chanokin / l2l-omniglot Goto Github PK

This repository contains an insect-inspired Spiking Neural Network (SNN) whose purpose is the correct recognition/classification of the presented input. This is done through unsupervised learning and competition. The network is expressed in the PyNN network description language and experiments are run with Graphics Processing Units (GPUs; through GeNN) and the SpiNNaker neuromorphic system as compute backends.

• The network consists of three layers (input, middle and output):

  • The Input layer represents an encoding mechanism, in this case we are using rank-order encoding of images through a procedure which resembles the computation in the mammalian retina. The code used to convert the images can be located here. Since the conversion is slow, the encoded images are attached to PyNN SpikeSourceArrays

  • The Middle layer is inspired by a region of insects' mushroom body. Its purpose is to perform a dimensional expansion and sparsify the representation. To achieve this the probability of input connectivity is low (~10%), we added a distance constraint so that neurons in the Middle layer can only connect to nearby Input neurons. Neurons who share input regions will compete through mutual inhibition (or a soft winner-takes-all [sWTA] per sub-population).

  • The Output layer is another region of insects' mushroom body. The main purpose in our experiment is to have a readout region for the classification of the input patten. The synapses coming from the Middle layer are modified through an unsupervised spike-timing-dependent plasticity (STDP) rule, together with a sWTA circuit to promote specialization of Output neurons.

• To tune the network hyper-parameters we use the Learning-to-Learn framework developed by colleagues from the TU Graz and the Jülich Supercomputing Centre.

• Currently working on:

  • Juwels super-computer (GPU and SLURM)
  • HBP neuromorphic platform (SpiNNaker machine and MPI)
  • Desktop (GPU and MPI)