Single-layer network architecture with $k$ multicompartmental neurons for outputting the sum of the canonical correlation subspace projections (CCSPs) $z=(z1,…,zk)$⁠. See algorithm 2. Here $a=Wxx$ and $b=Wyy$ are projections of the views $x=(x1,…,xm)$ and $y=(y1,…,yn)$ onto a common $k$-dimensional subspace. The output, $z=M-1(a+b)$⁠, is the sum of the CCSPs and is computed using recurrent lateral connections. The components of $a$⁠, $b$⁠, and $z$ are represented in three separate compartments of the neurons. Filled circles denote non-Hebbian synapses, and empty circles denote anti-Hebbian synapses. Importantly, each synaptic update depends only on variables represented locally.