8 -4.8

3 12.4 0 -12.4 -37.2

5 15.5 0 -15.5 -77.5

Sum -32.7 -119.5

w01 = 0 – (0.01) * (-32.7) = 0.327

w11 = 0 – (0.01) * (-119.5) = 1.195

With these new weights, let us update the above table.

X1 Y Y^ Y^-Y (Y^-Y)*X1

1 4.8 1.522 -3.278 -3.278

3 12.4 3.912 -8.488 -25.464

5 15.5 6.302 -9.198 -45.99

Sum -20.964 -74.732

w02 = 0.327 – (0.01) * (-20.964) = 0.537

w12 = 1.195 – (0.01) * (-74.732) = 1.942

With these new weights, let us update the above table.


X1 Y Y^ Y^-Y (Y^-Y)*X1

1 4.8 2.479 -2.321 -2.321

3 12.4 6.363 -6.037 -18.111

5 15.5 10.247 -5.253 -26.265

Sum -13.611 -46.697

As we can see, the sum of errors is decreasing as we are updating the weights.
We can continue to update the weights like the above manner until the sum of
errors become minimum (i.e. reaches almost a constant value)

Finally I want the plots of :


• Loss function (y-axis) vs w0 (x-axis)
• Loss function (y-axis) vs w0 (y-axis)
• 3D-plot of Loss function w.r.t. w0 & w1 similar to below image

You might also like