1.6 Learning and training in the case of artificial neural networks
Learning is the process of changing the behaviour due to experience. If for people and animals this process cannot to observed directly, it is possible to study it for artificial neural networks by seing each learning step as a cause and effect relationship.
The procedure used to execute the learning process is called learning algorithm and has the role of modifying the weights of the network in order to achieve the goal. One of the most interesing characteristics of an artificial neural network is its ability to learn and improve its performance. The most important learning techniques are: supervised and unsupervised learning.
Neural Networks Learning Algorithms
A learning …show more content…
The initial values of the weights are close to zero.
The Delta Rule
Considering the following learning set:
{(u^1,y ̂_1), (u^2,y ̂_2), ..., (u^q,y ̂_q), ..., (u^Q,Q)}, where u^(q,t)=(u_1^(q,t),u_2^(q,t),…,u_m^(q,t) ) ' and y ̂_ ^(q,t)=(y ̂_1^(q,t),y ̂_2^(q,t),…,y ̂_n^(q,t) ) '
The basic idea of discovering a learning rule is to define the performance of the network and to optimize this performance. We can define the performance of the network (the error) as:
E=∑_(t=1)^T▒E_t =1/2 ∑_(t=1)^T▒〖||y ̂^t-y^t || ^2 〗
Where k- is the index of the output; t- is the index of the learning pair y ̂_k^(,t)-is the desired output, corresponding to unit k when the input is vector u^t; y_k^t=f()=f(w_k^ ' u) for k=(1,n) ̅.
Step 4. We adjust the weights: w_ik≔w_ik+γ(y ̂_k-y_k )f '(〖w^ '〗_k u)u_i
Step 5. The total error is calculated:
E≔E+1/2||y ̂^ -y^ || ^2
Step 6. Discussion:
-if tE_max, then a new cycle should start: t≔1, E≔0 and go to Step
Learning is the process of changing the behaviour due to experience. If for people and animals this process cannot to observed directly, it is possible to study it for artificial neural networks by seing each learning step as a cause and effect relationship.
The procedure used to execute the learning process is called learning algorithm and has the role of modifying the weights of the network in order to achieve the goal. One of the most interesing characteristics of an artificial neural network is its ability to learn and improve its performance. The most important learning techniques are: supervised and unsupervised learning.
Neural Networks Learning Algorithms
A learning …show more content…
The initial values of the weights are close to zero.
The Delta Rule
Considering the following learning set:
{(u^1,y ̂_1), (u^2,y ̂_2), ..., (u^q,y ̂_q), ..., (u^Q,Q)}, where u^(q,t)=(u_1^(q,t),u_2^(q,t),…,u_m^(q,t) ) ' and y ̂_ ^(q,t)=(y ̂_1^(q,t),y ̂_2^(q,t),…,y ̂_n^(q,t) ) '
The basic idea of discovering a learning rule is to define the performance of the network and to optimize this performance. We can define the performance of the network (the error) as:
E=∑_(t=1)^T▒E_t =1/2 ∑_(t=1)^T▒〖||y ̂^t-y^t || ^2 〗
Where k- is the index of the output; t- is the index of the learning pair y ̂_k^(,t)-is the desired output, corresponding to unit k when the input is vector u^t; y_k^t=f()=f(w_k^ ' u) for k=(1,n) ̅.
Step 4. We adjust the weights: w_ik≔w_ik+γ(y ̂_k-y_k )f '(〖w^ '〗_k u)u_i
Step 5. The total error is calculated:
E≔E+1/2||y ̂^ -y^ || ^2
Step 6. Discussion:
-if tE_max, then a new cycle should start: t≔1, E≔0 and go to Step