I would take that with a grain of salt. "Might not be" coming from a university professor who oversees university projects funded by Intel.
The entire Open Neuromorphic community are super pro Loihi because it's Intel and are always on the look out for positions or funding from the big companies. The founder recently joined Qualcomm for example.
Keep listening past your cherry picked 32 minute section through to around the 40 minute mark. Glowing comments around akida, the ease of use, quick set up to convert CNN to SNN using MetaTF, boiler plate using edge impulse and the exceptional results.
There's a section in there where he mentioned they came second place in the hackathon and the part that the judges particularly liked was the fact they presented a $3,500 option as well as a $100 option. They go on to agree akida can be implemented at a significantly reduced price point to other technologies.
They also say results on Akida are "orders of magnitude better" when comparing to CPU platforms and there's also a question from one of the listeners around the fact akida is digital so (paraphrasing heavily here) is there any point in having asynchronous chips or analog anymore. I couldn't fully understand the discussion potentially due to a technical gap in my knowledge but also their English. Seemed to me they were discussing whether akida would make some chips redundant (why bother with them type attitude when akida is digital and power consumption results are so exceptional).
The presenter also mentions they didn't have accuracy improvements on existing technology but they also didn't spend time adjusting the code and the dataset so putting accuracy aside the power consumption results are very strong arguments for akida technology.
Strongly recommend people give this video a listen as it is a good indication of feedback from someone in the wild implementing akida and then discussing their findings.
FWIW I'm only about 40 minutes in so there could be more good or bad feedback I have not yet heard.
Hi Sera2g
Great summary. I noticed that to the right was the questions being asked by the audience and I was particularly taken by this series of questions by Gregor Lenz a prolific researcher who does a lot of stuff for NASA and Intel/Loihi. I have reproduced the questions but just his questions are gold. The answers are white gold but you will have to listen:
Gregor LenzWas this deployed to Akida? In MetaTF, you train conventional, non-spiking networks, and then convert post-training, is that correct?
Justin R
On a RASPBERRY ??
Justin R
Open NeuromorphicThanks for joining everyone!!
Please like
! It will help the video reach more people!
Gregor LenzWhats the input data like? Is it frames at fixed frequency or frames at fixed event count?
Gregor Lenzyou re saying 5W for Akida + Raspberry Pi, do you know how much power each board used?
bithighdoes spiking neural networks works only with event based cameras?
bithighprofessor anxie gave awesome lectures in THI
Wojciech Rogala
Justin R
Hi @Wojciech Rogala
Gregor Lenz If a synchronous edge chip that exploits sparsity like Akida burns just 5mW, is there a point of event-based, asynchronous systems?
bithighdo you think writing custom IP cores in HDL worths the effort for implementing custom snn?
zanz Use cases are huge!
I have highlighted the gold question which is answered by the presenter and the last response by 'znanz' is in line with the gentleman who appeared on the podcasts with Nandan from Infineon and Edge Impulse.
My opinion only DYOR
Fact Finder