Pushing the Boundaries of AI, Cheaply and Efficiently: Murat Onen Explains
Finding Genius Podcast
Richard Jacobs
4.4 • 1K Ratings
🗓️ 27 November 2022
⏱️ 32 minutes
🧾️ Download transcript
Summary
Large-scale AI models that enable next-generation applications like natural language processing and autonomous systems require intensive training and immense power. The monetary and environmental expense is too great.
This is where analog deep learning comes into play. The concept behind it is to develop a new type of hardware that can accelerate the training of neural networks, achieving a cheaper, more efficient, and more sustainable way to move forward with AI applications.
Murat Onen, a postdoctoral researcher in the Department of Electrical Engineering and Computer Science at MIT, explains.
Tune in to explore:
- Conventional vs. novel methods of training neural networks
- The difference between GPUs and CPUs and why it matters
- Analog vs. digital machine operations
- About how long it will take to have small and full-scale systems that outperform conventional AI models
Press play for the full conversation.
Episode also available on Apple Podcasts: http://apple.co/30PvU9C
Transcript
Click on a timestamp to play from that location
| 0:00.0 | So the goal is basically used to use CPUs for everything, for a long while, and they're excellent, |
| 0:05.5 | they're very flexible architectures in digital. And then the field of deep learning started |
| 0:10.8 | getting traction and we said, oh, okay, like these matrix algebra that we want to do actually |
| 0:15.5 | are much more efficient to be done on a GPU instead of a CPU, right? So we did that transition |
| 0:21.8 | more than a decade ago, I guess, by now. And reason we did that was, again, to be able to |
| 0:26.2 | handle more complex deep learning applications for artificial intelligence. |
| 0:30.1 | Forget frequently asked questions. How about advice from a real genius? 95% of people in any profession |
| 0:39.2 | are good enough to be qualified in license. 5%? Go above and beyond. They become very good at what |
| 0:44.3 | they do, but only 0.1% are real geniuses. Richard Jacobs has made his life's mission to find them |
| 0:51.7 | for you. He hunts down and interviews geniuses in every field, sleep science, cancer, |
| 0:56.3 | stem cells, ketogenic diets, and more. Here come the geniuses. This is the Finding Genius Podcast. |
| 1:02.8 | The Richard Jacobs. |
| 1:07.9 | Hello, this is Richard Jacobs with the Finding Genius Podcast, that part of the Finding Genius |
| 1:12.8 | Foundation. My guest today is Mirat Honen, PhD. He's a postdoctoral researcher in the Department |
| 1:18.8 | of Electrical Engineering and Computer Science at MIT. He's a Massachusetts Institute of Technology. |
| 1:24.5 | He's working on a project where he's creating artificial synapses that are 10,000 times faster |
| 1:30.4 | than biological ones. So this sounds, I don't know, like it has amazing potential. We'll see. |
| 1:35.6 | But Mirat, thank you for coming. Thank you very much for having me. |
| 1:38.8 | Yeah. Tell me about this project, like how did it start and what's your involvement in it? |
| 1:44.6 | Yeah, I mean, I would say that the broader field that we work in is called as analogue deep learning. |
| 1:51.7 | And the core concept here is to come up with a new type of hardware that can accelerate |
| 1:58.2 | training of neural networks, right? So basically going back to the problem itself, |
... |
Please login to see the full transcript.
Disclaimer: The podcast and artwork embedded on this page are from Richard Jacobs, and are the property of its owner and not affiliated with or endorsed by Tapesearch.
Generated transcripts are the property of Richard Jacobs and are distributed freely under the Fair Use doctrine. Transcripts generated by Tapesearch are not guaranteed to be accurate.
Copyright © Tapesearch 2026.