Neurons and gradient descent

• Neural nets are interconnected networks of simple processing units, a.k.a neurons
• It remains just a parallel, the artificial neuron is just an approximation! The output y was modelled as a weighted linear combination of inputs xj. Moreover, a transfer function, or activation function was used to make a decision.

  Gradient descent–how to find the right weights?

  

• We want to find the minimum of an error function
• We start with an initial guess of the parameter b
• We change its value in the direction of maximal slope (导数Derivatives are slopes)
• We continue until reaching a minimum

  Steps

  

• To update a weight b, we remove to its value the derivative
• a is called a learning rate»It decides by how much you multiply the step vector»Large a can lead to faster convergence than small» too large a can lead to disaster
• r is the iteration
• Why the minus sign? Because we want to move towards a minimum!

  Momentum

  

  Numerical example

  Learning rate controls oscillation and speed Momentum uses a bit of the previous step

  the models using them multi-layer perceptron convolutional neural networks ( deep learning)

  Multilayer perceptron (MLP)

  (from linear classifier to nonlinear version)

• It’s a feed forward network: it goes from inputs to outputs, without loops
• Every neuron includes an activation function (e.g. sigmoid, see earlier slides)

  Chain rule

• How to compute gradient descent on a cascade级联 of operations?
• How to reach all trainable weights?
  1. The forward pass: you put a data point in and obtain the prediction
  2. The Backward pass: you update parameters by back-propagating errors exercise Describe the architecture of the following network. ● How many layers are there? 2 ● How many neurons per layer? 3,1 ● How many trainable parameters (weights and biases)?10

     Convolutional Neural Networks(CNN)

• they proposed a network learning spatial filters on top of spatial filters. They were (and are) called convolutional neural networks
• The CNN was considered interesting, but very hard to train. It needed● Loads of training data > nobody had them● A lot of computational power > same

  Change

  big data graphic processing

  How to work

• They are conceptually very similar to MLPs
• But their weights (b) are 2D convolutional filters
• For this, they are very well suited for images
• In convolutional neural networks, the filters are spatial (on 2D grids). ● local : they convolve the values of the image in a local window● shared: the same filter is applied everywhere in the image: why shared? “Recycling” the same operation allows to have much less parameters to learn!

  Learn

• By “learns”, I mean adapt filters weights to minimize prediction error (i.e. backpropagation)

  Steps

• Convolutional filters start with random numbers
• Iteratively they are improved: each coefficient is updated in the direction of largest gradient of the cost function
• At the end, the filters become quite meaningful!

  Summary

• Perceptrons are “neuron-inspired” linear discriminants
• Multilayer perceptrons are trainable, nonlinear discriminants
• Feed-forward neural networks in general can be used for classification, regression and feature extraction
• There is a large body of alternative neural nets
• Key problems in the application of ANNs are choosing the right size and good training parameters
• Convolutional neural networks have a constrained architecture encoding prior knowledge of the problem
• Deep learning is concerned with constructing extremely large neural networks that depend on:● special hardware (GPUs), to be able to train them● specific tricks (e.g. rectified linear units) to prevent overfitting