no code implementations • 4 May 2022 • Kaveena Persand, Andrew Anderson, David Gregg
The removal of these weight slices from a single layer causes mismatching number of feature maps between layers of the network.
no code implementations • 25 Jan 2022 • Syed Asad Alam, Andrew Anderson, Barbara Barabasz, David Gregg
The choice of points impacts the numeric accuracy of the algorithm, but the optimal set of points for small convolutions remains unknown.
no code implementations • 21 May 2020 • Yuan Wen, Andrew Anderson, Valentin Radu, Michael F. P. O'Boyle, David Gregg
We optimize the trade-off between execution time and memory consumption by: 1) attempting to minimize execution time across the whole network by selecting data layouts and primitive operations to implement each layer; and 2) allocating an appropriate workspace that reflects the upper bound of memory footprint per layer.
no code implementations • 3 Apr 2020 • Kaveena Persand, Andrew Anderson, David Gregg
In most cases our method finds better selections than even the best individual pruning saliency.
no code implementations • 9 Jan 2020 • Andrew Anderson, Jing Su, Rozenn Dahyot, David Gregg
Hardware-Software Co-Design is a highly successful strategy for improving performance of domain-specific computing systems.
no code implementations • 11 Jun 2019 • Kaveena Persand, Andrew Anderson, David Gregg
The result is that it is difficult to separate the effectiveness of the saliency metric from the wider pruning algorithm that surrounds it.
no code implementations • 22 Mar 2019 • Andrew Anderson, Jonathan Dodge, Amrita Sadarangani, Zoe Juozapaitis, Evan Newman, Jed Irvine, Souti Chattopadhyay, Alan Fern, Margaret Burnett
We present a user study to investigate the impact of explanations on non-experts' understanding of reinforcement learning (RL) agents.
no code implementations • 15 Jan 2019 • Miguel de Prado, Jing Su, Rabia Saeed, Lorenzo Keller, Noelia Vallez, Andrew Anderson, David Gregg, Luca Benini, Tim Llewellynn, Nabil Ouerhani, Rozenn Dahyot and, Nuria Pazos
In this work, we present a modular AI pipeline as an integrating framework to bring data, algorithms, and deployment tools together.
no code implementations • 27 Sep 2018 • Andrew Anderson, David Gregg
Our approach yields very fast implementations of bit-precise custom DNN operations, which often match or exceed the performance of operations quantized to the sizes supported in native arithmetic.
no code implementations • 29 Mar 2018 • Barbara Barabasz, Andrew Anderson, Kirk M. Soodhalter, David Gregg
We propose several methods for reducing FP error.
no code implementations • 3 Oct 2017 • Andrew Anderson, David Gregg
We show experimentally that significant gains are possible versus the state of the art vendor libraries by using a principled analytic solution to the problem of layout selection in the presence of data format transformations.
no code implementations • 8 Sep 2017 • Andrew Anderson, Aravind Vasudevan, Cormac Keane, David Gregg
We present two novel GEMM-based algorithms that require just $O(MHW)$ and $O(KW)$ additional space respectively, where $M$ is the number of channels in the result of the convolution.
no code implementations • 6 Apr 2017 • Aravind Vasudevan, Andrew Anderson, David Gregg
A common approach to implementing convolutional layers is to expand the image into a column matrix (im2col) and perform Multiple Channel Multiple Kernel (MCMK) convolution using an existing parallel General Matrix Multiplication (GEMM) library.