Q: Can we say feedforward is better than other deep learning algorithms or vice versa by itself without knowing their application?[YM]
A: Feed-forward networks can be trained most easily and provide effcient solutions for inference. Therefore, in the following sessions we will consider only architectures which can be transformed (unrolled) into deep feed-forward networks.

Q: As we increase neural network hidden layers, the demand for computitional resources also increases. what is recommended?[YM]
A: Try to find a good compromise by carrying systematic  experiments with different network layouts. In general,  the size of the hidden layer should be as small as possible to force the model to generalize, i.e. to avoid overfitting.

Q: How activation functions work?[YM]
A: It  maps the aggregated input activity (i.e. the weighted sum) to the desired output range. Because  it plays a major role in determining the gradient of the loss function it should be easily derivable.

Q: What are the challenges of neural network model? [DM]
A: They have to be trained efficiently and effectively on as little as possible amounts of data that can be collected with as little as possible human (annotation) effort.
    
Q: which Deep learning model is more suitable for motion and Gesture detection and realtime conversation for vision support translation for example in sign language translation in to text [AH]  
A: I do not know. This is such a narrowly defined reserach question, you will hardly find an expert who can give you a reliable answer. All I can recommend is: check the existing literature for similar studies.