winner-take-all: one construct, many systems

In competitive learning, a set of units compete to respond to each input and only the single most strongly activated unit, the winner, is allowed to adjust its weights toward that input, so the units gradually divide the input space among themselves. Rumelhart and Zipser's 1985 Feature discovery by competitive learning formalised this winner-take-all rule and showed it can discover useful feature detectors without supervision. The referent is a selection dynamic in which a marginal lead in activation captures the entire learning signal, the same winner-take-all construct that reappears in cortical circuits and in economic markets.[1]

Unit

Which single competing unit captures the response and the update

Recovery means

Not applicable

Measured by

Observed as one unit dominating each input region and as the partition of inputs across units

In computational neuroscience a winner-take-all is a recurrent circuit in which mutual (lateral) inhibition lets the most strongly driven neuron suppress all the others, so the network's output reports the single strongest input. Maass's 2000 On the computational power of winner-take-all proved that this simple competitive operation is surprisingly powerful, able to compute functions that would otherwise need several layers of threshold units. It is the same winner-take-all construct as in competitive learning and superstar markets, here realised as a neural circuit and read out as which neuron wins.[2]

Unit

Which single neuron remains active after lateral competition

Recovery means

Not applicable

Measured by

Characterised by which unit survives inhibitory competition and by the circuit's selection accuracy

In economics a winner-take-all market is one where rewards depend on relative rather than absolute performance, so a slight edge in talent or luck translates into a hugely disproportionate share of the rewards and a few top performers capture most of the gains. Frank and Cook's The winner-take-all society argued that such markets have spread well beyond sports and entertainment, concentrating income and drawing too many contestants toward a handful of big prizes. It is the same winner-take-all construct (a marginal lead capturing a disproportionate or total share), expressed in a market rather than a circuit or a learning rule.[3]

Unit

The share of rewards captured by the top performer(s)

Recovery means

Not applicable

Measured by

Indexed by how steeply reward rises with rank and by the concentration of rewards at the top