Intercalation cells for multi-task learning
An electro-chemical random-access memory (ECRAM) cell includes a substrate and a plurality of source-drain pairs positioned on a top surface of the substrate, each source-drain pair comprising a source and a drain. A channel at least partially overlays the substrate and the plurality of source-drain pairs, and a transfer layer at least partially overlays the channel. A gate at least partially overlays the transfer layer, the gate at least partially controlling a channel between each source-drain pair.
1. An electro-chemical random-access memory (ECRAM) cell comprising:
a substrate having a top surface;
a plurality of source-drain pairs positioned on the top surface of the substrate, each of the source-drain pairs comprising a source and a drain;
a channel at least partially overlaying the substrate and the plurality of source-drain pairs;
a transfer layer at least partially overlaying the channel; and
a gate at least partially overlaying the transfer layer, the gate at least partially controlling a channel between each source-drain pair.
2. The ECRAM cell of claim 1 , wherein each source-drain pair is configured to enable sensing of data stored in the ECRAM cell via a different electrical path.
3. The ECRAM cell of claim 1 , wherein all source-drain pairs have substantially identical separation lengths within a gate pattern.
4. The ECRAM cell of claim 1 , wherein data stored in the ECRAM cell is a weight of a neural network.
5. The ECRAM cell of claim 1 , wherein the gate is a square gate and the source-drain pairs have a parallel layout.
6. The ECRAM cell of claim 1 , wherein the gate is a square gate and each source-drain pair is parallel to at least one other source-drain pair and perpendicular to at least one other source-drain pair.
7. The ECRAM cell of claim 1 , wherein the gate is a circular gate and each source-drain pair is placed at a polygonal corner of a polygon within the circular gate.
8. The ECRAM cell of claim 1 , wherein timing is delayed from programming to mitigate a transient effect.
9. The ECRAM cell of claim 8 , wherein a delay time is varied for each source-drain pair to avoid interference with other paths through the source-drain pairs.
10. A multi-task learning system comprising:
an input layer;
a pre-trained shared layer coupled to the input layer; and
one or more processing layers coupled to the pre-trained shared layer;
wherein the pre-trained shared layer comprises an array of ECRAMs, each ECRAM comprising;
a gate; and
a plurality of source-drain pairs, each source drain-pair comprising a source and a drain and providing a connection to a task of one of the one or more processing layers, the gate at least partially controlling a channel between each source-drain pair.
11. The multi-task learning system of claim 10 , wherein the input layer comprises a second ECRAM array.
12. The multi-task learning system of claim 10 , wherein at least one of the processing layers comprises an ECRAM array.
13. The multi-task learning system of claim 10 , wherein the ECRAM array of the pre-trained shared layer is used only for a forward pass from the input layer to at least one of the processing layers.
14. The multi-task learning system of claim 10 , the multi-task learning system being further configured to apply no bias during error back-propagation to disable the ECRAM.
15. The multi-task learning system of claim 10 , wherein at least one of the one or more processing layers incorporates a non-volatile memory (NVM) crossbar array for online training and support of a plurality of tasks.
16. The multi-task learning system of claim 10 , wherein at least one of the one or more processing layers serves as an output layer.
17. The multi-task learning system of claim 10 , wherein pre-trained weights of the pre-trained shared layer are used for detection of edge features and higher-level features are extracted via individual tasks using the one or more processing layers, and wherein the individual tasks are trained online with back-propagation.
18. The multi-task learning system of claim 10 , each ECRAM further comprising:
a substrate, the plurality of source-drain pairs are positioned on a top surface of the substrate;
a channel at least partially overlaying the substrate and the plurality of source-drain pairs; and
a transfer layer at least partially overlaying the channel, the gate at least partially overlaying the transfer layer.
19. The multi-task learning system of claim 10 , wherein each source-drain pair is configured to enable sensing of data stored in the ECRAM cell via a different electrical path and wherein all source-drain pairs have substantially identical separation lengths within a gate pattern.
20. The multi-task learning system of claim 10 , wherein timing is delayed from programming to mitigate a transient effect and wherein a delay time is varied for each source-drain pair to avoid interference with other paths through the source-drain pairs.