Restoring physiological oscillations using neuroprosthetic spike-timing-dependent plasticity in computer model of neocortex

TitleRestoring physiological oscillations using neuroprosthetic spike-timing-dependent plasticity in computer model of neocortex
Publication TypeConference Paper
Year of Publication2011
AuthorsNeymotin, SA., Kerr CC., Chadderdon GL., Francis JT., & Lytton WW.
Conference NameSociety for Neuroscience
Abstract

Alpha oscillation is a major component of neocortical oscillation that emerges over the first several years of life and likely contributes to cortical information processing. We hypothesized that a periodic input signal could be used to train cortical oscillation using spike-timing-dependent plasticity (STDP). Our model consisted of 470 event-based integrate-and-fire cells arranged to reflect sensory cortex layers. Cells received Poisson inputs and inputs from other cells via AMPA, NMDA, and GABAA synapses. STDP was implemented on AMPA synapses using a basic model with bilateral exponential decay (40 ms maximal interspike difference, 10 ms tau). At baseline, excitatory (E) and inhibitory (I) cells fired between 0.5-2 and 2-8 Hz, respectively. Our primary measurement was the E cell multiunit activity vector (MUA) power spectrum. Prior to training, spectrum was relatively flat with a low amplitude peak at  7 Hz. We provided training input to layer 4 E cells, as would be produced by thalamic or subcortical white matter stimulation. To avoid epileptic activity from emerging, we found that we needed to balance E->E STDP learning with comparable rate of learning E->I synapses. Even so, epileptic activity typically emerged after long periods of training. We trained a naive network with inputs at various frequencies. A 7 Hz training signal was ineffective in changing the spectrum, whereas training signals of 8,10 Hz produced sharp post-training spectral peaks at 12.5,15 Hz. Subsequent application of white noise with STDP tended to erase the spectral peaks. We also looked at retraining following synaptic erasure, as would be done experimentally using zeta inhibitory protein (ZIP, an inhibitor of PKM-zeta). To simulate ZIP application, we set synaptic weights of AMPA synapses to 20% of baseline level, producing a reduction in firing rates and a fairly flat spectrum. After ZIP erasure, the network showed very few intrinsic correlations and could be retrained by utilizing larger STDP increments for E->E synapses than for E->I synapses. Under these conditions, we were able to achieve a spectral profile with strong similarity to the original after 4 minutes of training at 8Hz. We were able to obtain a similar result using a 30 Hz signal, suggesting that the specific spectral profile emerges naturally from the network. Our model suggests that a neuroprosthesis could be used to train naive cortex or to retrain cortex after damage or after erasure. We suggest that these neuroprosthetic procedures could be used to repurpose cortical areas in order to utilize them to augment lost or compromised function in damaged areas.