CAVIAR: A 45k Neuron, 5M Synapse, 12G Connects/s AER Hardware Sensory–Processing– Learning–Actuating System for High-Speed Visual Object Recognition and Tracking
read more
Citations
Deep learning in neural networks
Convolutional networks and applications in vision
The SpiNNaker Project
The SpiNNaker project
Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing
References
Self Organization And Associative Memory
Speed of processing in the human visual system.
The synaptic organization of the brain
Face recognition: a convolutional neural-network approach
Related Papers (5)
A million spiking-neuron integrated circuit with a scalable communication network and interface
Neuromorphic Silicon Neuron Circuits
Frequently Asked Questions (13)
Q2. What future works have the authors mentioned in the paper "Connects/s aer hardware sensory–processing– learning–actuating system for high-speed visual object recognition and tracking" ?
The authors plan to miniaturize it by about 3–4 orders of magnitude within the next few years, by increasing the numbers of synapses and neurons per AER-module and by integrating more modules into a smaller physical volume. Assuming timing delays similar to those reported in this paper, preliminary results [ 91 ] suggest that these systems could perform sophisticated object recognition with delays around 100 s. With such developments, the authors will be able to provide a modular and scalable platform for real-time implementations of neural models of a really challenging complexity. Coupling this massive preprocessing power with flexible back-ends of conventional procedural computation will enable solutions to a host of practical applications.
Q3. How many connections can a CAVIAR system perform per second?
The CAVIAR system consists of about 45k spiking neurons and 5M synapses; and it can perform up to 12G connections/operations per second.
Q4. How fast does the jAER software send AEs?
Monitor PCBs connect to a host computer through a high-speed USB2.0 connection, sending AEs at a speed of up to a peak rate of 6 Meps.
Q5. What is the ability of the WTA chip to learn to classify spatio–temporal?
It is capable of both spike-based learning (or spike-timing-dependent plasticity [82]) to learn to classify spatio–temporal spike patterns and rate-based Hebbian learning to learn spatio–temporal activity patterns.
Q6. What is the way to visualize a slow-motion sequence of events?
For high-speed phenomena, one can configure a time slice of very short duration (down to a few microseconds) and visualize a slow-motion recorded sequence of events offline.
Q7. How many neurons can specialize on the input pattern?
One could have expected that there are still only four different input patterns and that maximally four neurons could specialize on exactly those four patterns, but since this real-world input is changing its state in a continuous fashion rather than just assuming four discrete states, some of the neurons have become selective for transitory states “between” the four positions.
Q8. How long does the delay line train take to be tapped?
Those delay lines are tapped at three different delays (approximately 0 s, 200 ms, and 400 ms) and the resulting 2 2 3 spike trains are passed on to the learning chip.
Q9. How many categories can be categorized by the learning classifier chip?
The task of the learning classifier chip is now to provide a good representation of at most 32 categories (since there are 32 neurons) from the repeated spatio–temporal pattern.
Q10. What is the servo system for changing the central view point?
The authors also implemented a fully electronic (without mirrors, mechanical parts, or motors) servo system for changing the central view point.
Q11. How many synaptic connections can be processed in a single chip?
Their convolution chips are very efficient in this sense, because for each input event, they can process up to synaptic connections in 330 ns connections/s/chip.
Q12. What is the effect of the global inhibitory neuron on the other quadrants?
this neuron will activate the global inhibitory neurons 2 of the remaining three quadrants the most if its quadrant receives the highest input rate out of the four quadrants.
Q13. What is the simplest way to track the movement of an object in space?
The learning chip can then, for example, track the 3-D movement of an object in space by programming the same feature shape at different sizes in the different convolution chips.