Post on 27-Dec-2015
Bump attractors
• Can explain sustained (but bounded) activity in populations of neurons.
• A useful description of working memory by combining self-sustained activity with sensitivity to external inputs.
Firing-Rate Model
τr time constantri (t) firing rate of neuron i Ii (t) input current for neuron i Jij synaptic strength between neurons i,j Θ current bias
Homogeneity
• Homogeneity is required to create a continuum of bump attractors. – Even a small amount of heterogeneity
destroys continuum, leaving only a few discrete attractors.
• Biologically implausible: number and strength of synaptic connections is variable.
Solutions
• Fine tuning properties of each neuron.
• Network learns to tune itself through an activity-dependent mechanism.– “Activity-dependent scaling of synaptic
weights, which up- or downregulates excitatory inputs so that the long term average firing rate is similar for each neuron” (Renart, Song, Wang 2003).
Synaptic Scaling (Renart, Song, Wang 2003)
τg time constant [large] g(θ) factor that multiplies excitatory synaptic conductances to neuron θ r(θ) instantaneous firing rate of neuron θ rtg(θ) target firing rate of neuron θ
Project Outline
1. Simulate network of firing-rate neurons.
2. Observe bump attractors.
3. Show how loss of symmetry destroys continuum of bump attractors.
4. Restore symmetry by activity-dependent scaling of synaptic weights.
Recurrent Networks
• Three layer network with feedback: input layer, intermediate layer, and output layer
• Intermediate layer in this instance are SAM cells
Spike Accumulation Model (SAM)
• Variable of importance – the accumulated potential of SAM cells in intermediate layer– The accumulated potential = a constantly-
updating characterization of a constant stream of sensory input
Membrane Time Constant
• Sequential memory dependent on membrane time constant (τ) of SAM cells– Different response to stimuli 1 2 than 2
1 due to difference in neuronal response– Activity of the entire population of SAM cells
important
Output
• Sequential learning decoded due to input from SAM cells AND efferent copy of activity of output layer to intermediate layer– How do we decode this? How does the
efferent feedback loop aid decoding?– Activation of output cell:
Learning
• How will the network learn a series of tones?
• The input to the SAM cells are determined by a comparison of the output of the output cells to the output of the input cells
Encode Specificity and Timing
• Specificity – tone 1 vs. tone 2– Auditory tuning within SAM cells could
determine the frequency of the input
• Timing - tone 1 tone 2 OR tone 2 tone 1– The membrane time constant and the efferent
copy of the output will help determine the order of the tones
Supervised learning with lateral interaction and back propagation
within a neural network4/27/11
Sunmee Park, Jing Wang, Rizwan Huq
Computational neuroscience 2011
Motivation
• Natural vision can differentiate various handwritten digits
• Can we mimic the vision system? (Of course, in a simpler way)
• Supervised learning: given input/output– Traditional backpropagation neural
network(NN) – Biophysically appropriate inter-layer
communications
Methods
…
- Hand-written digits set from 0~ 9, collected by USPS/NIST database
- 8 bit grayscale images (1100 examples of each class)
Input dataset
Deliverables
• A simulated neural system capable of decoding handwritten image data and identifying the input digit.
• System performance curves, for which we will reserve training examples.
• Analysis of the impact inter-layer communications.
Kuhn, et al. 2004: Neuronal Integration of Synaptic Input in the Fluctuation-Driven Regime
Project Team: Tommy Tea and Christina Hahn
Paper Recap
• Visual synaptic bombardment leads to changes in conductance. This in turn increases fluctuations in membrane potential and these fluctuations modify the firing rate.
• Firing rate decreases because of shunted membrane potential fluctuations, and increases because of shorter membrane time constants, allowing for faster membrane potential fluctuations.