New #neuromorphic #SpikingNeuralNetworks paper with Pengfei Sun, Zhe Su, nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpqwmycn7r6cnhp5rlg63rp3vm40vex70hl5gk7dzzrdtcqzdr29mhq070g2l (nprofile…0g2l), nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpqnvp92ljgp9mrztfh5gmwld2an0lmvyn74ykryh3sqsfyqp2jjyss35wyuq (nprofile…wyuq) and Danyal Akarca:
https://neural-reckoning.org/pub_dual_memory_pathways.html
Long story short, we found a neat new trick for augmenting spiking neural networks with a tiny separate memory buffer, that gives a huge boost to performance on tasks with temporal structure at a very low memory cost (state-of-the-art performance with 40-60% fewer parameters). It's also very neuromorphic hardware-friendly and gives us 4x increased throughput and 5x improvement in energy efficiency compared to state-of-the-art.
We think this is (a) a great example of where an abstraction inspired by biology can lead to networks that are both algorithmically effective and hardware-efficient, (b) shows the power of doing hardware-software co-design for neuromorphic computation.
Dan Goodman
npub1pm…8q5qg
2026-06-18 16:34:36 UTC
Author Public Key
npub1pmxyaxv6dm8xnd99c80fqyvtjmsmp68gv026tdyfsyz4000ttk0s48q5qgSeen on
wss://relay.ditto.pubPublished at
2026-06-18 16:34:36 UTCEvent JSON
{
"id": "8cd7b2aea6e9f7ce466899e5baeaada4d4c15aa79c03e9d96393b6d2e8cbd533",
"pubkey": "0ecc4e999a6ece69b4a5c1de90118b96e1b0e8e863d5a5b489810557bdeb5d9f",
"created_at": 1781800476,
"kind": 1,
"tags": [
[
"p",
"76c989f87ac4ee1a0fe8d44618b3757b326f3effa22de688436af001346a2eee",
"wss://relay.ditto.pub"
],
[
"p",
"9b02557e480976312d37a236efb55d9bffb6127ea92c325e3004124005529121",
"wss://relay.ditto.pub"
],
[
"p",
"568061e6dd0da1602911857cac17a4aad65f5de419299f6cda9af2320a512a0a",
"wss://relay.ditto.pub"
],
[
"p",
"8a6a09bf64610523ce0144029918925b55528244c28973b6dd3ae14c8f89e7a0",
"wss://relay.ditto.pub"
],
[
"t",
"neuromorphic"
],
[
"t",
"spikingneuralnetworks"
],
[
"proxy",
"https://neuromatch.social/users/neuralreckoning/statuses/116772076021474725",
"activitypub"
],
[
"client",
"Mostr",
"31990:6be38f8c63df7dbf84db7ec4a6e6fbbd8d19dca3b980efad18585c46f04b26f9:mostr",
"wss://relay.ditto.pub"
]
],
"content": "New #neuromorphic #SpikingNeuralNetworks paper with Pengfei Sun, Zhe Su, nostr:nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpqwmycn7r6cnhp5rlg63rp3vm40vex70hl5gk7dzzrdtcqzdr29mhq070g2l, nostr:nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpqnvp92ljgp9mrztfh5gmwld2an0lmvyn74ykryh3sqsfyqp2jjyss35wyuq and Danyal Akarca:\n\nhttps://neural-reckoning.org/pub_dual_memory_pathways.html\n\nLong story short, we found a neat new trick for augmenting spiking neural networks with a tiny separate memory buffer, that gives a huge boost to performance on tasks with temporal structure at a very low memory cost (state-of-the-art performance with 40-60% fewer parameters). It's also very neuromorphic hardware-friendly and gives us 4x increased throughput and 5x improvement in energy efficiency compared to state-of-the-art.\n\nWe think this is (a) a great example of where an abstraction inspired by biology can lead to networks that are both algorithmically effective and hardware-efficient, (b) shows the power of doing hardware-software co-design for neuromorphic computation.",
"sig": "15c4a7e51e2a5e9b36ab2066196ed3bcb35607e0cb653ee21e3bb55009b65f7d81f21019736b71bce2cc8e90914b43e990bdc5de9ffb6db98090cc17868b04f3"
}