Important Interview Questions On Convolution Neural Network- Deep Learning

[@krishnaik06]

Interview What is the advantage of using CNN over ANN for image classification even though we can solve image classification using ANN?

[@tejas5872]

Availability of various filters to detect many features of the image and also we can have good clarity of the image as o/p when we do max pooling but when we use ANN we create the image to multi-dimensional array and the features of the image are not recognised. This is the reason CNN is useful and also using CNN we extract the exact image leaving noise @krish

[@harsh7704]

@krish @tejas 1) Main reason of not using ANN is: Large computation power while taking images as inputs and training ANN model 2) Images are made up of pixels. At particular regions in the image we have similar pixels. As for classification and detections, these ROIs are only needed which can later help to classify images or detect objects. So we need to do some kind of image processings to extract these important information from these ROIs. In cnn we use filters and pooling layers for that and ann doesnt't have this smarter solution to classify images

[@harsh7704]

@krish I am adding some extra questions: 1) When do we use Auto encoders? Can you tell about variational Autoencoders? 2) Which one you preferr: bigger cnn filter size or smaller? 3) Why have you specially chosen VGG 16 if you have other transfer learning models? 4) What are the o/ps from initial cnn layers? 5) You have cameras in drone, Can your yolo works (detect pedestrians) from cameras over drones ? 6) What's the difference between image classification, object localization and object detection? 7) Define CNN pipeline for face recognition? 8) Which one to prefer GANs or Autoencoders? 9) Is there some rule, at which place to use specific activation functions in any neural nets model? 10) Which is the new transfer learning model in demand today. Any recent one which provide even more accuracy/won Imagenet?

[@sreelatha2010]

When we increase the size of image the number of trian able parameters increases drastically in ANN while in CNN capatures the spatial features from an image which ANN fail to do. ANN do not scale well with input size

[@rohandawar484]

Vanilla neural networks (MLPs) can learn extremely complex functions, but their architecture does not exploit what we know about how the brain reads and processes images. For this reason, although MLPs are successful in solving many complex problems, they haven't been able to achieve any major breakthroughs engineered to work

[@mallutalksaboutaiwithab381]

Hi Krish, I agree with your statement regarding Keras Tuner. But as you know Keras tuner is just an hyper parameter optimization framework only and the intention behind the interviewer is to know the candidates understanding about the relationship between CNN selection with task, dataset size, Number of classes etc. Since its a mutually connected areas we have to give an answer which connects all this areas. Keras tuner is kind of black box for us. when we ans any question in an interview we have to give an answer which can satisfy the interviewer. Here interviewer will ask connected qn like, "What if am using Pytorch, and I don't want to use Keras tuner anyhow.". So friends, Insted of keras tuner I would suggest to explain in depth by connecting the relationships of Data, classes, and task while ans qn related to selection of network and hyper parameter selection.

[@dibyaranjanjena5650]

In CNN we apply different filters also known as kernels, so while applying kernels we took consideration into the neighboring pixels which helps the network to learn different features, eventually these are converted into a 1-D array in the feed forward network and we get the classification results.

[@tejas5872]

Availability of various filters to detect many features of the image and also we can have good clarity of the image as o/p when we do max pooling but when we use ANN we create the image to multi-dimensional array and the features of the image are not recognised. This is the reason CNN is useful

[@arijit07]

One of the advantages of using CNN over ANN is that it reduces no of parameters i.e weights and bias .

[@amarnathreddysurapureddy9111]

Cnn : Specifically designed to capture the spatial structure of images. Less parameters. Ann : each pixel independent. Cnn : lines, edges detecting(spatial)

[@saikiranreddymekala1346]

ANN are location specific where as CNN are location unspecific. Eg: If a CNN learn features of Ears it can locate they anywhere in the image. If we do data-agumentation techniques then there is a decrease in accuracy of ANN models where as CNN models still predict with higher accuracy.

[@abdulqadar9580]

Just Amazing way of teching Sir. Love from Pakistan.

[@jagadeesh_7924]

We use cnn for image classification insted ann because cnn works wonderful for images Cnn works as In first layer it try to detect edges and in second layer it try to detect shapes like that... And also it uses max pooling, min pooling etc.. So it provides the location invariant and scale invariant like features.

[@keshavbansal5148]

This explanation style works very well!!!

[@noorulaintariq2379]

Thankyou so much for putting so much hard work….I am sure this will help me in my thesis Defense…Thankyou again :)

[@kushshri05]

Categorical cross entropy is used when our true labels are one hot encoded or they are not ordinal classes While Sparse categorical cross entropy is used where our true labels are ordinal in nature

[@Ranjithbhat444]

Beautiful. Too much learnt here. Please use the same technique but one request is please consider answering the questions completely. Would be helpful for research enthusiasts. Lots of respect 🫡

[@sunilyadav3098]

In CNN, images will be having higher input dimensions, so to represent such data tensorflow framework can be used to represent huge data. But in ANN it will be very difficult to handle and process such huge dimensions. Therefore, CNN is used instead.

[@ArunKumar-sg6jf]

categorical cross entropy is used for multi class with label encoding and sparse cross entropy for muti class with one hot encoding

[@ashishn.c.7913]

medium.com/@sanjivgautamofficial/categorical-cross-entropy-vs-sparse-categorical-cross-entropy-b6a24de2b7f0 In this it is other way around. Which one is correct?

[@raneshmitra8156]

This white board format is quite good for theoretical understanding but for Mathematical explanation your older technique is much better....

[@nivedithagd3371]

We can use ANN for images but if we have large number of images and large number of pixels it uses more then million wights when we are doing back proportion computational power require more to overcome this problem we are using CNN

[@aravindraj2205]

Thanks krish for this video :)

[@louerleseigneur4532]

Thanks Krish

[@nilakantas5153]

Thank you sir

[@anshulzade6355]

The white board thing is great man keep doing this!, Any recommendations for how to study in depth and stay updated with the latest shit going on in DS and that too understanding it throughly!!. Keep it up

[@VIVEK26]

great video sir

[@xxx-sm5gj]

CNNs are inherently translation invariant, meaning they can recognize objects regardless of their position or orientation within the image. This property is achieved through the use of convolutional and pooling layers, which aggregate information across spatial dimensions and make the network more robust to variations in object position and orientation.

[@harikk7872]

Continue with the same

[@SuperMsmystery]

ANN will not recognise features specific to images like angles,light/darkness,inversion etc.

[@pullmed]

there used to be a reason why people use the words "average pooling" instead of "mean pooling". But looks like no one cares anymore.

[@asiftandel8750]

Yes Please use white board

[@Mohith7548]

Hey! Do anyone have answers to the questions?

[@vivekd9563]

What’s your telegram group sir ?

[@abhinandanshirahatti8859]

what is keras tuner

[@milindsai5254]

Keras tuner is used to choose the optimal number of hidden layers and neurons required for ANN and for CNN, it helps in deciding the optimal number of filters or feature maps, convolution layers, etc required. Check out Krish sirs video on the same topic.

[@ShahidAli-ed2qw]

Hey bro provide me telegram link

[@chetanmundhe8619]

When we use max pooling, mean and min pooling, can anyone answer plz,

[@milindsai5254]

Max pooling gives better results for the images with a black background and white object (Ex: MNIST dataset). Min pooling gives a better result for images with white background and black object . The average or mean pooling method smooths out the image and hence the sharp features may not be identified when this pooling method is used. Source: medium.com/@bdhuma/which-pooling-method-is-better-maxpooling-vs-minpooling-vs-average-pooling-95fb03f45a9

[@rrrfamilyrashriderockers6891]

CNN can learn by edges