I was learning about inception module from deeplearning.ai by Andrew Ng, wherein we use 1x1 convolution to reduce computational cost. for example, if we directly apply 5x5 convolution we need to multiply for about 120M times whereas by introducing 1x1 convolution we reduce it by the factor of 10.

My question what effect does this have on the feature maps that are being learned or will the quality of feature maps differ if we had not used 1x1 convolution.

below is the screenshot of how we reduce multiplications from 120M to about 12M:

enter image description here