Are binary synapses superior to graded weight representations in stochastic attractor networks?