The proton structure function F2 is measured from deep inelastic positron-proton scattering in the kinematic range 5 ≤ Q² ≤ 40 GeV² and 1 · 10-4 ≤ x ≤ 6 · 10-3 using the '94 ZEUS data. This measurement extends the kinematic region so far studied to lower values of x and Q². For the event selection a neural network is used which efficiently identifies the scattered positron in the final state and reduces the contribution from background events significantly. For the training of the neural network a new approach to error function minimization is developed. When compared to the F2 measurement using the '93 ZEUS data the errors on the F2 are typically reduced by a factor 2. It is further shown that the value of the predicted power behavior of F2 as a function of the x at fixed Q² from the BFKL leading-order approximation does not match the data for 5 ≤ Q² ≤ 40 GeV² when fitting a function F2 ∼ x-λ. The double-leading-logarithmic approximation of the DGLAP evolution equation reproduces the Q² behavior of λ within errors. The predicted Q² behavior of F2 at fixed values of y from the BFKL and the DGLAP evolution equations are compared and shown to differ significantly in their curvature. It is demonstrated that the current errors on F2 need to be reduced by at least an additional factor 2 at lower Q² and by a factor ∼ 4 at larger Q² to uniquely distinguish between the predicted behavior of F2 from both evolution equations.