Implementing a foveal-pit inspired filter in a Spiking Convolutional Neural Network: a preliminary study

Abstract

We have presented a Spiking Convolutional Neural Network (SCNN) that incorporates retinal foveal-pit inspired Difference of Gaussian filters and rank-order encoding. The model is trained using a variant of the backpropagation algorithm adapted to work with spiking neurons, as implemented in the Nengo library. We have evaluated the performance of our model on two publicly available datasets -one for digit recognition task, and the other for vehicle recognition task. The network has achieved up to 90% accuracy, where loss is calculated using the cross-entropy function. This is an improvement over around 57% accuracy obtained with the alternate approach of performing the classification without any kind of neural filtering. Overall, our proof-of-concept study indicates that introducing biologically plausible filtering in existing SCNN architecture will work well with noisy input images such as those in our vehicle recognition task. Based on our results, we plan to enhance our SCNN by integrating lateral inhibition-based redundancy reduction prior to rank-ordering, which will further improve the classification accuracy by the network.

Foveal-pit inspired filtering of DVS spike response

“…The proposed model demonstrates the effect of applying neural filtering to real DVS data generated from a neuromorphic vision sensor. This builds upon our previous work [15] that depicted the results of foveal-pit inspired filtering for synthetically generated datasets like MNIST [3] and Caltech [16]. Our model achieves a promising performance of 92.5% using the unshifted off-center parasol ganglion cell and an accuracy of 100% in the circular-shifted scenario, which is an improvement of 35% over the classification using unfiltered DVS responses.…”

“…The SCNN is trained end-to-end using a spiking approximation of the backpropogation algorithm adapted for SNNs. This is done by minimizing the NLL loss using the procedure described in Gupta et al [15] with a duration of 3 epochs and a mini-batch size of 20.…”

“…In our previous work [15], we had demonstrated the effect of foveal-pit inspired filtering for synthetically generated datasets like MNIST [3] and Caltech [16]. In this work, we present the results of applying similar neural filtering to data generated by the DVS.…”

“…The SCNN architecture is inspired by two previous works viz. Diehl et al [7] and Kheradpisheh et al [6], while the model is implemented as in Gupta et al [15]. Each input is presented to the network for a total duration of 60 timesteps and the predictions are assigned based on the voltages measured from the output neurons.…”