Search

Neurons are very complicated computational devices, incorporating numerous non-linear processes, particularly in their dendrites. Biophysical models capture these processes directly by explicitly modelling physiological variables, such as ion channels, current flow, membrane capacitance, etc. However, another option for capturing the complexities of real neural computation is to use cascade models, which treat individual neurons as a cascade of linear and non-linear operations, akin to a multi-layer artificial neural network. Recent research has shown that cascade models can capture single-cell computation well, but there are still a number of sub-cellular, regenerative dendritic phenomena that they cannot capture, such as the interaction between sodium, calcium, and NMDA spikes in different compartments. Here, we propose that it is possible to capture these additional phenomena usingparallel, recurrentcascade models, wherein an individual neuron is modelled as a cascade of parallel linear and non-linear operations that can be connected recurrently, akin to a multi-layer, recurrent, artificial neural network. Given their tractable mathematical structure, we show that neuron models expressed in terms of parallel recurrent cascades can themselves be integrated into multi-layered artificial neural networks and trained to perform complex tasks. We go on to discuss potential implications and uses of these models for artificial intelligence. Overall, we argue that parallel, recurrent cascade models provide an important, unifying tool for capturing single-cell computation and exploring the algorithmic implications of physiological phenomena.

New neurons are continuously generated in the subgranular zone of the dentate gyrus throughout adulthood. These new neurons gradually integrate into hippocampal circuits, forming new naive synapses. Viewed from this perspective, these new neurons may represent a significant source of “wiring” noise in hippocampal networks. In machine learning, such noise injection is commonly used as a regularization technique. Regularization techniques help prevent overfitting training data and allow models to generalize learning to new, unseen data. Using a computational modeling approach, here we ask whether a neurogenesis-like process similarly acts as a regularizer, facilitating generalization in a category learning task. In a convolutional neural network (CNN) trained on the CIFAR-10 object recognition dataset, we modeled neurogenesis as a replacement/turnover mechanism, where weights for a randomly chosen small subset of hidden layer neurons were reinitialized to new values as the model learned to categorize 10 different classes of objects. We found that neurogenesis enhanced generalization on unseen test data compared to networks with no neurogenesis. Moreover, neurogenic networks either outperformed or performed similarly to networks with conventional noise injection (i.e., dropout, weight decay, and neural noise). These results suggest that neurogenesis can enhance generalization in hippocampal learning through noise injection, expanding on the roles that neurogenesis may have in cognition.

The connection patterns of neural circuits in the brain form a complex network. Collective signalling within the network manifests as patterned neural activity and is thought to support human cognition and adaptive behaviour. Recent technological advances permit macroscale reconstructions of biological brain networks. These maps, termed connectomes, display multiple non-random architectural features, including heavy-tailed degree distributions, segregated communities and a densely interconnected core. Yet, how computation and functional specialization emerge from network architecture remains unknown. Here we reconstruct human brain connectomes using in vivo diffusion-weighted imaging and use reservoir computing to implement connectomes as artificial neural networks. We then train these neuromorphic networks to learn a memory-encoding task. We show that biologically realistic neural architectures perform best when they display critical dynamics. We find that performance is driven by network topology and that the modular organization of intrinsic networks is computationally relevant. We observe a prominent interaction between network structure and dynamics throughout, such that the same underlying architecture can support a wide range of memory capacity values as well as different functions (encoding or decoding), depending on the dynamical regime the network is in. This work opens new opportunities to discover how the network organization of the brain optimizes cognitive capacity. The relationship between brain organization, connectivity and computation is not well understood. The authors construct neuromorphic artificial neural networks endowed with biological connection patterns derived from diffusion-weighted imaging. The neuromorphic networks are trained to perform a memory task, revealing an interaction between network structure and dynamics.

