Add locally connected layer

[kloudkl]

The DeepFace network architecture of Facebook [1] used locally connected layer to accelerate the training process while achieving very high classification accuracy. Whether such a network is able to compete with the one proposed by Krizhevsky et. al. on more generic datasets has not been explored since its publication. If it is the case, a lot of time to train models can be saved.

Please refer to #216 for more papers.

[1] Yaniv Taigman (Facebook), Ming Yang (Facebook), Marc'Aurelio Ranzato (Facebook), Lior Wolf (Tel Aviv University). DeepFace: Closing the Gap to Human-Level Performance in Face Verification. CVPR 2014, Columbus, Ohio.
[2] Krizhevsky, A., Sutskever, I. and Hinton, G. E. ImageNet Classification with Deep Convolutional Neural Networks. NIPS 2012: Neural Information Processing Systems, Lake Tahoe, Nevada.

[kloudkl]

Related issue: #259.

[benjaminklein]

Hi, I wanted to ask if you implemented the locally connected layer and willing to share the code? Thank you.

[lixin7895123]

Hi, it's been six months from Apr 5, how about your implemention of locally connected layer?
Is it nearly done?Thank you!
@kloudkl

[shelhamer]

Follow up at #1271 for an in-progress locally-connected layer.

[rodrigob]

Seems like the push request has now moved to #3068