via PACIFIC NORTHWEST NATIONAL LABORATORY
As anyone with a green thumb knows, pruning can promote thriving vegetation. A snip here, a snip there, and growth can be controlled and directed for a more vigorous plant.
The same principle can be applied to machine learning algorithms. Removing bits and pieces along coding branches in those algorithms can reduce complexity in decision trees and increase predictive performance.
Researchers at the U.S. Department of Energy’s Pacific Northwest National Laboratory (PNNL) have done just that. Exploring with binarized neural networks (BNNs), they used pruning principles to significantly reduce computation complexity and memory demands. BNNs are close cousins to deep neural networks, which require large amounts of computation. But BNNs differ in a significant way: they use single bits to encode each neuron and parameter, using much less energy and power for computation.
Pruning for faster growth
Researchers recognized the potential value of BNNs for machine learning starting in about 2016. If constructed–or pruned–just the right way, they consume less computing energy and are nearly as accurate as deep neural networks. That means BNNs have more potential to benefit resource-constrained environments, such as mobile phones, smart devices, and the entire Internet of Things ecosystem.
This is where pruning comes into play. As neural networks research has grown in recent years, pruning has gained more interest among computing researchers.
“Pruning is currently a hot topic in machine learning,” said PNNL computer scientist Ang Li. “We can add software and architecture coding to push the trimming towards a direction that will have more benefits for the performance of computing devices. These benefits include lower energy needs and lower computing costs.”
Pruning for precision
Li was among a group of PNNL researchers who recently published results in the Institute of Electrical and Electronics Engineers Transactions on Parallel and Distributed Systems showing the benefits of selective pruning. The research demonstrated that pruning redundant bits of the BNN architecture led to a custom-built out-of-order BNN, called O3BNN-R. Their work shows a highly-condensed BNN model–which already could display high-performing supercomputing qualities–can be shrunk signi?cantly further without loss of accuracy.
“Binarized neural networks have the potential of making the processing time of neural networks around microseconds,” said Tong “Tony” Geng, a Boston University doctoral candidate who, as a PNNL intern, assisted Li on the O3BNN-R project.
“BNN research is headed in a promising direction to make neural networks really useful and be readily adopted in the real-world,” said Geng, who will rejoin the PNNL staff in January as a postdoctoral research fellow. “Our finding is an important step to realize this potential.”
Their research shows this out-of-order BNN can prune, on average, 30 percent of operations without any accuracy loss. With even more fine tuning–in a step called “regularization at training”–the performance can be improved an additional 15 percent.
Pruning for power
In addition to this out-of-order BNN’s contributions to the Internet of Things, Li also pointed to potential benefits to the energy grid. Implementation of a modified BNN could also provide a boost to existing software that guards against cyberattacks when deployed in the power grid by helping existing sensors detect and respond to an attack, said Li.
“Basically,” said Li, “we are accelerating the speed of processing in hardware.”
The Latest Updates from Bing News & Google News
Go deeper with Bing News on:
Binarized neural networks
- The Battle of AI Processors Begins in 2018on January 16, 2021 at 4:00 pm
Neural networks are capable of performing a wide range of tasks ... operations for pruned, Int6, and binarized DNNs, respectively. ... On Ternary-ResNet, the Stratix 10 FPGA can deliver 60% better ...
- High Speed Connected Component Labeling as a Killer Application for Image Recognition Systems by Dynamically Reconfigurable Processoron January 10, 2021 at 4:00 pm
The spatial correlations in binarized input image pixels are searched and a unique ID is labeled to the group of the pixels which are neighboring each other. The connections of the pixels can not be ...
- Network structure of the mouse brain connectome with voxel resolutionon December 26, 2020 at 1:21 am
Studies examining the structural architecture of the brain have advanced our knowledge of how information is processed and integrated across distributed and specialized neural circuits. Current ...
- XMOS announces world's lowest cost, most flexible AI processoron February 13, 2020 at 5:48 am
Binarized Neural Networks Employs deep neural networks using binary values for activations and weights instead of full precision values, dramatically reducing execution time. By using Binary Neural ...
Go deeper with Google Headlines on:
Binarized neural networks
Go deeper with Bing News on:
- The 26 cannabis startups set to soar in 2021on January 15, 2021 at 10:39 am
A leading-edge research firm focused on digital transformation. Good Subscriber Account active since Free subscriber-exclusive audiobook! “No Rules Rules: Netflix and the Culture of Reinvention ...
- The Drilldown: Frosty temperatures blamed for higher oil priceson January 15, 2021 at 9:56 am
BNN Bloomberg reports that households outside Europe ... CBC News reports that Nova Scotia’s Offshore Energy Research Association wants to learn how the province could use government policy to build ...
- The Drilldown: Frosty temperatures blamed for increase in oil priceson January 15, 2021 at 9:55 am
The Lead Lower temperatures across the northern hemisphere are causing oil prices to rise, especially in Europe, CTV News reports. “Europe is the largest consumer of oil products as we speak,” says ...
- BRAIN Biotechnology Research & Information Network AGon January 14, 2021 at 4:00 pm
When autocomplete results are available use up and down arrows to review and enter to select. Touch device users, explore by touch or with swipe gestures.
- B.R.A.I.N. Biotechnology Research & Information Network AGon January 14, 2021 at 2:33 am
1 Day BNN 0.87% DJIA -0.03% Russell 2K -0.75% Health Care/Life Sciences 0.79% ...