# Conclusion

The purpose of this chapter was to provide a general review of neural networks’ working principles, discuss their evolution through time and interrogate on ways to reduce the data needed to perform object recognition, while ensuring good results are achieved.

From the research it is clear that NNs have evolved greatly since their inception: some key developments, such as the use of backpropagation, deeper networks, ReLU functions and transfer learning have revolutionised the original model, and allow today to surpass human accuracy on certain tasks. These models, however, require large amounts of data, often in the order of thousands of examples per class.

The siamese network architecture was discussed, along with the successful triplet loss. As confirmed by the literature, this type of network effectively allows to perform one or few-shot recognition, alleviating the burden of collecting and preparing large size datasets. In addition, this type of network can recognise new, unseen classes without the need for retraining.

Even though siamese networks have achieved a discrete success, they have only been used in a handful of occasions for object recognition. Further research is felt required, since there is still margin to improve. Also, a deeper investigation would be helpful to understand the limitations of these networks.

Finally, other approaches for performing one or few-shot recognition were discussed. k-nearest neighbour is simple to implement and makes a good baseline method. Matching networks improve this model, providing new insights for training a fully differentiable nearest-neighbour algorithm.