Pushing AI Over the Edge: A Scalable Approach to the Mass Deployment of DL
Details
Deep learning has captured everyone’s imagination as a deus ex algorithmus. A future of intelligent machines interacting with us and teaching us all about what it means to be human used to be the domain of the science fiction novels and Hollywood blockbusters, now it seems to be on the 8 o'clock news.
It seems no one doubts that we are on the verge of having this exciting technology at the tip of our fingers - but are we? Much like “A good science fiction story should be able to predict not the automobile but the traffic jam.”, at Hailo we focus not on the self-driving automobile but on the computational bottleneck. To truly bring about the promise of ubiquitous deep-learning we need to solve the problem of efficient hardware to run all the underlying computations.
In this meetup we will address a couple of the issues surrounding the design of efficient deep learning processors:
-
We’ll analyze what makes neural networks a unique computational problem and raise the question - Are current solutions good enough? We claim that the answer is NO and that the underlying issues can't be engineered away. Instead, a complete re-design of the computational foundations is needed.
-
8-bit deep learning: A simple way to make deep learning more efficient is to convert the network to do calculations in 8-bits - a process called Quantization. We’ll overview how this is achieved and showcase two new methods for improving quantization developed at Hailo.
Speakers –
Avi Baum, CTO & Co-Founder
Alex Finkelstein, Machine Learning Engineer
SAME SAME BUT DIFFERENT: A valuable research paper by Hailo's Machine Learning team published on arXiv on quantization of neural networks. In the paper we explore a new approach to reducing quantization error using inversely proportional factorization.
Read the full article here: https://arxiv.org/abs/1902.01917
While FREE, please sign up to save your spot.