Diary/2019-3-11
R-Wonc'19
というのを聴講.
https://usability-research.r-ccs.riken.jp/r-wonc19/
- Giacomo Indiveri, Neural processing and learning electronic circuits for building neuromorphic cognitive agents
- intro
- driven by AI
- driven by Big-Data https://www.nature.com/articles/d41586-018-01683-1
- Bee brain specs
- 1mg, 1mm^3, 960,000 neurons, 10^{-15} J/spike energy/op
- Neuromorphic processor chips
- spikes in, and spikes out
- analog subthreshold circuits
- inhomogeneous, imprecise, and noisy,
- massively parallel
- DYNAP-SEL-Dynamic Neuromorphic Asynch Processor with Self Learning
- Adaptive CardioRespiratory Pacemaker EU project
- Mike Davies, A New Era of Neuromorphic Computing
- cf. https://newsroom.intel.com/editorials/intel-creates-neuromorphic-research-community/
- Background (pict.)
- The Engineering Perspective (pict.)
- LOIHI: https://www.computer.org/csdl/magazine/mi/2018/01/mmi2018010082/13rRUIJcWtw
- https://www.researchgate.net/publication/322548911_Loihi_A_Neuromorphic_Manycore_Processor_with_On-Chip_Learning
- 128 neuromorphic cores, 128k neurons, 128M synapses
- 14nm FinFET
- cf. http://niceworkshop.org/wp-content/uploads/2018/05/Mike-Davies-NICE-Loihi-Intro-Talk-2018.pdf
- Loihi systems
- wolf mountain, Nahuku, Kapoho Bay, ...
- SNN algorithms discovery and development
- Speech Recognition: Keyword spotting
- Spiking LCA dynamics
- Spike-based LSTMs LSNN
- cf. Long short-term memory and learning-to-learn in networks of spiking neurons https://arxiv.org/abs/1803.09574
- cf. Biologically inspired alternatives to back-propagation through time for learning in recurrent neural nets https://arxiv.org/abs/1901.09049
- cf. Adaptive Control of a Robot Arm using Loihi https://royalsocietypublishing.org/doi/full/10.1098/rspb.2016.2134
- Graph search
- Olfacton-Inspired One Shot Learning
- cf. olfaction inspired machine learning https://arxiv.org/pdf/1802.05405.pdf
- Why Spikes: (pic.)
- Cerebellum
- a little brain
- cf. Cerebellar ataxia - https://www.youtube.com/watch?v=Txlvuu2byUY
- simulation by CPU, GPU, FPGA, PEZY
- GPU, FPGA, PEZY - realtime simulation
- human-scale cerebellum on K
- 68 billion neurons on 82,944 nodes
- MONET (in-house simulator)
- 600 time slower than realtime (10min. for 1s)
- application
- arm control
- reinforcement learning in cerebellum
- reinforcement learning can go in parallel (massively)
- - cf. Hybrid Reward Architecture for Reinforcement Learning https://arxiv.org/abs/1706.04208
- - cf. Hybrid Reward Architecture https://github.com/Maluuba/hra
- A Benchmarking and Programming Framework for Spiking Neuromorphic Computing Systems
A Survey of Neuromorphic Computing and Neural Networks in Hardware - https://arxiv.org/abs/1705.06963
- 3 examples (pict.)
- TENNLab sotware framework
- Types of Neuromorphic Implementations
- DANNA 2- fuly digital https://dl.acm.org/citation.cfm?id=3229894
- mr DANNA - mixed analog-digital
- SOEN - optoelectronic
- VLSI Research for Neuromorphic Computing in IBM Research
- spiking/non-spiking neural network, digital/analog implementation
- spiking/digital - TrueNorth
- non-spiking/digital - GPGPU, FPGA, FPU arrays
- spiking/analog - Spiking neural network chips w/ non-volatile memory arrays
- non-spiking/analog -
- A Scalable Multi-TeraOPS Deep Learning Processor Core for AI Training and Inference - https://ieeexplore.ieee.org/document/8502276, https://xpressdrivein.org/glo16/pdf/C04-2.PDF
- approximate computing by reduced precision computations
- A million spiking-neuron integrated circuit with a scalable communication network and interface - http://science.sciencemag.org/content/345/6197/668 http://paulmerolla.com/merolla_main_som.pdf
- 1M neurons, 256M synapses tileable 2D-onchip
- NVM synaptic array - eg. phase change memory (pict.)
- Analog multiply accumulation with non-volatile memory array
- NVM Weight Variation Imact on Analog Spiking Neural Network Chip - https://link.springer.com/chapter/10.1007/978-3-030-04239-4_61
- Lightweight Refresh Method for PCM-based Neuromorphic Circuits - https://www.semanticscholar.org/paper/Lightweight-Refresh-Method-for-PCM-based-Circuits-Ito-Ishii/5ae6ddfcf1820a7c1289d97197c09b598bc0724a
- AI on the Edge: Frontiers for Energy-Efficient Hardware Architectures
- "Structure" is a key
- Binary/Ternary DNN accelerator VLSI 2017
- Binary/Ternary, Reconfigurable in Memory
- cf. https://www.researchgate.net/publication/321930684_BRein_Memory_A_Single-Chip_BinaryTernary_Reconfigurable_in-Memory_Deep_Neural_Network_Accelerator_Achieving_14_TOPS_at_06_W
- Log-Quantized DNN accelerator with 3D SRAMs
- QUEST(Log QUantization, MIMD Parallel Engine, Die-STacking with SRAMs)
- cf. Convolutional Neural Networks using Logarithmic Data Representation - https://arxiv.org/pdf/1603.01025.pdf
- Dynamically reconfigurable processor with AI-MAC engine
- DRP(96-core, 333MHz) + AI-MAC(1024, 500MHz)
- The Era of "Intelligence at the Edge" will Begin
- Common Key Features: Mostly static, dataflow rich, (self) evolvable
- Procedure Oriented Computing -> Structure Oriented Computing: Reconfigurable HW, "Virtualized" Reconfigurable HW (post FPGA), Dataflow Oriented Machine (w/ reduced synth. cost)
- Towards biologically plausible learning of spike-based cognition
- CT-AuGMEnt
- Stochastic Computing for Brainware LSI
- 関連する?
- Large-Scale Neuromorphic Spiking Array Processors: A Quest to Mimic the Brain - https://www.frontiersin.org/articles/10.3389/fnins.2018.00891/full
- Adaptive motor control and learning in a spiking neural network realised on a mixed-signal neuromorphic processor - https://arxiv.org/pdf/1810.10801.pdf
- Dual Supervised Learning - https://arxiv.org/abs/1707.00415
- Implementation of a Liquid State Machine with Temporal Dynamics on a Novel Spiking Neuromorphic Architecture - https://www.osti.gov/servlets/purl/1405258