Minitaur, an Event-Driven FPGA-Based Spiking Network Accelerator

Daniel Neil, Shih-Chii Liu
2014 IEEE Transactions on Very Large Scale Integration (vlsi) Systems  
We demonstrate an interactive handwritten digit recognition system with a spike-based deep belief network running in real-time on SpiNNaker, a biologically inspired manycore architecture. Results show that during the simulation a SpiNNaker chip can deliver spikes in under 1 µs, with a classification latency in the order of tens of milliseconds, while consuming less than 0.3 W. Associated Track 8.1: Neural Networks and Systems: Spiking Neural Network circuits and systems I. DEMONSTRATION Deep
more » ... ral network architectures such as Deep Belief Networks (DBNs), are a very active research topic in machine learning and achieve state-of-the-art classification results. However, simulating large-scale DBNs has large computational demands, which means high energy requirements and long latencies, and thus limits their use in real-time applications for mobile or robotic platforms. A recently published study [1] developed the theory to map an offline-trained DBN into a spiking neural network composed of Leaky Integrateand-Fire (LIF) neurons, exploiting the so-called Siegert approximation for the firing rate of an LIF neuron. Spiking neural networks running on an appropriate hardware platform can allow asynchronous and massively parallel energy-efficient processing [2]. SpiNNaker [3] is a biologically-inspired manycore scalable architecture designed to enable large-scale simulations of heterogeneous models of spiking neurons efficiently and in real-time. In this demonstration, we show an application of SpiNNaker as a real-time event-based hardware prototyping platform to enable interactive handwritten digit recognition with a spiking deep belief network. II. DEMONSTRATION SETUP The demonstration setup consists of a four-chip SpiNNaker board [3], a wireless router, a laptop, and a mobile device which for the purpose of this demonstration is a tablet device, Fig 1. The user interface, which runs on the mobile device, allows the user to draw a digit and receive the classified output of the spiking deep belief network running on SpiNNaker. Fig. 1. Handwritten digit recognition setup. a) The user interface running on a mobile device, showing a finger-drawn digit (left) and a correct classification. b) A 4-chip SpiNNaker board. c) The output of the laptop monitor, showing network-classified similar digits from the training set, as well as the firing rates of neurons within each layer during classification III. VISITOR EXPERIENCE Visitors can draw digits using a tablet and the spikebased DBN will run on SpiNNaker in real-time producing a result within tens of milliseconds, while dissipating less than 0.3 W power. The classified output will be presented on the tablet as a normalised histogram of the spikes of the output population. In particular, they will see the conversion of static images to spike-trains, the propagation of neural spikes through multilayer spiking DBNs, and the resulting output of a correctly-classified handwritten digit. Additionally, the network recognizes different styles of handwriting accurately. REFERENCES [1] P. O'Connor, D. Neil, S.-C. Liu, T. Delbruck, and M. Pfeiffer, "Realtime classification and sensor fusion with a spiking deep belief network,"
doi:10.1109/tvlsi.2013.2294916 fatcat:juhmivazyrfw5jpj35jnw5giky