Deep Predictive Learning in Neocortex and Pulvinar

How do humans learn from raw sensory experience? Throughout life, but most obviously in infancy, we learn without explicit instruction. We propose a detailed biological mechanism for the widely embraced idea that learning is driven by the differences between predictions and actual outcomes (i.e., predictive error-driven learning). Specifically, numerous weak projections into the pulvinar nucleus of the thalamus generate top–down predictions, and sparse driver inputs from lower areas supply the actual outcome, originating in Layer 5 intrinsic bursting neurons. Thus, the outcome representation is only briefly activated, roughly every 100 msec (i.e., 10 Hz, alpha), resulting in a temporal difference error signal, which drives local synaptic changes throughout the neocortex. This results in a biologically plausible form of error backpropagation learning. We implemented these mechanisms in a large-scale model of the visual system and found that the simulated inferotemporal pathway learns to systematically categorize 3-D objects according to invariant shape properties, based solely on predictive learning from raw visual inputs. These categories match human judgments on the same stimuli and are consistent with neural representations in inferotemporal cortex in primates.

Operant conditioning of motor cortex neurons reveals neuron-subtype-specific responses in a brain-machine interface task

Operant conditioning is implemented in brain-machine interfaces (BMI) to induce rapid volitional modulation of single neuron activity to control arbitrary mappings with an external actuator. However, intrinsic factors of the volitional controller (i.e. the brain) or the output stage (i.e. individual neurons) might hinder performance of BMIs with more complex mappings between hundreds of neurons and actuators with multiple degrees of freedom. Improved performance might be achieved by studying these intrinsic factors in the context of BMI control. In this study, we investigated how neuron subtypes respond and adapt to a given BMI task. We conditioned single cortical neurons in a BMI task. Recorded neurons were classified into bursting and non-bursting subtypes based on their spike-train autocorrelation. Both neuron subtypes had similar improvement in performance and change in average firing rate. However, in bursting neurons, the activity leading up to a reward increased progressively throughout conditioning, while the response of non-bursting neurons did not change during conditioning. These results highlight the need to characterize neuron-subtype-specific responses in a variety of tasks, which might ultimately inform the design and implementation of BMIs.

High frequency neuronal bursting is essential for circadian and sleep behaviors in Drosophila

Circadian rhythms have been extensively studied in Drosophila, however, still little is known about how the electrical properties of clock neurons are specified. We have performed a behavioral genetic screen through the downregulation of candidate ion channels in the lateral ventral neurons (LNvs) and show that the hyperpolarization-activated cation current Ih is important for the behaviors that the LNvs command: temporal organization of locomotor activity and sleep. Using whole-cell patch clamp electrophysiology we demonstrate that small LNvs are bursting neurons, and that Ih is necessary to achieve the high frequency bursting firing pattern characteristic of both types of LNvs. Since firing in bursts has been associated to neuropeptide release, we hypothesized that Ih would be important for LNvs communication. Indeed, herein we demonstrate that Ih is fundamental for the recruitment of PDF filled dense core vesicles to the terminals at the dorsal protocerebrum and for their timed release, and hence for the temporal coordination of circadian behaviors.

Itinerant complexity in networks of intrinsically bursting neurons

Active neurons can be broadly classified by their intrinsic oscillation patterns into two classes characterized by periodic spiking or periodic bursting. Here we show that networks of identical bursting neurons with inhibitory pulsatory coupling exhibit itinerant dynamics. Using the relative phases of bursts between neurons, we numerically demonstrate that the network exhibits endogenous transitions among multiple modes of transient synchrony. This is true even for bursts consisting of two spikes. In contrast, our simulations reveal that identical singlet-spiking neurons do not exhibit such complexity in the network. These results suggest a role for bursting dynamics in realizing itinerant complexity in neural circuits.

Choice can be predicted from populations of bursting neurons in superficial layers of monkey V1

Primary visual cortex (V1) is absolutely necessary for normal visual processing, but whether V1 encodes upcoming behavioral decisions based on visual information is an unresolved issue, with conflicting evidence. Further, no study so far has been able to predict choice from time-resolved spiking activity in V1. Here, we hypothesized that the choice cannot be decoded with classical decoding schemes due to the noise in incorrect trials, but it might be possible to decode with generalized learning. We trained the decoder in the presence of the information on both the stimulus class and the correct behavioral choice. The learned structure of population responses was then utilized to decode trials that differ in the choice alone. We show that with such generalized learning scheme, the choice can be successfully predicted from spiking activity of neural ensembles in V1 in single trials, relying on the partial overlap in the representation between the stimuli and the choice. In addition, we show that the representation of the choice is primarily carried by bursting neurons in the superficial layer of the cortex. We demonstrated how bursting of single neurons and noise correlations between neurons with similar decoding selectivity helps the accumulation of the choice signal.HighlightsThe choice can be predicted from the spiking activity in the cortical column of V1 of the macaque.The information on choice and on stimuli is partially overlapping.Bursty neurons in the superficial layer of the cortex are the principal carriers of the choice signal.Correlated spike timing between neurons with similar decoding selectivity helps encoding.