Benchmarks for progress in neuromorphic computing.In order for...

Benchmarks for progress in neuromorphic computing.

In order for the neuromorphic research field to advance into the mainstream of computing, it needs to start quantifying gains, standardize on benchmarks and focus on feasible application challenges.
Mike Davies (Intel)
https://www.nature.com/articles/s42256-019-0097-1.epdf?author_access_token=DVi3hfPefL86m1i_Bts9NNRgN0jAjWel9jnR3ZoTv0MK4s3LnSWqkwc7kKuB1k_WC_R3JNRCJJch_gU0hHY_9FpiDFWj3Bi4OVrtVlUI-s_xyYAfOb01qXSqb9EKnvfNmjH6Ho2XGoTM0IBPIn_q1w%3D%3D

Box 1 SpikeMark
A benchmarking suite for spiking neuromorphic systems might include the following workloads:

• Classify spoken keywords using a specified pre-trained deep neural network converted to spiking form .
• Classify sequentially presented MNIST digits and TIMIT phonemes using offline-trained long short-term memory spiking neural networks .
• Detect hand gestures from the DVS Gesture event-based camera dataset , with a convolutional spiking neural network (SNN) trained offline with SNN backpropagation and online with deep continuous local learning.
• Solve least absolute shrinkage and selection operator (LASSO) problems with the spiking locally competitive algorithm.
• Solve Sudoku and map colouring constraint satisfaction problems using neural sampling.
• Identify the shortest path in a variety of graphs using spike-based temporal wavefront propagation.
• Perform pattern similarity matching with threshold phasor associative memory8.
• Control a modelled robotic arm subjected to nonlinear wear using an adaptive controller trained with the neural engineering framework.
• Solve simple cognitive problems with the Spaun-embodied brain model.
• Simultaneous localization and mapping (SLAM) using Bayesian learning and inference with a robotic head-direction environment model.

Today, no single neuromorphic platform has successfully run all of the above workloads. However, several of these examples have been run on multiple platforms, and none depends on exotic platform-specific features beyond typical leaky-integrate-and-fire spiking neural network functionality with local learning rules.

To what extent would the Akida platform or current CNN to SNN Converter/NPC be able to achieve these kind of benchmarks I wonder?
Would it be capable of performing all these various workloads?