Initial specificationsUp to Neural network architecture and design
Neural network algorithms and typesPosted by Giacomo Indiveri at February 27. 2015
Here are some properties of the neural networks that I propose as a staring point for our discussions and project definition:
- Spiking: we will investigate spiking neural network (as opposed to graded or analog)
- Configurable: we will aim for a flexible implementation that will allow us to explore different network topologies, including recurrent, multi-layer, convolutional, deep, etc.
- Dynamic synapses: the synapses will not be a simple multiplier, but will have complex dynamics and tunable non-linear characteristics, and tunable time constants
- Limited resolution: the algorithms will have to work with limited resolution synaptic weights (i.e. the synapses will produce analog continuous currents but multiplied by weight values that can take only a small number of bits).
- Real-time dynamics: both synapses and neurons will have the possibility of using biologically plausible time constants, for use cases that require real-time interaction with the environment (e.g. understanding an auditory scene in real-time).
- Possibility to speed-up computation: the synapses and neurons will allow the possibility to run faster than real-time (at the price of higher power consumption)
- Learning versus no-learning: as a short-term goal (to minimize risk) we will tackle problems that can be solved with off-line learning paradigms that down-load the weights to the VLSI device once the weights have been learned. As a long-term goal however we should include learning in the loop, and include the possibility to implement in-situ on-line learning.