Syncronised Oscillatory Networks: Applications to Graph Colouring

Satyam Mohla, Krishna Subramani, S. Lashkare, U. Ganguly

Abstract •The network left to evolve, during which •Simulations were done on a
the phases of the oscillators in each cluster benchmark graph, myciel3.col,
• Digital Computation is limited when dealing with certain classes of are almost synchronized, whereas the from DIMACS for graph colouring.
problems such as associative computing and optimising. phases of oscillators in different clusters •The simulation result are shown
• Graph colouring is a NP-Hard combinatorial optimization problem, with are far away. here, which split all the nodes into
real world applications in fault diagnosis, scheduling, resource allocation. 4 clusters: S1 = {1}, S2 = {3, 4,
• We simulate an Oscillatory Neural Network to approximate the solution of 11}, S3 = {2, 5}, and S4 = {6, 7, 8,
the graph colouring problem. •Thus the vertices are spit into several sub-sets and 9, 10}.
• We also demonstrate in theory the performance of such a device using all the vertices in a cluster can be assigned the •Examination shows
simulations and hardware data. same colour. We thus observe that the relative that it is indeed a
ordering of the phases approximates minimum
Introduction vertex-colour-sorting of the graph.
proper 4-coloring of
• The Graph Colouring Problem: the graph.
•For the graph shown aside, vertices are split into
• Colour the vertices of a graph with the minimum two sub-sets: S1 = {v2, v4, v6} and S2 = {v1, v3, v5},
number of colours, such that no two adjacent •To test the robustness of the
vertices have the same colour. which is a proper 2-coloring of the graph. model we added noise to the
• We use clustering phenomenon in a coupled system and observed de-
oscillator network to solve the problem dynamically. Proposed Device synchronization in the network
•Complement of Graph G¯(V, E¯) : which lead to less clustering and
•Experimentally obtained thus increased number of colours
•A graph with same vertices but no edges
waveform data from the device being detected
where a edge exists in G(V, E)
was used to analyse the
•
The Modified Kuramoto Synchronization Model: hardware performance of the Conclusions
proposed network.
• In this work, we have successfully established the capability of the device to act
• The coupled phase oscillators, (Kuramoto
model) is described by the equation •FFT was used to identify the dominant as a viable oscillator for dynamical computing networks.
harmonic components and to construct an • We also simulated noise and commented on the robustness of the device.
• +ve Kp is to make the phases of two equivalent waveform for simulation. • We envision such dynamical systems to provide foundational paradigms in the
connected vertices evolve together development of next-generation computational accelerators & kernels.
•Kuramoto’s Model was modified to
• Whereas –ve Kn is to make the phases of
two unconnected vertices evolve far away account for the additional harmonics, Further Work
introduced by the device. • To simulate our network on other experimental waveforms and benchmark the device
Methods Results • A spintronic oscillator can also be used in future hardware implementations for recognition;
however the dynamics involve solving LLG and NEGF (scf) & Kuramoto model fails there.
•By associating vertices to the •Our work highlights a physics-based (*A recognition ONN for MNIST classification using a spin oscillator and Kuramoto Model was also attempted as part of EE751 course)
phase oscillators and edges to the computing approach while also providing References
couplings, a graph can be direction for building customized analog • Wu, Jianshe, et al. "Clustering dynamics of nonlinear oscillator network: Application to graph
considered as a network of phase hardware for solving such problems coloring problem." Physica D: Nonlinear Phenomena 240.24 (2011): 1972-1978.
oscillators. efficiently. • Parihar, Abhinav, et al. "Vertex coloring of graphs via phase dynamics of coupled oscillatory
•We simulated the device both in presence networks." Scientific Reports 7 (2017).
•To identify the complete sub-graph/cluster in the complement, the focus
• Parihar, Abhinav, et al. "Computing with dynamical systems in the post-CMOS era." Photonics
here is on the clustering phenomenon of the modified Kuramoto model. and absence of noise. Society Summer Topical Meeting Series (SUM), 2016 IEEE. IEEE, 2016 APA
www.postersession.com