Stability and flexibility of odor representations in the mouse olfactory bulb.

Dynamic changes in sensory representations have been basic tenants of studies in neural coding and plasticity. In olfaction, relatively little is known about the dynamic range of changes in odor representations under different brain states and over time. Here, we used time-lapse in vivo two-photon calcium imaging to describe changes in odor representation by mitral cells, the output neurons of the mouse olfactory bulb. Using anesthetics as a gross manipulation to switch between different brain states (wakefulness and under anesthesia), we found that odor representations by mitral cells undergo significant re-shaping across states but not over time within state. Odor representations were well balanced across the population in the awake state yet highly diverse under anesthesia. To evaluate differences in odor representation across states, we used linear classifiers to decode odor identity in one state based on training data from the other state. Decoding across states resulted in nearly chance-level accuracy. In contrast, repeating the same procedure for data recorded within the same state but in different time points, showed that time had a rather minor impact on odor representations. Relative to the differences across states, odor representations remained stable over months. Thus, single mitral cells can change dynamically across states but maintain robust representations across months. These findings have implications for sensory coding and plasticity in the mammalian brain.

Single cortical neurons as deep artificial neural networks.

Utilizing recent advances in machine learning, we introduce a systematic approach to characterize neurons' input/output (I/O) mapping complexity. Deep neural networks (DNNs) were trained to faithfully replicate the I/O function of various biophysical models of cortical neurons at millisecond (spiking) resolution. A temporally convolutional DNN with five to eight layers was required to capture the I/O mapping of a realistic model of a layer 5 cortical pyramidal cell (L5PC). This DNN generalized well when presented with inputs widely outside the training distribution. When NMDA receptors were removed, a much simpler network (fully connected neural network with one hidden layer) was sufficient to fit the model. Analysis of the DNNs' weight matrices revealed that synaptic integration in dendritic branches could be conceptualized as pattern matching from a set of spatiotemporal templates. This study provides a unified characterization of the computational complexity of single neurons and suggests that cortical networks therefore have a unique architecture, potentially supporting their computational power.