In numpy.sum() there is parameter called "keepdims". What does it do?

An example showing keepdims in action when working with higher dimensional arrays. Let's see how the shape of the array changes as we do different reductions:

import numpy as np
a = np.random.rand(2,3,4)
a.shape
# => (2, 3, 4)
# Note: axis=0 refers to the first dimension of size 2
#       axis=1 refers to the second dimension of size 3
#       axis=2 refers to the third dimension of size 4

a.sum(axis=0).shape
# => (3, 4)
# Simple sum over the first dimension, we "lose" that dimension 
# because we did an aggregation (sum) over it

a.sum(axis=0, keepdims=True).shape
# => (1, 3, 4)
# Same sum over the first dimension, but instead of "loosing" that 
# dimension, it becomes 1.

a.sum(axis=(0,2)).shape
# => (3,)
# Here we "lose" two dimensions

a.sum(axis=(0,2), keepdims=True).shape
# => (1, 3, 1)
# Here the two dimensions become 1 respectively

@Ney @hpaulj is correct, you need to experiment, but I suspect you don't realize that summation for some arrays can occur along axes. Observe the following which reading the documentation

>>> a
array([[0, 0, 0],
       [0, 1, 0],
       [0, 2, 0],
       [1, 0, 0],
       [1, 1, 0]])
>>> np.sum(a, keepdims=True)
array([[6]])
>>> np.sum(a, keepdims=False)
6
>>> np.sum(a, axis=1, keepdims=True)
array([[0],
       [1],
       [2],
       [1],
       [2]])
>>> np.sum(a, axis=1, keepdims=False)
array([0, 1, 2, 1, 2])
>>> np.sum(a, axis=0, keepdims=True)
array([[2, 4, 0]])
>>> np.sum(a, axis=0, keepdims=False)
array([2, 4, 0])

You will notice that if you don't specify an axis (1st two examples), the numerical result is the same, but the keepdims = True returned a 2D array with the number 6, whereas, the second incarnation returned a scalar. Similarly, when summing along axis 1 (across rows), a 2D array is returned again when keepdims = True. The last example, along axis 0 (down columns), shows a similar characteristic... dimensions are kept when keepdims = True.
Studying axes and their properties is critical to a full understanding of the power of NumPy when dealing with multidimensional data.