Deep learning Flashcards
What is Machine learning
we provide examples to the machine and the model learns from those examples
what are the systems that are not based on learning
Usually, expert systems, where we have a set of rules defined by experts to allow classifying things where the system
Explain the 4 stages of machine learning
1-Expert systems feature designed by experts through imposing the rules
2-classic machine learning: we extract the features from the dataset then we perform the supervision
3-Representation machine Learning: we let the model extract its own features
4-Deep learning: less experts more examples
what is the difference between perception and multi layer network
one is made for linear problems while the other is made for non linear problem
what was the issue before in very big neural networkpo
The main problem is that we need to backpropagate the error hence
the issue is the it is impacted by the vanishing of the gradient
What is convolutional layers
it is simply a layer that looks at small subparts of the image depending on the size of the kernel
convolutional layers have been developed specifically for images, in other words, the pixels that belong to the same neighbourhood are related
however if we try to use all this data we ll end up with a huge vector , hence we may use subsampling strategy to diminishing the size of the network
unlike usual network , in convolutional layer we have that every neurons is a receptive fields such that every neuron actually sees only a portion of the image
the more we goo deeper into the network the wider it is the receptive field
what is a dense neural network
similar to a perceptron but every neuron layer is connected to another neuron layer
explain briefly Alex net
it was one of the first neural networks composed of several convolutional layers then a dense layer
How to fix the vanishing of the gradient
if the weight of the derivative is very low meaning when we have every small error it leads to vanishing gradient
To fi
why may want ot minimize the use of dense neural networks
we know that the more weights we have to optimize the mode data we need, hence, the dense layer is eponeential meaning we need way more data train the network with dense layer ; it is always a good idea to minimize the number of weights
what is the advantage of convolution layer compared to dense layer
the dense layer is exponential in terms of weights while it is not the case for the convolutional layer which keeps the number of weights reasonable
what is max pooling
one of the issues of the neural networks is the high amount of weight especially when we have large images
Max pooling diminishes the size of networks for every 4 neurons ze may keep the maximum value
we diminish the size of the network by 4
why we use several layers of convolutional and max pooling layers
the idea is that apply several layers of convolution and max pooling the more we go down the smaller is the our image in the end obtain a vector that describes our image and read for the learning
why using shortcuts
it help overocming the vanishing of the gradient
what is the difference bet
in regression we attemp to predict a single value while classification we predict a singe classs