Estimating value functions is a core component of reinforcement learning algorithms. Temporal difference (TD) learning algorithms use bootstrapping, i.e. they update the value function toward a learning target using value estimates at subsequent time-steps. Alternatively, the value function can be updated toward a learning target constructed by separately predicting successor features (SF)--a policy-dependent model--and linearly combining them with instantaneous rewards. We focus on bootstrapping targets used when estimating value functions, and propose a new backup target, the $\eta$-return mixture, which implicitly combines value-predictive knowledge (used by TD methods) with (successor) feature-predictive knowledge--with a parameter $\eta$ capturing how much to rely on each. We illustrate that incorporating predictive knowledge through an $\eta\gamma$-discounted SF model makes more efficient use of sampled experience, compared to either extreme, i.e. bootstrapping entirely on the value function estimate, or bootstrapping on the product of separately estimated successor features and instantaneous reward models. We empirically show this approach leads to faster policy evaluation and better control performance, for tabular and nonlinear function approximations, indicating scalability and generality., 18 pages, 6 figures, 2 tables. Preprint. Accepted by AAAI-22

Neurons can carry information with both the synchrony and rate of their spikes. However, it is unknown whether distinct subtypes of neurons are more sensitive to information carried by synchrony versus rate, or vice versa. Here, we address this question using patterned optical stimulation in slices of somatosensory cortex from mouse lines labelling fast-spiking (FS) and regular-spiking (RS) interneurons. We used optical stimulation in layer 2/3 to encode a 1-bit signal using either the synchrony or rate of activity. We then examined the mutual information between this signal and the interneuron responses. We found that for a synchrony encoding, FS interneurons carried more information in the first five milliseconds, while both interneuron subtypes carried more information than excitatory neurons in later responses. For a rate encoding, we found that RS interneurons carried more information after several milliseconds. These data demonstrate that distinct interneuron subtypes in the neocortex have distinct sensitivities to synchrony versus rate codes., In order to address whether distinct subtypes of neurons are more sensitive to information carried by synchrony versus rate, Prince et al. used optical stimulation in slices of somatosensory cortex from mouse lines labelling fast-spiking (FS) and regular-spiking (RS) interneurons. They demonstrated that FS and RS interneurons had differential sensitivity to changes in synchrony and rate, which advances our understanding of neural processing in the neocortex.

The representation of “what happened when” is central to encoding episodic and working memories. Recently discovered hippocampal time cells are theorized to provide the neural substrate for such representations by forming distinct sequences that both encode time elapsed and sensory content. However, little work has directly addressed to what extent cognitive demands and temporal structure of experimental tasks affect the emergence and informativeness of these temporal representations. Here, we trained deep reinforcement learning (DRL) agents on a simulated trial-unique nonmatch-to-location (TUNL) task, and analyzed the activities of artificial recurrent units using neuroscience-based methods. We show that, after training, representations resembling both time cells and ramping cells (whose activity increases or decreases monotonically over time) simultaneously emerged in the same population of recurrent units. Furthermore, with simulated variations of the TUNL task that controlled for (1) memory demands during the delay period and (2) the temporal structure of the episodes, we show that memory demands are necessary for the time cells to encode information about the sensory stimuli, while the temporal structure of the task only affected the encoding of “what” and “when” by time cells minimally. Our findings help to reconcile current discrepancies regarding the involvement of time cells in memory-encoding by providing a normative framework. Our modelling results also provide concrete experimental predictions for future studies.

Neurons in the dorsal visual pathway of the mammalian brain are selective for motion stimuli, with the complexity of stimulus representations increasing along the hierarchy. This progression is similar to that of the ventral visual pathway, which is well characterized by artificial neural networks (ANNs) optimized for object recognition. In contrast, there are no image-computable models of the dorsal stream with comparable explanatory power. We hypothesized that the properties of dorsal stream neurons could be explained by a simple learning objective: the need for an organism to orient itself during self-motion. To test this hypothesis, we trained a 3D ResNet to predict an agent’s self-motion parameters from visual stimuli in a simulated environment. We found that the responses in this network accounted well for the selectivity of neurons in a large database of single-neuron recordings from the dorsal visual stream of non-human primates. In contrast, ANNs trained on an action recognition dataset through supervised or self-supervised learning could not explain responses in the dorsal stream, despite also being trained on naturalistic videos with moving objects. These results demonstrate that an ecologically relevant cost function can account for dorsal stream properties in the primate brain.

