Has anyone investigated implementimg neural nets in FPGAs?
- posted
18 years ago
Has anyone investigated implementimg neural nets in FPGAs?
More importantly, has anybody found a use for neural nets?
John
I use mine daily....
Tryin to avoid some gardening work I had to check that out.
I only see upto p76 so 760 of 9880 refs.
Even after turning off a redundancy warning it goes to 980 out of 11600 so I guess the other 9-10k can't be reached or can they?
Is it possible to random access out there in the 9880 boonies, I don't usually go further than 1st few pages. And google says it won't go past
1000. Hint directly type in over "num=300" to jump but 1000 gate still there.Anyway to save the whole thing to file in 1 go? (after change settings to 100 items per page)
johnjakson at usa dot com
Yes,
I've very likely that the letter you receive are sorted by zip code using a neural nets. That's only one example in a million others
Nick
OK - Let me rephrase my question.
Has anyone who frequents this newsgroups and has a lot of experience with FPGAs, actually tried to make a functioning NN system and has some realworld experience with FPGA NN Systems. (Mean as it may sound , this is meant to rule out present or past graduate students that only got as far simulation.)
JJ wrote:
More likely than not, most regulars here probably have not dipped their finger into NNs as they are probably busy more in DSP, or embedded space.
But google fpga neural networks on the web did highlight at least a couple of may be interesting articles for you to start with. Those authors may or may not have been here too but AFAIK I haven't seen NN come up that often.
google groups too to see where they hang out
regards
johnjakson
Many (most?) classical neural networks utilise non-linear functions at the nodes, with fractional synapse weights and so on. In SW implementations, floating point is the order of the day.
Add in the back-propagation training algorithms and you have significant non-integer arithmetic to contend with.
For implementation in commodity FPGA HW, this will all hurt. Fixed point will help, but bring similar precision issues that arise in DSP.
Perhaps there is research on pure digital neural networks (binary weights, logical and/or/xor node functions) etc? Dig around in the evolvable hardware research literature, they've been banging on it for years.
It is my expectation (not experience), that there will be significant practical issues in implementing reasonable sized classical NNs on FPGA hardware. You will very quickly find yourself building either huge arrays of FPGAs, or diving into dynamic reconfig / multicontext FPGA territory to get the logic coverage required to implement what will be a very large (virtual) circuit.
Or, implement a couple of NN nodes in your FPGA fabric, with some sort of controller updating the weights and accumulating responses at each node. Use this to simulate the entire massive (and parallel) NN operation.
Surprise surprise it's time/area trade.
Lots of fun no doubt, but not trivial either.
John
Take a look at GenoByte,
News is old though, ca. 2000.
-- Pete
e wrote:
I have a PhD student (Jihan Zhu) working in this area. There have been many implementations of NN's in FPGAs over the years. The first seems to have been Cox, C.E. and E. Blanz, GangLion - a fast field-programmable gate array implementation of a connectionist classifier. IEEE Journal of Solid-State Circuits, 1992. 28(3): p. 288-299.
A very brief (4 page) survey appeared in FPL in 2003: J. Zhu and P. Sutton, "FPGA Implementations of Neural Networks - a Survey of a Decade of Progress," in Proceedings of 13th International Conference on Field Programmable Logic and Applications (FPL 2003), Lisbon, Sep 2003.
It's online at
Peter
I saw a paper maybe 10 years ago on a neural net set up in an FPGA (small by todays standards), It used bit serial arithmetic and was able to fit more nodes than one might have thought. IIRC, this was an early xilinx 4K family.
-- --Ray Andraka, P.E. President, the Andraka Consulting Group, Inc.
You might google "Federico Faggin" (my boss at Zilog 26 years ago) or Synaptics, his later company... Peter Alfke
ElectronDepot website is not affiliated with any of the manufacturers or service providers discussed here. All logos and trade names are the property of their respective owners.