(Remark: I updated this answer completely after the OP provided more information about formatting requirements.)
The following solution works by setting up a dedicated environment called myalgo
. The "outer" equation
environment assures that there's some vertical separation between the algo material and the text above and below it, and it also provides an equation number. The "inner" minipage
environment, whose width is set to 0.9\columnwidth
, assures that each line starts flush-left within the minipage
. Instances of myalg
may be cross-referenced via the usual \label
-\ref
mechanism.

\documentclass[twocolumn]{article}
\newenvironment{myalgo}{%
\begin{equation}\begin{minipage}{0.9\columnwidth}\obeylines}{%
\end{minipage}\end{equation}}
\begin{document}
\noindent
Generally, for a simple supervised learning the \dots
\begin{myalgo} \label{alg:1}
\qquad\textit{repeat until convergence}
\{
$\displaystyle W_{i+1} = W_i - \alpha \frac{\partial}{\partial W_i} f(W_i)$
\}
\end{myalgo}
Here $W_i$ are network parameters of each \dots
A cross-reference to algorithm \ref{alg:1}.
\end{document}