Hossein Askari

I am a Phd student at Ecole Polytechnique de Montreal . I am working under Prof. Jean-Pierre David supervision. Previously, I was an ASIC Verification engineer at Microsemi (Now MicroChip) where I worked on the next generation of OTN (Optical Transport Network) processors. And before that, I was a Software Engineer at Tru Simulation + Training.

My major areas of interests are:
  • Making Deep Neural Networks more computationally efficient
  • Deep Learning Acceleration

Email  /  Google Scholar




Papers

prl

U-Net Fixed Point Quantization For Medical Image Segmentation
MICCAI 2019
/

In this work, we present a fixed point quantization method for the U-Net architecture, a popular model in medical image segmentation. We then applied our quantization algorithm to three different datasets and comapred our results with the existing work. Our quantization method is more flexible (different quantization level is possible) compared to existing work.


prl

Towards code generation for ARM Cortex-M MCUs from SysML activity diagrams
ISCAS 2016
/

In this work, we redefined New Activity Calculus (NuAC) terms to support code generation for ARM Cortex-M processors and we present an automated SysML activity diagram to RTX (Keil Real-Time Operating System) code generator that uses mapping rules expressed in NuAC.


prl

Formal modeling, verification and implementation of a train control system
International Conference on Microelectronics (ICM) 2015
/ /

In this paper, we verified a train control system for safe speed and acceleration limits. We verified different properties of the model. For this task, we verified model properties using the NuSMV model checker ( a symbolic model checker tool). We then implement the algorithm of the verified model on an ARM CortexM platform.


prl

Duplication Avoidance for Energy Efficient Wireless Sensor Networks
CSNDSP 2012

Aduplication avoidance method is introduced for wireless sensor network that aims for high coverage area and low control overheads. To preventi energy holes creation the periodic selected transmitters are distributed uniformly.




Projects

prl

Project Logic Brain : A Binary Neural Networks Accelerator on FPGA
Rapid Prototyping Course Project
/

In this project, we designed a Binary Neural Network Accelerator. We used Intel Cyclone V FPGA as the target platform. With no modifications, the accelerator is capable of accelerating MLP Networks and with some modifications, it is capable of accelerating CNNs. We have shown that compared to a software model (that runs on a NIOS II processor @100Mhz), our implementation can run upto 14K times faster (@100Mhz).


prl

ICLR 2018 Reproducibility Challenge
Training and Inference with Integers in Deep Neural Networks
/ (original paper) /

We reproduce Wu et al.'s ICLR 2018 submission Training And Inference With Integers In Deep Neural Networks. The proposed `WAGE' model reduces floating-point precision with only a slight reduction in accuracy. The paper introduces two novel approaches which allow for the use of integer values by quantizing weights, activation, gradients and errors in both training and inference. We reproduce the WAGE model, trained on the CIFAR10 dataset. The methodology demonstrated in this paper has applications for use with Application Specific Integrated Circuit (ASICs).


prl

Celebrity Face Generation using GAN
Deep Learning Course Assignment
/

We implemented a GAN with latent dimension 100, in order to generate novel faces from the CelebFaces Attributes Dataset. The dataset downloaded contains nearly 10000 images. A DCGAN architecture was used.


prl

Neural Turing Machine implementation
Deep Learning Course Assignment
/

We implemented the copy task of a Neural Turing Machine that was proposed by Deep Mind. The original paper can be found here. We implemented the MLP_NTM and LSTM_NTM and we compared the results with a LSTM model.


prl

Cats VS. Dogs (Kaggle challenge)
Deep Learning Course Assignment
/

In this assignment, we implemented a CNN model to detect if the image contains a dog or a cat. We also had to provide some missclassified images and propose methods to prevent such scenraios. Finally, we implementd a focus detection algorithm to show what features in the original image are really important for the network to correctly classify the image.


prl

Implementation of BinaryConnect
(theano) / (pytorch)

BinaryConnect introduced a quantization method to use -1 and +1 as the only values for parameters in a Neural Network. This dramatically reduces the footprint needed to store the parameters for a Neural Net. For most models, this can be achieved with negligible drop in accuracy.