Storage capacity of networks with discrete synapses and sparsely encoded memories

Data-driven synapse classification reveals a logic of glutamate receptor diversity

The rich diversity of synapses facilitates the capacity of neural circuits to transmit, process and store information. We used multiplex super-resolution proteometric imaging through array tomography to define features of single synapses in mouse neocortex. We find that glutamatergic synapses cluster into subclasses that parallel the distinct biochemical and functional categories of receptor subunits: GluA1/4, GluA2/3 and GluN1/GluN2B. Two of these subclasses align with physiological expectations based on synaptic plasticity: large AMPAR-rich synapses may represent potentiated synapses, whereas small NMDAR-rich synapses suggest "silent" synapses. The NMDA receptor content of large synapses correlates with spine neck diameter, and thus the potential for coupling to the parent dendrite. Overall, ultrastructural features predict receptor content of synapses better than parent neuron identity does, suggesting synapse subclasses act as fundamental elements of neuronal circuits. No barriers prevent future generalization of this approach to other species, or to study of human disorders and therapeutics.

Biomolecular Neuristors from Functionalized Lipid Membranes

Modeled after biological neurons, neuristors are emerging hardware that generate recurring voltage spikes in response to electrical stimulation. This type of excitability could enable transistor-free spiking neural networks for efficient signal processing and computing. Yet, prior neuristors consist of circuits containing numerous devices, thus complicating fabrication and increasing size, power usage, and cost. In contrast, we show that a single, 5nm-thick lipid membrane functionalized with voltage-activated peptides functions as a two-terminal, ultra-low power (fW-pW) artificial neuristor in response to supplied current. Specifically, the biomolecular membrane generates stochastic voltage oscillations (10-150 mV) in response to DC currents (|5-40| pA), and is capable of generating two distinct types of action potentials - fast (~1-50 ms) and slow (~1-2 s) spikes via distinct physical mechanisms. This discovery showcases the inherent multifunctionality and modularity of engineered biomembranes, and it contributes to an expanding suite of ionic and biomolecular devices designed with synapse and neuron functionalities for emerging computing architectures.

Quiet Trajectories as Neural Building Blocks

Our concept of the neural mechanisms of working memory has recently undergone an upheaval, because of two transformative concepts: multivariate neural state trajectories and the activity-silent hypothesis. I will argue that putting these concepts together raises the difficult problem of "quiet trajectories," where future neural activity is not fully determined by current activity. However, this also promises new building blocks for neural computation.