Neural architecture search for fusing multiple networks into one
One or more embodiments of the present disclosure include systems and methods that use neural architecture fusion to learn how to combine multiple separate pre-trained networks by fusing their architectures into a single network for better computational efficiency and higher accuracy. For example, a computer implemented method of the disclosure includes obtaining multiple trained networks. Each of the trained networks may be associated with a respective task and has a respective architecture. The method further includes generating a directed acyclic graph that represents at least a partial union of the architectures of the trained networks. The method additionally includes defining a joint objective for the directed acyclic graph that combines a performance term and a distillation term. The method also includes optimizing the joint objective over the directed acyclic graph.
1. A computer implemented method, comprising:
obtaining multiple trained networks, wherein each of the trained networks is associated with a respective task and has a respective architecture;
generating a directed acyclic graph that represents at least a partial union of the architectures of the trained networks;
defining a joint objective for the directed acyclic graph that combines a performance term and a distillation term; and
optimizing the joint objective over the directed acyclic graph.
2. The computer implemented method of claim 1 , further comprising using weights of the trained networks to initialize nodes of the directed acyclic graph.
3. The computer implemented method of claim 1 , wherein the performance term is selected from the group consisting of latency, accuracy, and a number of parameters.
4. The computer implemented method of claim 1 , wherein defining the joint objective comprises regressing outputs of the trained networks.
5. The computer implemented method of claim 1 , wherein optimizing the joint objective comprises using a differentiable architecture search.
6. The computer implemented method of claim 1 , wherein the directed acyclic graph comprises a lattice.
7. A non-transitory computer-readable medium having executable instructions stored thereon that, when executed by a processor, perform the operations of:
generating a directed acyclic graph that represents at least a partial union of multiple trained networks, wherein each of the trained networks is associated with a respective task and has a respective architecture;
defining a joint objective for the directed acyclic graph, wherein the joint objective represents a combination of objectives for each of the trained networks; and
optimizing the joint objective over the directed acyclic graph.
8. The non-transitory computer-readable medium of claim 7 , wherein the instructions, when executed by the processor, further perform the function of using weights of the trained networks to initialize nodes of the directed acyclic graph.
9. The non-transitory computer-readable medium of claim 7 , wherein optimizing the joint objective comprises using a differentiable architecture search.
10. The non-transitory computer-readable medium of claim 7 , wherein the directed acyclic graph comprises a search space lattice comprising nodes connected by edges.
11. The non-transitory computer-readable medium of claim 10 , wherein the instructions, when executed by the processor, further perform the function of selecting paths in the search space lattice using edge weights of the search space lattice.
12. The non-transitory computer-readable medium of claim 11 , wherein selecting the paths in the search space lattice comprises using a Viterbi algorithm.
13. A non-transitory computer-readable medium having executable instructions stored thereon that, when executed by a processor, perform the operations of:
decoding one or more architectures by selecting paths in a directed acyclic graph network comprising nodes and edges using edge weights of the directed acyclic graph,
wherein the directed acyclic graph represents at least a partial union of multiple trained networks, wherein each of the trained networks is associated with a respective task and has a respective architecture,
wherein a joint objective has been optimized over the directed acyclic graph, wherein the joint objective is defined for the directed acyclic graph and represents a combination of objectives for each of the trained networks.
14. The non-transitory computer-readable medium of claim 13 , wherein the directed acyclic graph network is a lattice search space.
15. The non-transitory computer-readable medium of claim 13 , wherein the joint objective has been optimized using a differentiable architecture search.
16. The non-transitory computer-readable medium of claim 13 , wherein the joint objective is defined using regression.