In the pursuit for building hardware accelerators to compute optimization problems researchers realize that the challenges in achieving this objective lie not only in implementing the hardware but also in the formulating the computing fundamentals of such designs. Neural network algorithms are considered most suited for this task, as there is usually a direct description of distributed computing entities, called “neurons”, and their interactions which can be mapped to both electronic and non-electronic hardware. In this regard, coupled oscillator systems have been studied where individual oscillators correspond to neurons and the information is encoded in either phase or frequency. But as is the case with neural networks, the computational power of the network depends on complexity of interactions among oscillators, and it is a challenge to implement oscillator networks with complex simultaneous interactions among multiple oscillators. Sinusoidal oscillators with assumption of weak linear phase coupling, akin to Kumamoto models, have been studied in theory but implementing such oscillators with weak couplings and encoding information in phase or frequency have been a challenge. Examples of using novel devices for making neural network hardware include memristor based neuromorphic synapses [1] and spin-torque oscillator (STO) based systems [2]. In our work, we use relaxation oscillators coupled using passive elements - capacitances or resistances - without the assumption of weak linear phase couplings. Our theoretical models are derived from circuit implementations, instead of the other way round, which means there are only engineering challenges in implementing the hardware, and no modeling discrepancies. We have explored two kinds of implementations - (a) simple pairwise coupling scheme with information encoded as frequency for pattern matching and associative computing, and (b) complex global coupling with information encoded in phase for the NP-hard graph coloring problem. We have been demonstrated in theory, using simulations and experimental implementations using VO2 devices, the working of such coupled relaxation oscillator networks.
[1]
Vijaykrishnan Narayanan,et al.
Synchronized charge oscillations in correlated electron systems
,
2014,
Scientific Reports.
[2]
Suman Datta,et al.
Neuro inspired computing with coupled relaxation oscillators
,
2014,
2014 51st ACM/EDAC/IEEE Design Automation Conference (DAC).
[3]
S. Datta,et al.
Pairwise coupled hybrid vanadium dioxide-MOSFET (HVFET) oscillators for non-boolean associative computing
,
2014,
2014 IEEE International Electron Devices Meeting.
[4]
Wei Yang Lu,et al.
Nanoscale memristor device as synapse in neuromorphic systems.
,
2010,
Nano letters.
[5]
Suman Datta,et al.
Exploiting Synchronization Properties of Correlated Electron Devices in a Non-Boolean Computing Fabric for Template Matching
,
2014,
IEEE Journal on Emerging and Selected Topics in Circuits and Systems.
[6]
Kaushik Roy,et al.
Coupled Spin Torque Nano Oscillators for Low Power Neural Computation
,
2015,
IEEE Transactions on Magnetics.
[7]
Suman Datta,et al.
Synchronization of pairwise-coupled, identical, relaxation oscillators based on metal- insulator phase transition devices: A model study
,
2014
.