Neurons integrate from thousands of synapses whose strengths span an order of magnitude. Intriguingly, in mouse neocortex, the few 'strong' synapses are formed between similarly tuned cells, suggesting they determine spiking output. This raises the question of how other computational primitives, including 'background' activity from the many 'weak' synapses, short-term plasticity, and temporal factors contribute to spiking. We used paired recordings and extracellular stimulation experiments to map excitatory postsynaptic potential (EPSP) amplitudes and paired-pulse ratios of synaptic connections formed between pyramidal neurons in layer 2/3 (L2/3) of barrel cortex. While net short-term plasticity was weak, strong synaptic connections were exclusively depressing. Importantly, we found no evidence for clustering of synaptic properties on individual neurons. Instead, EPSPs and paired-pulse ratios of connections converging onto the same cells spanned the full range observed across L2/3, which critically constrains theoretical models of cortical filtering. To investigate how different computational primitives of synaptic information processing interact to shape spiking, we developed a computational model of a pyramidal neuron in the excitatory L2/3 circuitry, which was constrained by our experiments and published in vivo data. We found that strong synapses were substantially depressed during ongoing activation and their ability to evoke correlated spiking primarily depended on their high temporal synchrony and high firing rates observed in vivo. However, despite this depression, their larger EPSP amplitudes strongly amplified information transfer and responsiveness. Thus, our results contribute to a nuanced framework of how cortical neurons exploit synergies between temporal coding, synaptic properties, and noise to transform synaptic inputs into spikes.

Learning a sequence of tasks without access to i.i.d. observations is a widely studied form of continual learning (CL) that remains challenging. In principle, Bayesian learning directly applies to this setting, since recursive and one-off Bayesian updates yield the same result. In practice, however, recursive updating often leads to poor trade-off solutions across tasks because approximate inference is necessary for most models of interest. Here, we describe an alternative Bayesian approach where task-conditioned parameter distributions are continually inferred from data. We offer a practical deep learning implementation of our framework based on probabilistic task-conditioned hypernetworks, an approach we term posterior meta-replay. Experiments on standard benchmarks show that our probabilistic hypernetworks compress sequences of posterior parameter distributions with virtually no forgetting. We obtain considerable performance gains compared to existing Bayesian CL methods, and identify task inference as our major limiting factor. This limitation has several causes that are independent of the considered sequential setting, opening up new avenues for progress in CL.