COUPLED NETWORKS FOR PHYSICS-BASED MACHINE LEARNING
A system for physics-based learning and computation including two networks, each having a plurality of identical edges and feedback circuitry to compare the voltage drop at a given edge of the first network with the voltage drop at the corresponding edge of the second network. In both networks, at least one corresponding node is designated for input and at least one corresponding node is designated for output. In the first network, the at least one output node remains free and produces output voltage in response to the input voltage(s). In the second network, the at least one output node is clamped at voltage(s) closer to the desired value for the specified input voltage(s). Feedback circuitry compares voltages across corresponding edges and adjusts their effective resistances in order to learn.
1 - 8 . (canceled)
9 . A learning system, comprising:
at least one feedback circuit associated with (i) a first edge in a first network and (ii) a second edge in a second network, wherein the second edge corresponds to the first edge, and wherein a feedback circuit is configured to:
determine an electrical state of the first edge in the first network;
determine an electrical state of the second edge in the second network, wherein the second network receives a clamped input; and
iteratively adjust, based on the electrical state of the first edge and the electrical state of the second edge, an electrical characteristic of the first edge and the second edge, so as to cause an output associated with the first edge to approach the clamped input.
10 . The learning system of claim 9 , wherein the electrical state of the first edge is associated with at least one of: a first voltage input, a first current input, and a first physical input, and wherein the electrical state of the second edge is associated with least one of: a second voltage input, a second current input, and a second physical input.
11 . The learning system of claim 9 , wherein the first network comprises a first plurality of nodes, and wherein the second network comprises a second plurality of nodes corresponding to the first plurality of nodes.
12 . The learning system of claim 9 , wherein the first network and the second network are structurally identical.
13 . The learning system of claim 9 , wherein the clamped input comprises a voltage determined from training data associated with a task.
14 . The learning system of claim 9 , wherein iteratively adjusting the electrical characteristic solves at least one of a classification task, an allosteric task, a regression task, and a sequential task.
15 . The learning system of claim 9 , wherein the at least one feedback circuit comprises at least one of: a capacitor and a transistor.
16 . The learning system of claim 9 , wherein the at least one feedback circuit comprises a free transistor and a clamped transistor.
17 . The learning system of claim 9 , wherein an effective resistance, ΔR i , of one or more edges iteratively adjusts according to:
Δ
R
i
=
{
+
δ
R
if
❘
"\[LeftBracketingBar]"
Δ
V
i
C
❘
"\[RightBracketingBar]"
>
❘
"\[LeftBracketingBar]"
Δ
V
i
F
❘
"\[RightBracketingBar]"
,
-
δ
R
otherwise
.
}
,
or
Δ
R
i
=
{
+
δ
R
if
XOP
[
Δ
V
i
C
>
Δ
V
i
F
,
0
<
Δ
V
i
C
]
-
δ
R
otherwise
.
}
.
wherein
Δ
V
i
F
corresponds to the electrical state of the first network and
Δ
V
i
C
corresponds to the electrical state of the second network.
18 . The learning system of claim 17 , wherein
V
→
C
O
corresponds to the clamped input, and wherein
V
→
C
O
is changed according to:
V
→
C
O
=
η
V
→
D
+
(
1
-
η
)
V
→
F
O
wherein {right arrow over (V)} D is a desired output,
V
→
F
O
is the first input, and 0<η≤1.
19 . The learning system of claim 9 , wherein the first network and the second network have a same number of edges.
20 . The learning system of claim 19 , wherein the feedback circuit comprises a transistor, and wherein the electrical characteristic varies with a voltage applied to the transistor.
21 . The learning system of claim 9 , wherein the first network and the second network are linear networks or non-linear networks.
22 . The learning system of claim 9 , wherein the electrical characteristic of the first edge or the second edge adjusts according to a difference in the electrical state of the first network and the electrical state of the second network.
23 . A learning method, comprising:
determining an electrical state of a first edge in a first network;
determining an electrical state of a second edge in a second network, wherein the second edge corresponds to the first edge, and wherein the second network receives a clamped input; and
iteratively adjusting, based on the electrical state of the first edge and the electrical state of the second edge, an electrical characteristic of the first edge and the second edge, so as to cause an output associated with the first edge to approach the clamped input.
24 . The learning method of claim 23 , wherein the electrical state of the first edge is associated with at least one of: a first voltage input, a first current input, and a first physical input, and wherein the second electrical state is associated with least one of: a second voltage input, a second current input, and a second physical input.
25 . The learning method of claim 23 , wherein the first network comprises a first plurality of nodes, and wherein the second network comprises a second plurality of nodes corresponding to the first plurality of nodes.
26 . The learning method of claim 23 , wherein the clamped input comprises at least one voltage determined from training data associated with a task.
27 . A non-transitory computer readable storage medium comprising instructions stored thereon, which when executed by a processor cause a computing device to:
determine a first difference between a first edge of a first network and a second edge of a second network, wherein the second edge corresponds to the first edge; and
iteratively adjust, based on the first difference, an electrical characteristic of at least one of the first edge of the first network and the second edge of the second network so as to cause a first output of the first network to approach a clamped input provided at the second edge.