Rematerialization and offloading are two well known strategies to save memory during the training phase of deep neural networks, allowing data scientists to consider larger models, batch sizes or higher resolution data.

Rematerialization and offloading are two well known strategies to save memory during the training phase of deep neural networks, allowing data scientists to consider larger models, batch sizes or higher resolution data. As these two resources are independent, it is appealing to consider the simultaneous combination of both strategies to save even more memory. We precisely model the costs and constraints corresponding to Deep Learning frameworks such as PyTorch or Tensorflow, we propose optimal algorithms to find a valid sequence of memory-constrained operations and finally, we evaluate the performance of proposed algorithms on realistic networks and computation platforms. Our experiments show that the possibility to offload can remove one third of the overhead of rematerialization, and that together they can reduce the memory used for activations by a factor 4 to 6, with an overhead below 20%.

This paper presents a novel method for the reconstruction of a neural network connectivity using calcium fluorescence data. We introduce a fast unsupervised method to integrate different networks that reconstructs structural connectivity from neuron activity. Our method improves the state-of-the-art reconstruction method General Transfer Entropy (GTE). We are able to better eliminate indirect links, improving therefore the quality of the network via a normalization and ensemble process of GTE and three new informative features. The approach is based on a simple combination of networks, which is remarkably fast. The performance of our approach is benchmarked on simulated time series provided at the connectomics challenge and also submitted at the public competition.