hide
Free keywords:
-
Abstract:
Input selectivity, the ability of cortical neurons to respond differently to distinct stimuli, is one of the fundamental components of neural computation. While originally input selectivity was exclusively attributed to differences in the incoming excitatory currents arriving in a neuron, it is now well-known that inhibitory currents also vary between different stimuli [1]. Moreover, the incoming E and I currents cancel each other out over periods of time (E/I balance), which is hypothesized to be essential for efficient neural coding [2]. Here, we aim to examine if the balance and co-tuning of excitatory and inhibitory inputs emerge when neuronal connections are trained explicitly for an input discrimination task. Specifically, we use a single postsynaptic leaky-integrate-and-fire neuron that receives input from several presynaptic populations of E and I neurons (Fig. 1A). We train the feedforward E and I connections using surrogate gradient learning [3] to enable the postsynaptic neuron to discriminate which presynaptic input group is active. To achieve this, we train a classifier to identify which presynaptic group was active from the spikes the postsynaptic neuron emits and use the classification loss to train the feedforward connections with surrogate gradient learning. After the training process, we observe that increasing the constraint on postsynaptic firing rate forces E/I balance, which is consistent with previous claims about efficient coding. Moreover, we find that while inhibitory input selectivity does emerge, it is consistently inversely correlated with the excitatory input selectivity (i.e., a group with relatively strong E feedforward weights will develop weak I weights and vice versa) (Fig. 1B, 1C). This type of connectivity has been investigated in theoretical studies [4] and can emerge via synaptic plasticity mechanisms. Our findings suggest that while E/I balance and inhibitory input selectivity are necessary components of neural computation, the co-tuning of excitatory and inhibitory currents may not be necessary for discriminating simple inputs. We finally hypothesized that in more complex discrimination tasks that require temporal precision, E/I co-tuning is more likely to emerge via training.