The visual system of mammals is comprised of parallel, hierarchical specialized pathways. Different pathways are specialized in so far as they use representations that are more suitable for supporting specific downstream behaviours. In particular, the clearest example is the specialization of the ventral (“what”) and dorsal (“where”) pathways of the visual cortex. These two pathways support behaviours related to visual recognition and movement, respectively. To-date, deep neural networks have mostly been used as models of the ventral, recognition pathway. However, it is unknown whether both pathways can be modelled with a single deep ANN. Here, we ask whether a single model with a single loss function can capture the properties of both the ventral and the dorsal pathways. We explore this question using data from mice, who like other mammals, have specialized pathways that appear to support recognition and movement behaviours. We show that when we train a deep neural network architecture with two parallel pathways using a self-supervised predictive loss function, we can outperform other models in fitting mouse visual cortex. Moreover, we can model both the dorsal and ventral pathways. These results demonstrate that a self-supervised predictive learning approach applied to parallel pathway architectures can account for some of the functional specialization seen in mammalian visual systems.

What is the purpose of dreaming? Many scientists have postulated a role for dreaming in learning, often with the aim of improving generative models. In this issue of Patterns, Erik Hoel proposes a novel hypothesis, namely, that dreaming provides a means to reduce overfitting. This hypothesis is interesting both for neuroscience and for the development of new machine-learning systems.

Scientists have long conjectured that the neocortex learns the structure of the environment in a predictive, hierarchical manner. According to this conjecture, expected, predictable features are differentiated from unexpected ones by comparing bottom-up and top-down streams of information. It is theorized that the neocortex then changes the representation of incoming stimuli, guided by differences in the responses to expected and unexpected events. In line with this conjecture, different responses to expected and unexpected sensory features have been observed in spiking and somatic calcium events. However, it remains unknown whether these unexpected event signals occur in the distal apical dendrites where many top-down signals are received, and whether these signals govern subsequent changes in the brain’s stimulus representations. Here, we show that both somata and distal apical dendrites of cortical pyramidal neurons exhibit distinct unexpected event signals that systematically change over days. These findings were obtained by tracking the responses of individual somata and dendritic branches of layer 2/3 and layer 5 pyramidal neurons over multiple days in primary visual cortex of awake, behaving mice using two-photon calcium imaging. Many neurons in both layers 2/3 and 5 showed large differences between their responses to expected and unexpected events. Interestingly, these responses evolved in opposite directions in the somata and distal apical dendrites. These differences between the somata and distal apical dendrites may be important for hierarchical computation, given that these two compartments tend to receive bottom-up and top-down information, respectively.

The connection patterns of neural circuits in the brain form a complex network. Collective signaling within the network manifests as patterned neural activity, and is thought to support human cognition and adaptive behavior. Recent technological advances permit macro-scale reconstructions of biological brain networks. These maps, termed connectomes, display multiple non-random architectural features, including heavy-tailed degree distributions, segregated communities and a densely interconnected core. Yet, how computation and functional specialization emerge from network architecture remains unknown. Here we reconstruct human brain connectomes usingin vivodiffusion-weighted imaging, and use reservoir computing to implement these connectomes as artificial neural networks. We then train these neuromorphic networks to learn a cognitive task. We show that biologically realistic neural architectures perform optimally when they display critical dynamics. We find that performance is driven by network topology, and that the modular organization of large-scale functional systems is computationally relevant. Throughout, we observe a prominent interaction between network structure and dynamics, such that the same underlying architecture can support a wide range of learning capacities across dynamical regimes. This work opens new opportunities to discover how the network organization of the brain optimizes cognitive capacity, conceptually bridging neuroscience and artificial intelligence.

