Oscillator based neural network apparatus
A neural network scheme is described that uses unsupervised learning in oscillator neural networks. Training occurs by varying the weights in proportion to the output from a frequency detector. Inputs and initial weights are split into plurality of inputs and plurality of weights. These split inputs and weights can be analog or digital. Oscillators generate signals having frequencies that represent difference in inputs, initial weights, and adjusted factors. Frequency detectors are used to compare the oscillator frequencies with a synchronized frequency of all oscillators. The output of the frequency detectors are used to generate the adjusted factors, and in turn generate trained weights.
1. An apparatus comprising:
a plurality of adders or subtractors, each of which is to receive an input voltage, a corresponding initial weight voltage, and a corresponding adjustment voltage;
a plurality of buffers, wherein each buffer is coupled to an adder or subtractor of the plurality of adders or subtractors;
a plurality of oscillators, wherein each oscillator is coupled to an output of an individual buffer of the plurality of buffers;
an averager coupled to the plurality of oscillators, wherein the average is to generate a signal with a frequency being an average of frequencies of outputs of the plurality of oscillators; and
a plurality of frequency detectors, wherein each frequency detector is coupled to an individual oscillator and the averager, and also coupled to an individual adders or subtractor of the plurality of adders or subtractor, wherein an output of each adder or subtractor corresponds to a sum of the input voltage, the corresponding initial weight voltage, and the corresponding adjustment voltage.
2. The apparatus of claim 1 , wherein each oscillator is a current controlled or voltage controlled oscillator.
3. The apparatus of claim 1 , wherein each buffer is a voltage buffer comprising an amplifier.
4. The apparatus of claim 1 comprising a peak detector coupled to an output of the averager.
5. The apparatus of claim 1 , wherein each frequency detector generates a signal indicative of a difference between the average of frequencies and the frequency of the signal from the individual oscillator.
6. The apparatus of claim 1 , wherein each adder or subtractor is to subtract the corresponding initial weight voltage and the corresponding adjustment voltage from the input voltage.
7. The apparatus of claim 1 , wherein each frequency detector is coupled to a charge pump which is coupled to an individual adder or subtractor.
8. The apparatus of claim 1 , wherein each adder or subtractor is an analog adder or subtractor.
9. The apparatus of claim 1 , wherein the corresponding adjustment voltage, once stabilized at or near a constant value, is a trained weight.
10. The apparatus of claim 9 comprising a memory to store the trained weight.
11. An apparatus comprising:
a plurality of digital adders or subtractors, each of which is to receive an input code, a corresponding initial weight code, and a corresponding adjustment code;
a plurality of digital-to-analog converters (DACs), wherein each DAC is coupled to an individual digital adder or subtractor of the plurality of digital adders or subtractors;
a plurality of oscillators, wherein each oscillator is coupled to an output of an individual DAC of the DACs;
an averager coupled to the plurality of oscillators, wherein the average is to generate a signal with a frequency being an average of frequencies of outputs of the plurality of oscillators; and
a plurality of frequency detectors, wherein each frequency detector is coupled to an individual oscillator and the average; and
a plurality of analog-to-digital converters (ADCs), wherein each ADC is coupled to an individual digital adder or subtractor, wherein an output of each digital adder or subtractor corresponds to a sum of the input code, the corresponding initial weight code, and the corresponding adjustment code.
12. The apparatus of claim 11 , wherein each frequency detector is coupled to a charge pump which is coupled to an individual ADC.
13. The apparatus of claim 11 , wherein each oscillator is a current controlled or voltage controlled oscillator.
14. The apparatus of claim 11 comprising a peak detector coupled to an output of the averager.
15. The apparatus of claim 11 , wherein the corresponding adjustment code, once stabilized at or near a constant value, is a trained weight.
16. The apparatus of claim 15 comprising memory to store the trained weight.
17. A system comprising:
a memory;
a processor with a neural network circuitry which includes:
a plurality of adders or subtractors, each of which is to receive an input voltage, a corresponding initial weight voltage, and a corresponding adjustment voltage;
a plurality of buffers, wherein each buffer is coupled to an adder or subtractor of the plurality of adders or subtractors;
a plurality of oscillators, wherein each oscillator is coupled to an output of an individual buffer of the plurality of buffers;
an averager coupled to the plurality of oscillators, wherein the average is to generate a signal with a frequency being an average of frequencies of outputs of the plurality of oscillators; and
a plurality of frequency detectors, wherein each frequency detector is coupled to an individual oscillator and the averager, and also coupled to an individual adders or subtractor of the plurality of adders or subtractor, wherein an output of each adder or subtractor corresponds to a sum of the input voltage, the corresponding initial weight voltage, and the corresponding adjustment voltage; and
a wireless interface to allow the processor to communicate with another device.
18. The system of claim 17 , wherein the neural network circuitry comprises a peak detector coupled to an output of the averager.
19. The apparatus of claim 17 , wherein the corresponding adjustment voltage, once stabilized at or near a constant value, is a trained weight.
20. The system of claim 19 , wherein the memory is to store the trained weight.