Context-dependent computation with spike-based learning, for compact low-power embedded systems

Advances in technology are allowing the integration of
ever more sensory and computing devices in embedded systems and mobile computing
platforms. By collecting and processing the large amount of diverse and highly infor-
mative data produced by the embedded sensors, it is possible to realize electronic de-
vices that are endowed with cognitive abilities, for example to autonomously adapt to the
needs of the context and the specific usage, thus optimizing resources. However, the ap-
plication of conventional signal processing and machine learning methods to implement
such cognitive abilities is typically restricted to specific problem domains such as auditory
scene analysis, gesture, or character recognition, and requires prohibitively large amount
of resources in terms of memory, computational power and energy. To overcome these
limitations, we propose a spike-based neural processing model, optimally suited for neu-
romorphic Very Large Scale Integration (VLSI) implementation, that can robustly collect
multi-modal information using a pool of randomly connected neurons and autonomously
learn through continuous stochastic learning to trigger appropriate outputs depending on
the context. The model has two major components. One is a recurrent neural network
with distributed plastic synapses which represents the “memory” of the system, i.e., the
internal state, and implements state-dependent computation through attractor dynamics.
The second is a pool of neurons that receive spikes from both the input layer and the
recurrent network to provide a distributed and highly diverse representation of the mem-
ories and state-transitions to learn. These components and the distributed stochastic
learning implemented by the synapse models are optimally suited for implementation in
both ultra low-power hybrid analog/digital neuromorphic circuits and future nano-scale
implementations of spiking neural network circuits because they minimize the amount of
device power and size requirements, while being robust to the noise, variability and lim-
ited precision figures that affect these technologies. Furthermore, as the network doesn’t
require computationally demanding and task-specific pre-processing modules, it can be
used as a general-purpose computing unit and applied to a wide variety of application
scenarios without compromising the system’s scalability. In this project we will develop
the model, validate it on state-of-the-art mixed analog and digital neuromorphic com-
puting VLSI architectures, and apply it to state-dependent classification tasks in typical
real-world scenarios.