The units in artificial neural networks (ANNs) can be thought of as abstractions of biological neurons, and ANNs are increasingly used in neuroscience research. However, there are many important differences between ANN units and real neurons. One of the most notable is the absence of Dale’s principle, which ensures that biological neurons are either exclusively excitatory or inhibitory. Dale’s principle is typically left out of ANNs because its inclusion impairs learning. This is problematic, because one of the great advantages of ANNs for neuroscience research is their ability to learn complicated, realistic tasks. Here, by taking inspiration from feedforward inhibitory interneurons in the brain we show that we can develop ANNs with separate populations of excitatory and inhibitory units that learn just as well as standard ANNs. We call these networks Dale’s ANNs (DANNs). We present two insights that enable DANNs to learn well: (1) DANNs are related to normalization schemes, and can be initialized such that the inhibition centres and standardizes the excitatory activity, (2) updates to inhibitory neuron parameters should be scaled using corrections based on the Fisher Information matrix. These results demonstrate how ANNs that respect Dale’s principle can be built without sacrificing learning performance, which is important for future work using ANNs as models of the brain. The results may also have interesting implications for how inhibitory plasticity in the real brain operates.

Recently, deep learning has unlocked unprecedented success in various domains, especially using images, text, and speech. However, deep learning is only beneficial if the data have nonlinear relationships and if they are exploitable at available sample sizes. We systematically profiled the performance of deep, kernel, and linear models as a function of sample size on UKBiobank brain images against established machine learning references. On MNIST and Zalando Fashion, prediction accuracy consistently improves when escalating from linear models to shallow-nonlinear models, and further improves with deep-nonlinear models. In contrast, using structural or functional brain scans, simple linear models perform on par with more complex, highly parameterized models in age/sex prediction across increasing sample sizes. In sum, linear models keep improving as the sample size approaches ~10,000 subjects. Yet, nonlinearities for predicting common phenotypes from typical brain scans remain largely inaccessible to the examined kernel and deep learning methods., Schulz et al. systematically benchmark performance scaling with increasingly sophisticated prediction algorithms and with increasing sample size in reference machine-learning and biomedical datasets. Complicated nonlinear intervariable relationships remain largely inaccessible for predicting key phenotypes from typical brain scans.

Behavioral flexibility is important in a changing environment. Previous research suggests that systems consolidation, a long-term poststorage process that alters memory traces, may reduce behavioral flexibility. However, exactly how systems consolidation affects flexibility is unknown. Here, we tested how systems consolidation affects: (1) flexibility in response to value changes and (2) flexibility in response to changes in the optimal sequence of actions. Mice were trained to obtain food rewards in a Y-maze by switching nose pokes between three arms. During initial training, all arms were rewarded and mice simply had to switch arms in order to maximize rewards. Then, after either a 1 or 28 d delay, we either devalued one arm, or we reinforced a specific sequence of pokes. We found that after a 1 d delay mice adapted relatively easily to the changes. In contrast, mice given a 28 d delay struggled to adapt, especially for changes to the optimal sequence of actions. Immediate early gene imaging suggested that the 28 d mice were less reliant on their hippocampus and more reliant on their medial prefrontal cortex. These data suggest that systems consolidation reduces behavioral flexibility, particularly for changes to the optimal sequence of actions.

