The paper written by the research scientist at Sony Neuromorphic Computing refers to Intel Loihi in the conclusion?
Impact of spiking neurons leakages and network recurrences on event-based spatio-temporal pattern recognition
Mohamed Sadek Bouanane1 , Dalila Cherifi1 , Elisabetta Chicca2,3 , Lyes Khacef2,3,∗ 1 Institute of Electrical and Electronic Engineering, Univesity of Boumerdes, Algeria. 2Bio-Inspired Circuits and Systems (BICS) Lab. Zernike Institute for Advanced Materials, University of Groningen, the Netherlands. 3Groningen Cognitive Systems and Materials Center (CogniGron), University of Groningen, the Netherlands
V. CONCLUSION In this work we explored the effect of spiking neurons synaptic and membrane leakages, network explicit recurrences and time constants heterogeneity on event-based spatiotemporal pattern recognition. The main findings of our work can be summarized as follows: • Neural leakages are only important when there are both temporal information in the data and explicit recurrent connections in the network. • Neural leakages do not necessarily lead to sparser spiking activity in the network. • Time constants heterogeneity slightly improves performance on data with a rich temporal structure and does not affect performance on data with a spatial structure. This work supports the identification of the right level of model abstraction of biological evidences needed to build efficient application-specific neuromorphic hardware. This is a crucial analysis for advancing the field beyond state-of-theart, especially when constrains on resources are critical (e.g. edge computing). In fact, when using digital neuromorphic architectures, IF neurons have been shown to be 2 × smaller and more power-efficient than formal Perceptrons [23]. It is nevertheless not clear how this gain evolves when adding a multiplier to implement a LIF or CUBA-LIF neuron. Further works will focus on implementing these two architectures in FPGAs for fast prototyping. In addition, IF neurons give the possibility to implement a digital asynchronous processing purely driven by the input, since there is no inherent temporal dynamics in the spiking neurons. On the other hand, LIF and CUBA-LIF neurons require algorithmic time-steps where the leakage is updated regardless of the presence of input spikes. Further works will explore the impact of both paradigms in energy-efficiency on the Loihi neuromorphic chip [13]. Furthermore, it is important to mention that our results only hold in benchmarking so far. In a real-world scenario such as continuous keyword spotting, there can be more noise in the data but also in void. Hence, when using the IF neurons that do not have any leakage, this noise can accumulate and create false positives and degrade the performance. Indeed, the low-pass filtering effect of the spiking neurons leakages has been shown to eliminate high frequency components from the input and enhance the noise robustness of SNNs, especially in real-world environments [34]. In addition, given that the LIF model achieved a superior performance when compared to the CUBA-LIF, it is important to investigate where the latter could perform better. More complex tasks could show such a gain for the CUBA-LIF neuron, because of its current compartment which is an extra state that gives more potential for spatio-temporal feature extraction. Finally, spiking neural networks in neuromorphic hardware can be used beyond fast and efficient inference, by adding adaptation through local synaptic plasticity [43], [44], [45]. In this context, the impact of the leakage can be different, as the inherent temporal dynamics is required in some plasticity mechanisms [46], [47] for online learning.
ACKNOWLEDGEMENTS We would like to thank Dylan Muir for the fruitful discussion on DynapCNN. We would like to thank the Institute of Electrical and Electronic Engineering at the University of Boumerdes for supporting this work. We would also like to acknowledge the financial support of the CogniGron research center and the Ubbo Emmius Funds of the University of Groningen."
Not sure where Sony comes into the frame but they find quite a lot of issues with Loihi that need to be resolved.
My opinion only DYOR
FF
AKIDA BALLISTA