One Exact Step

One Exact Step

The shared table renderer now carries the whole update register. Each row displays old, gradient, eta times gradient, and new value.

highlighted = computed this step

One exact step

The table applies the same rule to every shown parameter. The eta*gradient column is exact, and the new column is recomputed from old minus that column.

wnew=wold−η∇ww_{\text{new}}=w_{\text{old}}-\eta\nabla w
One exact stepExact update table with required boundary note.gradient descent update, η=1/2parameteroldgradientη·gradientneww111-2-12w121-4-23b1-1-2-10w211001w22-100-1b20000v11-4-23v21001c0-2-11w_new = w_old - η·gradientone exact step with η=1/2; NOT convergence, NOT the right η, NOT learning

Changed entries

The changed entries are w11->2, w12->3, b1->0, v1->3, and c->1.

w11=2,w12=3,b1=0,v1=3,c=1w_{11}=2,\quad w_{12}=3,\quad b_1=0,\quad v_1=3,\quad c=1
One exact stepExact update table with required boundary note.gradient descent update, η=1/2parameteroldgradientη·gradientneww111-2-12w121-4-23b1-1-2-10w211001w22-100-1b20000v11-4-23v21001c0-2-11w_new = w_old - η·gradientone exact step with η=1/2; NOT convergence, NOT the right η, NOT learning

Zero-gradient entries

Parameters with gradient 0 keep the same value in this step: w21, w22, b2, and v2.

η⋅0=0\eta\cdot0=0
One exact stepExact update table with required boundary note.gradient descent update, η=1/2parameteroldgradientη·gradientneww111-2-12w121-4-23b1-1-2-10w211001w22-100-1b20000v11-4-23v21001c0-2-11w_new = w_old - η·gradientone exact step with η=1/2; NOT convergence, NOT the right η, NOT learning