Learning matches Nengo learning more closely #139
Conversation
hunse
force-pushed
the
learn-nengo
branch
2 times, most recently
from
1bc9372 to
57ba8df
Compare
|
Okay, I fixed the problem from #89 with overflowing (when initializing |
|
I tried using the config system to set the weight exponent for learning connections. It wasn't pretty, and finally I gave up when core Nengo tests were failing because they added SPA-specific parameters which I then tried to copy but didn't exist in the new config objects because they weren't initialized. You can see the progress here. So now I've switched to just having an attribute on the Model to set this (see the new test). We can revisit the config stuff later if we wish. Fingers crossed everything passes now. |
|
Ok, I think I've addressed all of Dan's comments. |
|
Everything looks good on the emulator, but |
|
Ok, I've fixed As part of that, I removed the fudge factor in the learning overflow. This had originally been to make the emulator more like the chip, but I don't think it actually did. It made the chip learn significantly more slowly than the emulator. With it removed, the chip learns slightly faster. I'm still not sure why this is; it's possible it has to do with how stochastic rounding happens on the chip. Anyway, the emulator and chip are closer now, so overall it's a win I think. The two new commits can probably get squashed into existing ones if we want. I just kept them separate for now. Finally, as part of our testing for this, we should probably look at the adaptive control notebook and check it on the emulator and the chip. |
There was a problem hiding this comment.
I added a commit to fix the adaptive control notebook, and added a changelog entry.
I can't connect to the hardware at the moment, but I trust you that tests are passing there now 😄 . For the same reason I couldn't test the adaptive control notebook on hardware.
I just had one more question up above about the change to wgtExp. Not blocking or anything, that's just for my own curiosity.
Assuming that all looks good to you, this looks good to me.
|
Confirmed that the tests pass and notebook looks good on hardware 👌 |
hunse
force-pushed
the
learn-nengo
branch
2 times, most recently
from
673bd67 to
9cb8614
Compare
|
Ok, I think I've tested almost all the new added lines. There's one emulator error for spikes being dropped in the trace that is not being tested. I'm not quite sure how to test this: I think we'd need some sort of input that's spiking really fast (possibly just having a number of neurons converge on the same input axon in the post population would do it). I'd definitely want to do some sort of test that makes sure the emulator matches the chip in this respect, so I think for now I'd recommend just leaving this untested. There's also a line in |
drasmuss
force-pushed
the
learn-nengo
branch
2 times, most recently
from
bd27bcd to
d66394d
Compare
Before, we were checking that the number of learning axons was greater than the first learning index, which doesn't make sense.
By reducing `vth_max` for these connections. This is most common on learning connections with zeroed initial weights, so we'll test in `test_multiple_pes`.
This commit ensures that both the emulator and chip match Nengo in terms of basic PES learning. The emulator now does learning much more accurately, which allows us to better map PES learning parameters (i.e., learning rate, synapse decay) to the chip. This commit also makes several changes to learning tests: - The PES learning test has been simplified and explicitly compares PES learning on Nengo Loihi with core Nengo. - `test_multiple_pes` was fixed. It was not passing when using Nengo as the simulator, so it had been designed with the old Loihi learning rates in mind and needed fixing now that Loihi learning is closer to Nengo learning. - Several test tolerances have been adjusted. Note that some tolerances were changed for the chip; in general, the chip seems to learn slightly faster than the emulator (I'm not quite sure why), but this difference seems to be less apparent for more dimensions/neurons. And adds additional tests: - Add test for PES error clipping. - Add test for learning overflow. - Add test for PES learning with 'simreal' emulator. - Add test for trace increment clip warning. - Add test for discretize_weights.lossy_shift. - Add test for learning trace dropping Co-authored-by: Daniel Rasmussen <[email protected]>
This improves both the match between emulator and Loihi learning, and emulator and Nengo learning. The result is that everything matches Nengo better.
Based off of #137. Replaces #98.
TODO: I haven't tested this in extreme cases where we actually see overflow in the weights. I think some of these cases were referenced in #98. I (or someone) should look at these cases.