Feedforward networks with ternary synapses, {-1,0,1}, and unconstrained thresholds can be formed without implementing the floating point multiplier in neurons. The reduced complexity of the neurons, combined with the fact that all synapses are 2-bit deep, makes these networks very attractive for implementation in hardware. Results on the MONK's benchmark, predicting the onset of diabetes, and a hand-written digit recognition task show that these multiplier-free networks are as effective generalisers as conventional continuous-weight feed forward networks.