Spike-timing-dependent plasticity and subcortical waves enhance alpha oscillations in a computer model of neocortex
Samuel Neymotin (SUNY Downstate / NYU-Poly), Cliff Kerr (SUNY Downstate / U Sydney), George Chadderdon (SUNY Downstate), Christopher Fietkiewicz (SUNY Downstate), William Lytton (SUNY Downstate / Kings County Hospital)
We used a previously published model consisting of 470 event-based integrate-and-fire cells (Neymotin et al., 2011). Excitatory and inhibitory cells were arranged in 13 types to reflect the laminar structure of sensory cortex. Each cell received external driving Poisson inputs, as well as inputs from other cells in the model, via AMPA, NMDA, and GABAA synapses. STDP was implemented at AMPA synapses, with a maximal spike-time difference of 40 ms and a decay time constant of 10 ms. After learning, we evaluated the spectral power in the network with application of white noise, to see if it had learned to produce activity in the trained bands, effectively retaining a memory of the training. The waves from lower centers were simulated by rhythmic inputs to layer 4 excitatory cells to simulate inputs via thalamus. At baseline, prior to STDP, the excitatory cell multiunit activity vector (MUA) power spectrum had a low-amplitude peak near 7 Hz and gradual attenuation towards higher frequencies (Fig. 1, orange).
When STDP was implemented only at synapses on excitatory cells, the network was prone to epileptic activity, including high-frequency oscillations and pathologically high firing rates. To provide balanced inhibition in the network, we subsequently implemented STDP for excitatory synapses on inhibitory cells. STDP caused a shift in the distribution of synaptic weights, with the strength of AMPA synapses on inhibitory cells increasing more than those on excitatory cells, preventing epileptic activity from emerging. After applying STDP with no training signal, the spectral peak was reduced (Fig. 1, black) due to the added inhibition in the network.
In order to train the network, we provided a subcortical training input consisting of shocks applied at several different frequencies. When a training signal of 7 Hz was applied, the activity of layer 4 excitatory cells increased, causing an increase in the spectral power. The post-training spectral peak in response to white noise was broad (7 - 22 Hz) and changed little from baseline (Fig. 1, red). A small increase in training signal rate to 8 Hz (Fig. 1, blue) produced a large increase in resultant overall spectral power. In this case, a sharp spectral peak emerged near 12.5 Hz, considerably higher than the training frequency. A smaller peak was produced at 8 Hz. Applying a 10 Hz training signal enhanced the amplitude of the spectral peak further and pushed the resultant peak upwards to form a broad profile that peaked at approximately 15 Hz (Fig. 1, green). The greater response to 8 and 10 Hz appeared to be associated with far greater spread of excitation throughout the initial, untrained network, suggesting that a resonant response can occur at these higher frequencies due to the continuation of activations for these time periods.
Our model suggests that learning rates should be properly balanced between excitatory to excitatory and excitatory to inhibitory synapses in order to avoid a transition to epileptic dynamics. Alterations in excitatory weights are necessary for learned dynamics, but learning must be modulated by interneurons. Finally, we predict that alpha/beta oscillations could emerge via STDP learning of training signals projected from subcortical areas.
Neymotin, SA and Lee, H and Park, E and Fenton, AA and Lytton, WW (2011).
Frontiers in Computational Neuroscience 5:19. doi:10.3389/fncom.2011.00019.