Trace-based neuromorphic architecture for advanced learning
A neuromorphic computing apparatus has a network of neuromorphic cores, with each core including an input axon and a plurality of neurons having synapses. The input axon is associated with an input data store to store an input trace representing a time series of filtered pre-synaptic spike events, and accessible by the synapses of the plurality of neurons of the core. Each neuron includes at least one dendritic compartment to store and process variables representing a dynamic state of the neuron. Each compartment is associated with a compartment-specific data store to store an output trace representing a time series of filtered post-synaptic spike events. Each neuron includes a learning engine to apply a set of one or more learning rules based on the pre-synaptic and post-synaptic spike events to produce an adjustment of parameters of a corresponding synapse to those spike events.
1. A neuromorphic computing apparatus comprising:
computing hardware, including processing and data storage circuitry;
a network of spiking neural network cores implemented in the computing hardware, each core including an input axon and a plurality of neurons, including synapses, wherein the input axon is common to the plurality of neurons;
wherein the input axon is associated with an input data store to store an input trace representing a time series of pre-synaptic spike events that are filtered according to a filtering function, and accessible by the synapses of the plurality of neurons of the core;
wherein each one of the plurality of neurons includes at least one dendritic compartment to store and process variables representing a dynamic state of the neuron, and wherein the at least one dendritic compartment is associated with a compartment-specific data store to store an output trace representing a time series of post-synaptic spike events that are filtered according to a filtering function; and
wherein each one of the plurality of neurons includes a learning engine to:
apply a set of one or more learning rules based on the pre-synaptic and post-synaptic spike events to produce an adjustment of parameters of a corresponding synapse to those spike events; and
compare the pre-synaptic or post-synaptic state variables against a learning exit condition, wherein upon satisfaction of the learning exit condition, the learning engine is to terminate a learning operation.
2. The apparatus of claim 1 , wherein the input data store is to store a plurality of input traces.
3. The apparatus of claim 1 , wherein the compartment-specific data store associated with each of the at least one compartment is to store a plurality of output traces.
4. The apparatus of claim 1 , wherein the input axon is a dedicated reward axon responsive to reinforcement-learning signaling, and wherein the learning engine is to read, process, and adjust, the reinforcement-learning signaling according to applicable learning rules.
5. The apparatus of claim 1 , wherein the input axon is associated with a reinforcement-learning data store that stores at least one reinforcement-learning trace representing a temporal sequence of reinforcement-learning state variables.
6. The apparatus of claim 1 , wherein the adjustment of the parameters of the corresponding synapse includes at least one adjustment selected from among: weighting adjustment, delay adjustment, tag adjustment, or any combination thereof.
7. The apparatus of claim 1 , wherein the learning engine is to access a forward-mapping data structure that associates fan-outs of the input axon with dendritic compartments that are responsive to stimuli arriving on the input axon.
8. The apparatus of claim 1 , wherein the learning engine is to access a backward-mapping data structure that associates each dendritic compartment to a corresponding set of fan-in input axons.
9. The apparatus of claim 1 , wherein the learning engine is to perform a learning cycle in response to a predefined passage of time steps, wherein the predefined passage of time steps represents a learning epoch.
10. The apparatus of claim 9 , wherein the input data store maintains historical values of the input trace, including time offsets of the pre-synaptic spike events, over a set of prior epochs, to facilitate historic reconstruction of the input trace.
11. The apparatus of claim 10 , wherein the historical values of the input trace are used to compute trace values for a current epoch at fanout synapses with nonzero network delay.
12. The apparatus of claim 9 , wherein the input data store update stored traces for decay only once per learning epoch.
13. The apparatus of claim 9 , wherein the learning engine applies learning rules unconditionally in response to passage of a predefined number of learning epochs.
14. The apparatus of claim 1 , wherein the learning engine is programmable to access a plurality of different learning rule profiles.
15. The apparatus of claim 1 , wherein the learning engine is programmable to apply a plurality of learning rules.
16. The apparatus of claim 1 , wherein the learning engine is to apply learning rules that are expressed as a sum-of-products semantic, with each product comprising a programmable series of trace values, synaptic parameters, and constants available to the learning engine.
17. The apparatus of claim 1 , wherein each neuron includes:
an input to accept pre-synaptic signaling from an input axon associated with the neuromorphic core;
a synapse to store and distribute input and feedback signaling for processing, the synapse including a set of pre-synaptic terminals; and
an output to carry post-synaptic signaling from the neuron to other neurons.
18. The apparatus of claim 1 , wherein the learning engine includes a spike timing-dependent plasticity (STDP) engine.
19. A method for operating a neuromorphic computing system, the method being executed by computing hardware, and comprising:
configuring a network of spiking neural network cores in the computing hardware, each core including an input axon and a plurality of neurons, including synapses, wherein the input axon is common to the plurality of neurons;
associating the input axon with an input, data store to store an input trace representing a time series of pre-synaptic spike events that are filtered according to a filtering function, and accessible by the synapses of the plurality of neurons of the core;
providing each one of the plurality of neurons with at least one dendritic compartment to store and process variables representing a dynamic state of the neuron;
associating the at least one dendritic compartment with a compartment-specific data store to store an output trace representing a time series of post-synaptic spike events that are filtered according to a filtering function; and
for each one of the plurality of neurons;
applying a set of one or more learning rules based on the pre-synaptic and post-synaptic spike events to produce an adjustment of parameters of a corresponding synapse to those spike events; and
comparing the pre-synaptic or post-synaptic state variables against a learning exit condition, wherein upon satisfaction of the learning exit condition, the learning engine is to terminate a learning operation.
20. The method of claim 19 , wherein the input axon is a dedicated reward axon responsive to reinforcement-learning signaling, and further comprising:
reading, processing, and adjusting, the reinforcement-learning signaling according to applicable learning rules.
21. The method of claim 19 , wherein the input axon is associated with a reinforcement-learning data store that stores at least one reinforcement-learning trace representing a temporal sequence of reinforcement-learning state variables.
22. At least one non-transitory machine-readable medium containing instructions that, when executed by computing hardware, cause the computing hardware to perform:
configuring a network of spiking neural network cores in the computing hardware, each core including an input axon and a plurality of neurons, including synapses, wherein the input axon is common to the plurality of neurons;
associating the input axon with an input data store to store an input trace representing a time series of pre-synaptic spike events that are filtered according to a filtering function, and accessible by the synapses of the plurality of neurons of the core;
providing each one of the plurality of neurons with at least one dendritic compartment to store and process variables representing a dynamic state of the neuron;
associating the at least one dendritic compartment with a compartment-specific data store to store an output trace representing a tune series of post-synaptic spike events that are filtered according to a filtering function; and
for each one of the plurality of neurons;
applying a set of one or more learning rules based on the pre-synaptic and post-synaptic spike events to produce an adjustment of parameters of a corresponding synapse to those spike events; and
comparing the pre-synaptic or post-synaptic state variables against a learning exit condition, wherein upon satisfaction of the learning exit condition, the learning engine is to terminate a learning operation.
23. The at least one non-transitory machine-readable medium of claim 22 , further comprising instructions for performing a learning cycle in response to a predefined passage of time steps, wherein the predefined passage of time steps represents a learning epoch.
24. The at least one non-transitory machine-readable medium of claim 23 , wherein the input data store maintains historical values of the input trace, including time offsets of the pre-synaptic spike events, over a set of prior epochs, to facilitate historic reconstruction of the input trace.
25. The at least one non-transitory machine-readable medium of claim 23 , further comprising instructions for applying learning rules unconditionally in response to passage of a predefined number of learning epochs.