##Deep Learning Typical Interview Quetions
Following are few theoretical questions people ask. You might answer all
Importance of random weight initialisation?
- To add non-linearity to the network, otherwise it becomes a simple high-dimensional linear model similar to like SVN.
Effect of Training a shallow-network vs Deep-network?
Why convolution, why not direct MLP?
- Less params than MLP so that they becomes practically implementable
- Spatial understanding and understanding spatial features
- Makes the system translational invariant
Accuracy measures and their uses?
- Precision
- Recall
- Accuracy
- F1-Score
- RoC Curve
- PR Curve
- mAP
Why MaxPool? Why not any other function?
What is ResNet? Advantages of ResNet? What problems were solved by ResNet?
- It solves Vanishing Gradient problem
What is Vanishing Gradient Problem in DNN?
Why use Relu? Why we canot use sigmoid or tanh activation function?
What is softmax? When to use softmax?
What is bias-vaiance problem?
Why 1x1 convolutions? How does it reduce computations?
- Depth-wise directionality reduction
- Makes input independent of size when compared to FC layer where input image has to be of fixed size due to FC layer is hard-coded stuff.
- Show calculation of how much 1x1 conv helps in reducing the computation. Refer to 1x1 conv.md in the same repo.