IBM developed a new chip from neural net to make AI 100 times more energy-efficient

IBM Artificial Synapses

Have you not upgraded your website to HTTPS yet? Upgrade NOW.

Google with its Chrome 68 update to show all HTTP websites as NOT SECURE. Avoid Google's penalty by installing an SSL Certificate. Get a DigiCert Standard SSL and secure your website at just $157/year. BUY NOW

ADVERTISEMENT
DAILY BRIEF
Get daily updates straight in your inbox.

IBM has brought the key features of a neural network directly into silicon in order to make it 100 times more efficient than the standard existing chips. It is expected that chips built in this way will magnify machine learning attributes in the coming years.

Neural networks are the main element for AI boost, gorge on data and are engaged in transcribing speech and describing images with almost cent percent accuracy. The neural nets are modeled loosely on human brain structure, and typically constructed in software and not hardware. Now, the software running on the conventional computer chips slows things down.

This is where IBM researchers have come up with a research paper published in the journal Nature, talking about microelectronic IBM artificial synapses. Their developed chips can mimic synapses that connect individual neurons in a human brain. Monitoring and boosting the strengths of these synaptic connections are required for the network to learn and grow.

A living brain hosts this process in form of growing or withering connections over time. This is easy to reproduce in software, although the opposite of easy and more to do the same in hardware so far. IBM, therefore, becomes the pioneer in coming up with the advanced science.

Its researchers have been inspired by neuroscience, and the approach includes using two types of synapses. These are short-term IBM artificial synapses for computation and long-term ones for memory. According to Michael Schneider, a researcher at National Institute of Science and Technology, this process addresses several key issues including low accuracy being the main one.

Related

A test was conducted on the new chips with two basic image-recognition tasks: color image classification and handwriting. It was observed that the system performed as accurately as a software-based deep neural network. The best part, it only consumed 1% of the energy.

The system will be useful not just for AI, but as well in other sectors, like IBM’s recent efforts to reinvent computer hardware. The company hopes that the new microelectronic components may be useful for the coming advancements.

The new method is potent enough to be a mark in history for future technological developments, but it may also remain underdeveloped due to ignorance from the industry leaders. This is where IBM needs to ideally market its innovation.

Adding to a probable negativity is the design of the chips that are still relatively clunky. It has five transistors and three other components, and there should be only one transistor in a chip. Also, certain aspects of the system have been only simulated for testing.

Given all that, IBM still needs to develop a complete chip, addressing all relevant issues, before marketing. The innovation is indeed prospectus in terms of computing and AI. IBM with required relevant actions in the coming days may just have written the first chapters of this history book!

Via: MIT Technology Review

IBM developed a new chip from neural net to make AI 100 times more energy-efficient