Recursive Training of 2D-3D ConvNet for Neuronal Boundary Detection

Kisuk Lee, Aleksandar Zlateski, Ashwin Vishwanathan and H. Sebastian Seung. In NIPS, 2015.



Efforts to automate the reconstruction of neural circuits from 3D electron microscopic (EM) brain images are critical for the field of connectomics. An important computation for reconstruction is the detection of neuronal boundaries. Images acquired by serial section EM, a leading 3D EM technique, are highly anisotropic, with inferior quality along the third dimension. For such images, the 2D max-pooling convolutional network has set the standard for performance at boundary detection. Here we achieve a substantial gain in accuracy through three innovations. Following the trend towards deeper networks for object recognition, we use a much deeper network than previously employed for boundary detection. Second, we incorporate 3D as well as 2D filters, to enable computations that use 3D context. Finally, we adopt a recursive architecture in which a first network generates a preliminary boundary map that is provided as input along with the original image to a second network that generates a final boundary map. Backpropagation training is accelerated by ZNN, a new implementation of 3D convolutional networks that uses multicore CPU parallelism for speed. Our hybrid 2D-3D architecture could be more generally applicable to other types of anisotropic 3D images, including video, and our recursive architecture for any image labeling problem.

ZNN - Fast 3D Sliding-Window ConvNets for Multi-Core Shared Memory Machines

Aleksandar Zlateski, Kisuk Lee and H. Sebastian Seung, 2015. (Manuscript in preparation)

ZNN (core architecture and main algorithm) was originally designed and developed by Aleksandar Zlateski. I additionally designed and developed ZNN front end and user interface, along with a modest contributoin to the core architecture. I also conducted initial experiments using ZNN reproducing the result of Masci et al. (2013) [1] to verify the correctness of ZNN implementation.



Our lab has recently released ZNN, a new ConvNet implementation that is optimized for training deep 3D ConvNets, and also specialized for dealing with 3D volume data such as serial EM images. The most striking feature of ZNN is that it exploits multi-core CPU to parallelize its computation, which is in stark contrast to the mainstream GPU-based implementations that have rapidly dominated the field of deep learning in recent years. ZNN also makes use of FFT-based convolution using fftw library. FFT-based convolution has recently been introduced to major deep learning frameworks including Torch [2,3] and Theano [4].

To the best of our knowledge, ZNN is the first deep learning implementation that is fully scalable with respect to the number of CPU cores on a single shared-memory machine. Note that there have been only a few deep learning implementations that make use of CPU instead of GPU to parallelize computation. Google’s DistBelief [5] can massively utilize tens of thousands CPU cores distributed across thousands of machines, and Jin et al. (2014) [6] has recently introduced a parallel implementation of deep learning for the Intel Xeon Phi many-core co-processor, but none of them is compatible, like ZNN, with the general purpose computing machines of moderate cost.


Experiment done on a machine with 32 physical cores, demonstrating the linear scale-up of ZNN performance as the number of threads increases.

Crowd Intelligence in EyeWire

Kisuk Lee and H. Sebastian Seung.



EyeWire [7], the world's first game to map the brain, is designed to speed up the process of analyzing EM images of neural tissue by combining machine and crowd intelligence. In this project, we aim to develop a probabilistic model for an optimal integration of the results from multiple players in EyeWire. As a preliminary step, we developed a weighted voting scheme by parameterizing each players's expertise, approximating the full probabilistic model with a simple logistic regression model in which the parameters are estimated by the stochastic gradient descent (SGD) method.

1. Masci, J., Giusti, A., Cireşan, D., Fricout, G. & Schmidhuber, J. 2013. A Fast Learning Algorithm for Image Segmentation with Max-Pooling Convolutional Networks. arXiv preprint.
2. Vasilache, N., Johnson, J., Mathieu, M., Chintala, S., Piantino, S. & LeCun, Y. 2014. Fast Convolutional Nets with fbfft: A GPU Performance Evaluation. arXiv preprint.
3. Mathieu, M., Henaff, M. & LeCun, Y. 2014. Fast Training of Convolutional Networks through FFTs. 2014. arXiv preprint.
5. Dean, J., Corrado, G.S., Monga, R., Chen, K., Devin, M., Le, Q.V., Mao, M.Z., Ranzato, M., Senior, A., Tucker, P., Yang, K. & Ng., A.Y. 2012. Large Scale Distributed Deep Networks. NIPS 2012.
6. Jin, L., Wang, Z., Gu, R., Yuan, C. & Huang, Y. 2014. Training Large Scale Deep Neural Networks on the Intel Xeon Phi Many-Core Coprocessor. IPDPSW '14. pp.1622-1630.
Unless otherwise stated, the content of this page is licensed under Creative Commons Attribution-ShareAlike 3.0 License