Advancement in research to create new neural network “synapse” technologies were highlighted IEEE (Institute of Electrical and Electronics Engineers) meet 2018. The key focus of the meet included emphasis on reduced power, ease of production, and better performance for optimum neural network technologies. A number of methodologies were witnessed at the IEEE International Electron Device Meeting, held in San Francisco. At present, there are innumerable possibilities for improvement including flash memory, resistive RAM, electrochemical RAM (ECRAM), magneto-resistive RAM (MRAM), and phase change memory (PCM), among others.
Instead of using logic and memory banks of regular computers for neural net connections, businesses and researchers are searching for ways to represent them in various kinds of non-volatile memory setups. This allows major computations to be made without move any data. Artificial intelligence based systems use resistive RAM, MRAM, flash memory, and phase change memory.
IBM Showcased Work on ECRAM Technology
ECRAM cells appear somewhat similar to CMOS transistors. It has a dielectric layer upon which a gate is setup, to cover semi-conducting channels with electrodes, which act as the source and drain respectively. Moreover, the dielectric of the ECRAM is lithium phosphorous oxynitride, an electrolyte which is utilized in thin-film, lithium-ion batteries. For ECRAMs, the silicon channel in the CMOS transistors is constructed from tungsten trioxide, which is primarily used for smart windows and other applications.
“To set the synapse’s weight in neural networks for the right level of resistance, a current is pulsed through the gate and electrodes. If the pulse is of a single polarity, it pushes lithium ions towards the tungsten layer, improving its conductivity. If the polarity is reversed, the ions move backwards to the lithium phosphate, thereby reducing conductivity.
Analyzing the weight of the synapse needs the setting up of voltage on the electrodes, and measuring the resultant current. The diversion the currents is one of the major benefits of ECRAM, as per a spokesperson of the IBM T.J. Watson Research Center. Changing of the phase and resistant memory setups have to both develop and measure conductivity by allowing a current to run through that path. This measuring of the cell might be able to cause the drift of conductivity.
In addition, it was stated that, although there are no ideal synapses for deep learning devices and neuro-morphic chips, it is evident from the various innovations for ones introduced at the IEDM that improvements will be made from present systems.