Methods and systems for digital neural processing with discrete-level synapes and probabilistic STDP
Certain embodiments of the present disclosure support implementation of a digital neural processor with discrete-level synapses and probabilistic synapse weight training.
1. An electrical circuit, comprising:
a digital neural processing unit with one or more synapses and a post-synaptic neuron circuit connected to the synapses,
wherein a weight of one of the synapses changes a value in discrete levels with a probability depending on a time elapsed between a pair of spikes originated from the post-synaptic neuron circuit and a pre-synaptic neuron circuit connected to the synapse.
2. The electrical circuit of claim 1 , wherein the probability exponentially decays as a function of the elapsed time.
3. The electrical circuit of claim 1 , wherein the post-synaptic neuron circuit further comprises:
at least one generator circuit triggered by one of the spikes to generate at least one pulse signal with a probability of each pulse to occur in a defined time interval exponentially decaying over time.
4. The electrical circuit of claim 3 , wherein one of the pulses causes the weight value to decrease.
5. The electrical circuit of claim 3 , wherein one of the pulses causes another weight value of another synapse connected to the neural processing unit to increase by a defined discrete level.
6. The electrical circuit of claim 3 , wherein each of the generator circuits comprises ring oscillators.
7. The electrical circuit of claim 1 , wherein the synapse further comprises:
a generator circuit triggered at least in part by one of the spikes to generate a pulse signal with a probability of each pulse to occur in a defined time interval exponentially decaying over time.
8. The electrical circuit of claim 7 , wherein one of the pulses and a gate signal cause the weight value to increase.
9. The electrical circuit of claim 7 , wherein one of the pulses and a gate signal cause the weight value to decrease.
10. The electrical circuit of claim 7 , wherein the generator circuit comprises ring oscillators.
11. The electrical circuit of claim 1 , wherein the synapse further comprises:
a digital-to-analog conversion circuit gated by one of the spikes generating an electrical current to be input into the post-synaptic neuron circuit, a value of the electrical current is based on the weight value of the synapse.
12. A method for implementing a digital neural processing unit, comprising:
connecting a post-synaptic neuron circuit of the neural processing unit with at least one synapse of the neural processing unit; and
changing, in discrete levels, a weight value of one of the synapses with a probability depending on a time elapsed between a pair of spikes originated from the post-synaptic neuron circuit and a pre-synaptic neuron circuit connected to the synapse.
13. The method of claim 12 , wherein the probability exponentially decays as a function of the elapsed time.
14. The method of claim 12 , further comprising:
triggering, by one of the spikes, at least one generator circuit within the post-synaptic neuron circuit to start generating at least one pulse signal with a probability of each pulse to occur in a defined time interval exponentially decaying over time.
15. The method of claim 14 , wherein one of the pulses causes the weight value to decrease.
16. The method of claim 14 , wherein one of the pulses causes another weight value of another synapse connected to the neural processing unit to increase by a defined discrete level.
17. The method of claim 14 , wherein each of the generator circuits comprises ring oscillators.
18. The method of claim 12 , further comprising:
triggering, at least in part by one of the spikes, a generator circuit within the synapse to start generating a pulse signal with a probability of each pulse to occur in a defined time interval exponentially decaying over time.
19. The method of claim 18 , wherein one of the pulses and a gate signal cause the weight value to increase.
20. The method of claim 18 , wherein one of the pulses and a gate signal cause the weight value to decrease.
21. The method of claim 18 , wherein the generator circuit comprises ring oscillators.
22. The method of claim 12 , further comprising:
generating, by a digital-to-analog conversion circuit within the synapse gated by one of the spikes, an electrical current to be input into the post-synaptic neuron circuit, wherein a value of the electrical current is based on the weight value of the synapse.
23. An apparatus for implementing a digital neural processing unit, comprising:
means for connecting a post-synaptic neuron circuit of the neural processing unit with at least one synapse of the neural processing unit; and
means for changing, in discrete levels, a weight value of one of the synapses with a probability depending on a time elapsed between a pair of spikes originated from the post-synaptic neuron circuit and a pre-synaptic neuron circuit connected to the synapse.
24. The apparatus of claim 23 , wherein the probability exponentially decays as a function of the elapsed time.
25. The apparatus of claim 23 , further comprising:
means for triggering, by one of the spikes, at least one generator circuit within the post-synaptic neuron circuit to start generating at least one pulse signal with a probability of each pulse to occur in a defined time interval exponentially decaying over time.
26. The apparatus of claim 25 , wherein one of the pulses causes the weight value to decrease.
27. The apparatus of claim 25 , wherein one of the pulses causes another weight value of another synapse connected to the neural processing unit to increase by a defined discrete level.
28. The apparatus of claim 25 , wherein each of the generator circuits comprises ring oscillators.
29. The apparatus of claim 23 , further comprising:
means for triggering, at least in part by one of the spikes, a generator circuit within the synapse to start generating a pulse signal with a probability of each pulse to occur in a defined time interval exponentially decaying over time.
30. The apparatus of claim 29 , wherein one of the pulses and a gate signal cause the weight value to increase.
31. The apparatus of claim 29 , wherein one of the pulses and a gate signal cause the weight value to decrease.
32. The apparatus of claim 29 , wherein the generator circuit comprises ring oscillators.
33. The apparatus of claim 23 , further comprising:
means for generating, by a digital-to-analog conversion circuit within the synapse gated by one of the spikes, an electrical current to be input into the post-synaptic neuron circuit, wherein a value of the electrical current is based on the weight value of the synapse.
34. A computer program product in a digital neural processing unit, comprising a computer-readable medium comprising code for:
interfacing a post-synaptic neuron circuit of the neural processing unit with at least one synapse of the neural processing unit; and
changing, in discrete levels, a weight value of one of the synapses with a probability depending on a time elapsed between a pair of spikes originated from the post-synaptic neuron circuit and a pre-synaptic neuron circuit connected to the synapse.
35. The computer program product of claim 34 , wherein the probability exponentially decays as a function of the elapsed time.
36. The computer program product of claim 34 , wherein the computer-readable medium further comprises code for:
triggering, by one of the spikes, at least one generator circuit within the post-synaptic neuron circuit to start generating at least one pulse signal with a probability of each pulse to occur in a defined time interval exponentially decaying over time.
37. The computer program product of claim 36 , wherein one of the pulses causes the weight value to decrease.
38. The computer program product of claim 36 , wherein one of the pulses causes another weight value of another synapse connected to the neural processing unit to increase by a defined discrete level.
39. The computer program product of claim 36 , wherein each of the generator circuits comprises ring oscillators.
40. The computer program product of claim 34 , wherein the computer-readable medium further comprises code for:
triggering, at least in part by one of the spikes, a generator circuit within the synapse to start generating a pulse signal with a probability of each pulse to occur in a defined time interval exponentially decaying over time.
41. The computer program product of claim 40 , wherein one of the pulses and a gate signal cause the weight value to increase.
42. The computer program product of claim 40 , wherein one of the pulses and a gate signal cause the weight value to decrease.
43. The computer program product of claim 40 , wherein the generator circuit comprises ring oscillators.
44. The computer program product of claim 34 , wherein the computer-readable medium further comprises code for:
generating, by a digital-to-analog conversion circuit within the synapse gated by one of the spikes, an electrical current to be input into the post-synaptic neuron circuit, wherein a value of the electrical current is based on the weight value of the synapse.