An adaptation of the email I just sent to those who got in touch:
A couple of points first - this guy was super nice in responding to my cold contact. Let’s not harass him or use his name unnecessarily. If you disagree then find a different academic from whom we can get a second opinion instead of going back to the same person. He has provided whatever feedback he can.
“... is an Assistant Professor in the Electrical and Computer Engineering department at Boise Sate University starting Fall 2010. His research interests are in Analog and Mixed-Signal circuit and system design. Earlier, he obtained his B. Tech. degree in Electrical Engineering from Indian Institute of Technology, Madras in 2002. He graduated with his MS and PhD from Boise State University in 2007 and 2010 respectively. In between, he has held senior design position in broadband communication system design in a start-up in India, design intern positions in Micron Imaging group (later Aptina) in Boise, ID and Lightwire Inc (now Cisco Systems) in Allentown, PA. During Spring 2010, he was a special lecturer in the ECE department at BSU teaching delta-sigma data converter design course. He has been on the organizing committee of the IEEE Midwest Symposium on Circuits and Systems and active in the local IEEE SSC and signal processing chapter.”
What they call autonomous learning is actually unsupervised learning using STDP and spiking neurons, analogous to the work by Masquelier et al. Implementing this feature is the next logical step in an FPGA, which was precisely done in this paper.
The digital neuron design is straightforward, as in existing literature. The implementation of triplet STDP with digital synapses is the basic novelty which differentiates it from TrueNorth. How this will scale to a large network with richer cognitive functionality is another question.
Overall, as an academic, I am not overtly excited with the paper, unless something more spectacular is presented.
Best regards,
PS: I understand that this personal opinion is not for a media article.
Response 2:
Yes, you can share the opinion with your investor forum.
I am not sure if a simple unsupervised learning chip makes much business sense; at most in robotic toys.
My present focus in this rather slow moving area (my primary area is hybrid CMOS photonic ICs) is to use existing hardware to explore algorithms, which need key breakthroughs. Chips are not the issue in my opinion.
Best
~~~~~~~~~~~~~~~~~~
So, my thoughts:
He addresses scalability - which they have mentioned in the roadshows. They have made the claim that Samsung etc are all interested if it can be shown to be scalable. Well. People would be interested in many things if they were possible. That statement is smoke and mirrors. Many large institutions are currently spending millions trying to make the same technology scalable. Of course the big companies are interested if it's scalable. Everyone knows that, the question is who is going to make it scalable.
He also says something similar to something that has been mentioned on HC previously (perhaps by Eshman but I can't remember - apologies) that this particular chip is actually doing 'unsupervised' learning, rather than autonomous learning. What use is learning that is unsupervised? It will have to supervised by programming, which takes us in a circle back to where other institutions are working - on the programming and algorithms. This is purely my thought process, happy to hear someone else’s thoughts on that part of His response.
Why has Peter’s paper not been submitted for review? Not critical, but worth considering.
Note: reference is made to work by Masquelier et al which may be a good lead to follow.
Cheers all, squeef.
BRN Price at posting:
32.5¢ Sentiment: None Disclosure: Held
(20min delay)