6.034 Artificial Intelligence - Recitations, fall 2004 online slides on learning

Next: Hidden-to-output Weights Previous: Learning in a Multilayer

Delta Rule: how to change weights

Weight update for hidden-to-output links:

\(w_{ji} = w_{ji} + (\eta * a_j * Err_i * g'(in_i))\), where

$w_{ji}$ = weight of link from node j to node i

$\eta$ = learning rate, eta, or sometimes denoted "r", for learning rate

$a_j$ = activation of node j (i.e., what the node 'outputs')
If hidden node, this is $a_j$. If input to input node, this is $I_k$.

$Err_i$ = Error at this node (target output minus actual output, how much we need to change ALL the weights to node i)

$g'(in_i)$ = derivative of the transfer function g

Let $\Delta_i$ = $Err_i$ * $g'(in_i)$ represent the error term.