摘要:
Certain aspects of the present disclosure provide methods and apparatus for spiking neural computation of general linear systems. One example aspect is a neuron model that codes information in the relative timing between spikes. However, synaptic weights are unnecessary. In other words, a connection may either exist (significant synapse) or not (insignificant or non-existent synapse). Certain aspects of the present disclosure use binary-valued inputs and outputs and do not require post-synaptic filtering. However, certain aspects may involve modeling of connection delays (e.g., dendritic delays). A single neuron model may be used to compute any general linear transformation x=AX+BU to any arbitrary precision. This neuron model may also be capable of learning, such as learning input delays (e.g., corresponding to scaling values) to achieve a target output delay (or output value). Learning may also be used to determine a logical relation of causal inputs.
摘要:
Certain aspects of the present disclosure provide methods and apparatus for spiking neural computation of general linear systems. One example aspect is a neuron model that codes information in the relative timing between spikes. However, synaptic weights are unnecessary. In other words, a connection may either exist (significant synapse) or not (insignificant or non-existent synapse). Certain aspects of the present disclosure use binary-valued inputs and outputs and do not require post-synaptic filtering. However, certain aspects may involve modeling of connection delays (e.g., dendritic delays). A single neuron model may be used to compute any general linear transformation x=AX+BU to any arbitrary precision. This neuron model may also be capable of learning, such as learning input delays (e.g., corresponding to scaling values) to achieve a target output delay (or output value). Learning may also be used to determine a logical relation of causal inputs.
摘要:
Certain embodiments of the present disclosure support implementation of a neural processor with synaptic weights, wherein training of the synapse weights is based on encouraging a specific output neuron to generate a spike. The implemented neural processor can be applied for classification of images and other patterns.
摘要:
Certain aspects of the present disclosure present a technique for primary visual cortex (V1) cell training and operation. The present disclosure proposes a model structure of V1 cells and retinal ganglion cells (RGCs), and an efficient method of training connectivity between these two layers of cells such that the proposed method leads to an autonomous formation of feature detectors within the V1 layer. The proposed approach enables a hardware-efficient and biological-plausible implementation of image recognition and motion detection systems.
摘要:
Certain embodiments of the present disclosure support techniques for power efficient implementation of neuron synapses with positive and/or negative synaptic weights.
摘要:
Certain embodiments of the present disclosure support techniques for storing synaptic weights separately from a neuro-processor chip into a replaceable storage. The replaceable synaptic memory gives a unique functionality to the neuro-processor and improves its flexibility for supporting a large variety of applications. In addition, the replaceable synaptic storage can provide more choices for the type of memory used, and might decrease the area and implementation cost of the overall neuro-processor chip.
摘要:
Certain aspects of the present disclosure present a technique for unsupervised training of input synapses of primary visual cortex (V1) simple cells and other neural circuits. The proposed unsupervised training method utilizes simple neuron models for both Retinal Ganglion Cell (RGC) and V1 layers. The model simply adds the weighted inputs of each cell, wherein the inputs can have positive or negative values. The resulting weighted sums of inputs represent activations that can also be positive or negative. In an aspect of the present disclosure, the weights of each V1 cell can be adjusted depending on a sign of corresponding RGC output and a sign of activation of that V1 cell in the direction of increasing the absolute value of the activation. The RGC-to-V1 weights can be positive and negative for modeling ON and OFF RGCs, respectively.
摘要:
Certain aspects of the present disclosure support a local competitive learning rule applied in a computational network that leads to sparse connectivity among processing units of the network. The present disclosure provides a modification to the Oja learning rule, modifying the constraint on the sum of squared weights in the Oja rule. This constraining can be intrinsic and local as opposed to the commonly used multiplicative and subtractive normalizations, which are explicit and require the knowledge of all input weights of a processing unit to update each one of them individually. The presented rule provides convergence to a weight vector that is sparser (i.e., has more zero elements) than the weight vector learned by the original Oja rule. Such sparse connectivity can lead to a higher selectivity of processing units to specific features, and it may require less memory to store the network configuration and less energy to operate it.
摘要:
Exemplary embodiments of the disclosure are directed to down-converting an RF signal of a transmitter to baseband, filtering the down-converted signal, and generating a composite signal based on the filtered down-converted signal and a transmission based-band signal.
摘要:
Certain embodiments of the present disclosure support implementation of a neural processor with synaptic weights, wherein training of the synapse weights is based on encouraging a specific output neuron to generate a spike. The implemented neural processor can be applied for classification of images and other patterns.