Benchmark Network II

Input type: Poisson spikes (NE15-Poissonian)

Network:

• Spiking Deep Belief Network (DBN)

• One input layer network

• Two hidden layers (500 neurons each)

• Fully connected decision layer (10 neurons)

• Current based LIF-exp neurons:

  Parameter	Values	Units
  Tau_m	5	s
  Tau_refrac	2	ms
  V_reset	0	mV
  V_rest	0	mV
  V_thresh	1	mV

Training:

• Off-line training
• Unsupervised standard Contrastive Divergence for first layers
• Decision layer is trained with supervision
• Siegert approximation as the activation function

Testing:

• 1.5 kHz input rate
• Weights are just transferred to the spiking network

Performance:

• 94.94% accuracy
• 16ms latency
• 1.88M Sopbs

Hardware Platform	Accuracy (%)	Sim Time (s)	Enery (KJ)	Ref
SpiNNaker	94.94	10, 000	2.97