Learning beyond sensations: How dreams organize neuronal representations.

Semantic representations in higher sensory cortices form the basis for robust, yet flexible behavior. These representations are acquired over the course of development in an unsupervised fashion and continuously maintained over an organism's lifespan. Predictive processing theories propose that these representations emerge from predicting or reconstructing sensory inputs. However, brains are known to generate virtual experiences, such as during imagination and dreaming, that go beyond previously experienced inputs. Here, we suggest that virtual experiences may be just as relevant as actual sensory inputs in shaping cortical representations. In particular, we discuss two complementary learning principles that organize representations through the generation of virtual experiences. First, "adversarial dreaming" proposes that creative dreams support a cortical implementation of adversarial learning in which feedback and feedforward pathways engage in a productive game of trying to fool each other. Second, "contrastive dreaming" proposes that the invariance of neuronal representations to irrelevant factors of variation is acquired by trying to map similar virtual experiences together via a contrastive learning process. These principles are compatible with known cortical structure and dynamics and the phenomenology of sleep thus providing promising directions to explain cortical learning beyond the classical predictive processing paradigm.

Learning cortical representations through perturbed and adversarial dreaming.

Humans and other animals learn to extract general concepts from sensory experience without extensive teaching. This ability is thought to be facilitated by offline states like sleep where previous experiences are systemically replayed. However, the characteristic creative nature of dreams suggests that learning semantic representations may go beyond merely replaying previous experiences. We support this hypothesis by implementing a cortical architecture inspired by generative adversarial networks (GANs). Learning in our model is organized across three different global brain states mimicking wakefulness, non-rapid eye movement (NREM), and REM sleep, optimizing different, but complementary, objective functions. We train the model on standard datasets of natural images and evaluate the quality of the learned representations. Our results suggest that generating new, virtual sensory inputs via adversarial dreaming during REM sleep is essential for extracting semantic concepts, while replaying episodic memories via perturbed dreaming during NREM sleep improves the robustness of latent representations. The model provides a new computational perspective on sleep states, memory replay, and dreams, and suggests a cortical implementation of GANs.

Thalamus mediates neocortical Down state transition via GABAB-receptor-targeting interneurons.

Slow-wave sleep is characterized by near-synchronous alternation of active Up states and quiescent Down states in the neocortex. Although the cortex itself can maintain these oscillations, the full expression of Up-Down states requires intact thalamocortical circuits. Sensory thalamic input can drive the cortex into an Up state. Here we show that midline thalamic neurons terminate Up states synchronously across cortical areas. Combining local field potential, single-unit, and patch-clamp recordings in conjunction with optogenetic stimulation and silencing in mice in vivo, we report that thalamic input mediates Down transition via activation of layer 1 neurogliaform inhibitory neurons acting on GABAB receptors. These results strengthen the evidence that thalamocortical interactions are essential for the full expression of slow-wave sleep, show that Down transition is an active process mediated by cortical GABAB receptors, and demonstrate that thalamus synchronizes Down transitions across cortical areas during natural slow-wave sleep.

Short-term depression and long-term plasticity together tune sensitive range of synaptic plasticity.

Synaptic efficacy is subjected to activity-dependent changes on short- and long time scales. While short-term changes decay over minutes, long-term modifications last from hours up to a lifetime and are thought to constitute the basis of learning and memory. Both plasticity mechanisms have been studied extensively but how their interaction shapes synaptic dynamics is little known. To investigate how both short- and long-term plasticity together control the induction of synaptic depression and potentiation, we used numerical simulations and mathematical analysis of a calcium-based model, where pre- and postsynaptic activity induces calcium transients driving synaptic long-term plasticity. We found that the model implementing known synaptic short-term dynamics in the calcium transients can be successfully fitted to long-term plasticity data obtained in visual- and somatosensory cortex. Interestingly, the impact of spike-timing and firing rate changes on plasticity occurs in the prevalent firing rate range, which is different in both cortical areas considered here. Our findings suggest that short- and long-term plasticity are together tuned to adapt plasticity to area-specific activity statistics such as firing rates.

Minimal Circuit Model of Reward Prediction Error Computations and Effects of Nicotinic Modulations.

Dopamine (DA) neurons in the ventral tegmental area (VTA) are thought to encode reward prediction errors (RPE) by comparing actual and expected rewards. In recent years, much work has been done to identify how the brain uses and computes this signal. While several lines of evidence suggest the interplay of the DA and the inhibitory interneurons in the VTA implements the RPE computation, it still remains unclear how the DA neurons learn key quantities, for example the amplitude and the timing of primary rewards during conditioning tasks. Furthermore, endogenous acetylcholine and exogenous nicotine, also likely affect these computations by acting on both VTA DA and GABA (γ -aminobutyric acid) neurons via nicotinic-acetylcholine receptors (nAChRs). To explore the potential circuit-level mechanisms for RPE computations during classical-conditioning tasks, we developed a minimal computational model of the VTA circuitry. The model was designed to account for several reward-related properties of VTA afferents and recent findings on VTA GABA neuron dynamics during conditioning. With our minimal model, we showed that the RPE can be learned by a two-speed process computing reward timing and magnitude. By including models of nAChR-mediated currents in the VTA DA-GABA circuit, we showed that nicotine should reduce the acetylcholine action on the VTA GABA neurons by receptor desensitization and potentially boost DA responses to reward-related signals in a non-trivial manner. Together, our results delineate the mechanisms by which RPE are computed in the brain, and suggest a hypothesis on nicotine-mediated effects on reward-related perception and decision-making.