Fast Training of Convolutional Networks through FFTs
Authors: Michael Mathieu; Mikael Henaff; Yann LeCun
ICLR 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To test our analysis, we ran a series of experiments comparing our method to the Cuda Conv GPU implementation of [8] and a custom implementation using the Torch 7 machine learning environment [4]. Both of these implementations compute convolutions using the direct method in the spatial domain. All experiments were performed on the same Ge Force GTX Titan GPU. We began by performing unit tests comparing the results of convolutions computed by our method to those computed by the Torch implementation for each of the three operations. We found that the differences in results for operations (1) and (2) to be of the order of 10 5 and for operation (3) to be of the order 10 4. |
| Researcher Affiliation | Academia | Michael Mathieu Courant Institute of Mathematical Sciences New York University mathieu@cs.nyu.edu Mikael Henaff Courant Institute of Mathematical Sciences New York University mbh305@nyu.edu Yann Le Cun Courant Institute of Mathematical Sciences New York University yann@cs.nyu.edu |
| Pseudocode | No | No pseudocode or clearly labeled algorithm blocks were found in the paper. |
| Open Source Code | No | The paper does not provide any statement or link for open-sourcing the code for the described methodology. |
| Open Datasets | Yes | Even using parallel computing environments, training a network on Image Net can take weeks [8]. [...] current datasets are of the order of millions [6, 2]. |
| Dataset Splits | No | The paper does not specify how the datasets were split into training, validation, and test sets with specific percentages or counts. |
| Hardware Specification | Yes | All experiments were performed on the same Ge Force GTX Titan GPU. |
| Software Dependencies | Yes | To test our analysis, we ran a series of experiments comparing our method to the Cuda Conv GPU implementation of [8] and a custom implementation using the Torch 7 machine learning environment [4]. |
| Experiment Setup | Yes | For all experiments, we chose 96 input feature maps and 256 output feature maps, which represents a typical configuration of a deep network s second layer. [...] All configurations have minibatches of size 128. |