Akida ballista's learning capacity needs explaining better IMO ?, page-7

Hi Hotty,

A few days ago, I posted about:

WO2020092691A1 AN IMPROVED SPIKING NEURAL NETWORKAvailable inPatent TranslateBibliographic data
https://worldwide.espacenet.com/patent/search/family/070458523/publication/WO2020092691A1?q=WO2020092691

This patent application discusses the mechanisms for learning using STDP:

[0121] FIG. 7 is an example of a Spike Time Dependent Plasticity (STDP) learning method, according to some embodiments. In STDP learning, spikes that contribute to an output event/spike can have their representative synaptic weights strengthened while spikes that did not contribute to an output event/spike can have their synaptic weights weakened.

The strengthening or weakening is carried out by the UP/DOWN counter 109 in Figure 1.

[0127] In some embodiments, a spiking neuron circuit emits a spike when its inputs drive its membrane potential value (e.g., counter 109) up to a threshold value. This can mean that when the neuron is driven to the threshold value and generates a spike, connections from its recently activated inputs are strengthened, while a number of its other connections are weakened. The can result in neurons learning to respond to patterns of inputs that they see repeatedly, thereby autonomously learning the features that characterize an input dataset.

The strengthening/weakening depends on the relative time of arrival of the spike (leading/lagging).

[0062]... In STDP learning, an input spike that precedes an output spike indicates the input spike contributed to the output spike. In STDP, this can causes the synapse weight to be strengthened.

Learning can be unsupervised:

[0130] In some embodiments, this modified STDP learning method is completely unsupervised. This is different than conventional diverse supervised training methods that are in use in neural networks. This means that embodiment herein can be presented with an unlabeled dataset, and without any additional information can learn to respond to different features that are present in the data. Learning can be an ongoing process.

Another form of learning is by weight swapping:

[0133] FIG. 8 illustrates a weight swapping step of the STDP learning method, according to some embodiments. FIG. 8 shows an example of the next step in the modified STDP learning process whereby ‘unused synaptic weights’ are swapped to‘unused inputs’ thereby strengthening the neurons’ response to the same or a similar input spike pattern in the future.