Motivation. Problem: With our linear methods, we can train the weights but not the basis functions: Activator Trainable weight. Fixed basis function

Size: px

Start display at page:

Download "Motivation. Problem: With our linear methods, we can train the weights but not the basis functions: Activator Trainable weight. Fixed basis function"

Dorthy Johns
5 years ago
Views:

1 Neural Networks

2 Motivation Problem: With our linear methods, we can train the weights but not the basis functions: Activator Trainable weight Fixed basis function

3 Flashback: Linear regression

4 Flashback: Linear regression Basis functions. Sometimes called features.

5 Motivation Problem: With our linear methods, we can train the weights but not the basis functions: Can we learn the basis functions?

6 Motivation Problem: With our linear methods, we can train the weights but not the basis functions: Can we learn the basis functions? Maybe reuse ideas from our previous technology? Adjustable weight vectors?

7 Historical notes This motivation is not how the technology came about Neural networks were developed in an attempt to achieve AI by modelling the brain (but artificial neural networks have very little to do with biological neural networks).

8 Truth in advertising We usually still want to extract features from the data, but in a different way... we do not have to guess at basis functions but just use features to extract what is most relevant in our data, based on our intuition about the data...

9 Feed-forward neural network Two levels of linear regression (or classification): Output layer Hidden layer

10 Flashback: Linear regression Basis functions. Sometimes called features.

11 Flashback: Linear classification Non-linear function assigning a class.

12 Feed-forward neural network

13 Network training If we can give the network a probabilistic interpretation, we can use our standard estimation methods for training.

14 Network training If we can give the network a probabilistic interpretation, we can use our standard estimation methods for training. Regression problems:

15 Network training If we can give the network a probabilistic interpretation, we can use our standard estimation methods for training. Classification problems:

16 Network training If we can give the network a probabilistic interpretation, we can use our standard estimation methods for training. Training by minimizing an error function.

17 Network training If we can give the network a probabilistic interpretation, we can use our standard estimation methods for training. Training by minimizing an error function. Typically, we cannot minimize analytically but need numerical optimization algorithms.

18 Network training If we can give the network a probabilistic interpretation, we can use our standard estimation methods for training. Training by minimizing an error function. Typically, we cannot minimize analytically but need numerical optimization algorithms. Notice: There will typically be more than one minimum of the error function (due to symmetries). Notice: At best, we can find local minima.

19 Example tanh activation function for 3 hidden layers Linear activation function for output layer

20 Example tanh activation function for 3 hidden layers Linear activation function for output layer

21 Example tanh activation function for 3 hidden layers Linear activation function for output layer

22 Example tanh activation function for 3 hidden layers Linear activation function for output layer

23 Example tanh activation function for 3 hidden layers Linear activation function for output layer

24 Parameter optimization Numerical optimization is beyond the scope of this class you can (should) get away with just using appropriate libraries.

25 Parameter optimization Numerical optimization is beyond the scope of this class you can (should) get away with just using appropriate libraries. These can typically find (local) minima of general functions, but the most efficient algorithms also need the gradient of the function to minimize.

26 Back propagation Back propagation is an efficient algorithm for computing both the error function and the gradient.

27 Back propagation Back propagation is an efficient algorithm for computing both the error function and the gradient. We just focus on this guy

28 Back propagation After running the feed-forward algorithm, we know the values z i and y k. we need to compute the δs (called errors).

29 kz i Error at destination node (output or hidden layer) Value at source node (hidden or input layer)

30 We compute the node values from input to output (the feed-forward algorithm)

31 We compute the error values from output to input (the back-propagation algorithm)

32 Back propagation The choice of error function and output activator typically makes it simple to deal with the output layer

33 Back propagation For hidden layers, we apply the chain rule

34 Calculating error and gradient Algorithm 1. Apply feed forward to calculate hidden and output variables and then the error. 2. Calculate output errors and propagate errors back to get the gradient. 3. Iterate ad lib.

35 Calculating error and gradient Algorithm 1. Apply feed forward to calculate hidden and output variables and then the error. 2. Calculate output errors and propagate errors back to get the gradient. 3. Iterate ad lib.

36 Calculating error and gradient Algorithm 1. Apply feed forward to calculate hidden and output variables and then the error. 2. Calculate output errors and propagate errors back to get the gradient. 3. Iterate ad lib.

37 Calculating error and gradient Algorithm 1. Apply feed forward to calculate hidden and output variables and then the error. 2. Calculate output errors and propagate errors back to get the gradient. 3. Iterate ad lib.

38 Calculating error and gradient Algorithm 1. Apply feed forward to calculate hidden and output variables and then the error. 2. Calculate output errors and propagate errors back to get the gradient. 3. Iterate ad lib.

39 Calculating error and gradient Algorithm 1. Apply feed forward to calculate hidden and output variables and then the error. 2. Calculate output errors and propagate errors back to get the gradient. 3. Iterate ad lib.

40 Calculating error and gradient Algorithm 1. Apply feed forward to calculate hidden and output variables and then the error. 2. Calculate output errors and propagate errors back to get the gradient. 3. Iterate ad lib.

41 Calculating error and gradient Algorithm 1. Apply feed forward to calculate hidden and output variables and then the error. 2. Calculate output errors and propagate errors back to get the gradient. 3. Iterate ad lib.

42 Overfitting The complexity of the network is determined by the hidden layer and overfitting is still a problem. Use e.g. test datasets to alleviate this problem.

Perceptrons and Backpropagation. Fabio Zachert Cognitive Modelling WiSe 2014/15

Perceptrons and Backpropagation Fabio Zachert Cognitive Modelling WiSe 2014/15 Content History Mathematical View of Perceptrons Network Structures Gradient Descent Backpropagation (Single-Layer-, Multilayer-Networks)