Jean Anne Incorvia: Using Magnetic Spin Textures for Cognitive Computing (Invited)
HTML-код
- Опубликовано: 8 сен 2024
- 2022 IEEE AtC-AtG Magnetics Conference Session 12
Jean Anne Incorvia, University of Texas at Austin, USA
For more information, please visit the website below.
www.atc-atg.or...
0:00 Speaker Introduction
0:01:16 Spintronics on Altermagnetism
0:01:40 Introduction
0:02:33 Outline
0:03:25 New Computing Era for Artificial Intelligence uses Deep Neural Networks (NNs) and Beyond-Von-Neumann Architectures
0:04:12 Neuromorphic Computing uses Artificial Neurons and Synapses in NNs and Other Architectures
0:04:34 Neuromorphic Computing can be implemented in CMOS e.g. as spiking neural networks
0:05:00 AI faces multiple challenges
0:05:54 Magnetic materials provide
0:06:57 Magnetic tunnel junction (MTJ): basic building block of MRAM
0:07:16 Magnetic Tunnel Junction Markets
0:07:45 Magnetic computing basic building block 2: domain walls
0:08:01 Magnetic computing basic building block 2: skyrmions
0:08:17 Magnetic computing state-of-the-art
0:08:44 Device of the focus here: three terminal domain walls-magnetic tunnel junctions (DW-MTJs)
0:09:41 Domain wall or skyrmions based magnetic devices face challenges of high current density and low TMR
0:10:36 DW-MTJ prototypes down to 250 nm feature size using spin-orbit torque (SOT) switching
0:11:18 Prototypes showing high TMR and minimal damage
0:12:11 With Lowered Switching Current Density and Maintained TMR, Device-to-Device and Cycle-to-Cycle Data can be obtained
0:13:59 Switching Current and Energy
0:15:00 Neuromorphic computing building blocks
0:15:05 Long MTJ allows DW-MTJ to act as an artificial synapse
0:15:15 Neural network crossbar array (one layer)
0:15:31 Best Synapse Behaviour Depends on the Task
0:16:27 Long MTJ allows DW-MTJ to act as an artificial synapse
0:16:48 Straight DW-MTJ artificial synapse shows multiple stable but stochastic states
0:17:32 Input straight DW-MTJ synapse data into CIFAR-100 Dataset Classification Test
0:17:48 Weights data converted to conductance levels with measured error across cycles
0:18:02 Input straight DW-MTJ synapse data into CIFAR-100 Dataset Classification Test
0:18:14 Straight DW-MTJ synapse does very well on inference of challenging dataset due to high stability of each weight level
0:19:40 Linearity and Symmetry of the straight DW-MTJ corroborated by micromagnetic simulations
0:20:19 Conductance “heat maps” show high linearity at 0 K and 300 K
0:21:04 Comparison to other resistive memory technologies shows DW-MTJ with noise than ECRAM but with very good linearity
0:21:37 Standard setup for matrix vector multiplication and backpropagation weight updates
0:21:47 STT vs. SOT: Higher stochasticity of SOT boosts accuracy of notches present
0:22:45 How long of a wire/how many notches are needed for higher accuracy?
0:23:49 Materials parameters can be tuned for high accuracy
0:24:24 Some take-away points
0:25:09 Best Synapse Behaviour Depends on the Task
0:25:27 Trapezoidal DW-MTJ shows stable and controllable asymmetric weights
0:26:42 Use asymmetry in DW response across wire as the metaplastic function
0:27:34 Apply metaplastic function to stream learning task
0:28:15 Metaplastic function prevents forgetting as new data is learned and reduces number of levels needed
0:29:53 Neuron design: Tunable-SKyrmion-based Oscillating NEuron (T-SKONE)
0:31:10 5 skyrmions array neuron transfer function
0:31:28 Let’s see what it can do: application to context-aware diagnosis of breast cancer tumours
0:31:59 Conclusions
0:32:40 Q&A