Sensory information–driven spikes are synchronized across cortical layers by distinct subtypes of interneurons., Synchronization of precise spike times across multiple neurons carries information about sensory stimuli. Inhibitory interneurons are suggested to promote this synchronization, but it is unclear whether distinct interneuron subtypes provide different contributions. To test this, we examined single-unit recordings from barrel cortex in vivo and used optogenetics to determine the contribution of parvalbumin (PV)– and somatostatin (SST)–positive interneurons to the synchronization of spike times across cortical layers. We found that PV interneurons preferentially promote the synchronization of spike times when instantaneous firing rates are low (12 Hz). Furthermore, using a computational model, we demonstrate that these effects can be explained by PV and SST interneurons having preferential contributions to feedforward and feedback inhibition, respectively. Our findings demonstrate that distinct subtypes of inhibitory interneurons have frequency-selective roles in the spatiotemporal synchronization of precise spike times.

The predominant focus in the neurobiological study of memory has been on remembering (persistence). However, recent studies have considered the neurobiology of forgetting (transience). Here we draw parallels between neurobiological and computational mechanisms underlying transience. We propose that it is the interaction between persistence and transience that allows for intelligent decision-making in dynamic, noisy environments. Specifically, we argue that transience (1) enhances flexibility, by reducing the influence of outdated information on memory-guided decision-making, and (2) prevents overfitting to specific past events, thereby promoting generalization. According to this view, the goal of memory is not the transmission of information through time, per se. Rather, the goal of memory is to optimize decision-making. As such, transience is as important as persistence in mnemonic systems.

Background Abnormal accumulation of amyloid β1–42 oligomers (AβO1–42), a hallmark of Alzheimer’s disease, impairs hippocampal theta-nested gamma oscillations and long-term potentiation (LTP) that are believed to underlie learning and memory. Parvalbumin-positive (PV) and somatostatin-positive (SST) interneurons are critically involved in theta-nested gamma oscillogenesis and LTP induction. However, how AβO1–42 affects PV and SST interneuron circuits is unclear. Through optogenetic manipulation of PV and SST interneurons and computational modeling of the hippocampal neural circuits, we dissected the contributions of PV and SST interneuron circuit dysfunctions on AβO1–42-induced impairments of hippocampal theta-nested gamma oscillations and oscillation-induced LTP. Results Targeted whole-cell patch-clamp recordings and optogenetic manipulations of PV and SST interneurons during in vivo-like, optogenetically induced theta-nested gamma oscillations in vitro revealed that AβO1–42 causes synapse-specific dysfunction in PV and SST interneurons. AβO1–42 selectively disrupted CA1 pyramidal cells (PC)-to-PV interneuron and PV-to-PC synapses to impair theta-nested gamma oscillogenesis. In contrast, while having no effect on PC-to-SST or SST-to-PC synapses, AβO1–42 selectively disrupted SST interneuron-mediated disinhibition to CA1 PC to impair theta-nested gamma oscillation-induced spike timing-dependent LTP (tLTP). Such AβO1–42-induced impairments of gamma oscillogenesis and oscillation-induced tLTP were fully restored by optogenetic activation of PV and SST interneurons, respectively, further supporting synapse-specific dysfunctions in PV and SST interneurons. Finally, computational modeling of hippocampal neural circuits including CA1 PC, PV, and SST interneurons confirmed the experimental observations and further revealed distinct functional roles of PV and SST interneurons in theta-nested gamma oscillations and tLTP induction. Conclusions Our results reveal that AβO1–42 causes synapse-specific dysfunctions in PV and SST interneurons and that optogenetic modulations of these interneurons present potential therapeutic targets for restoring hippocampal network oscillations and synaptic plasticity impairments in Alzheimer’s disease.

Systems neuroscience seeks explanations for how the brain implements a wide variety of perceptual, cognitive and motor tasks. Conversely, artificial intelligence attempts to design computational systems based on the tasks they will have to solve. In the case of artificial neural networks, the three components specified by design are the objective functions, the learning rules, and architectures. With the growing success of deep learning, which utilizes brain-inspired architectures, these three designed components have increasingly become central to how we model, engineer and optimize complex artificial learning systems. Here we argue that a greater focus on these components would also benefit systems neuroscience. We give examples of how this optimization-based framework can drive theoretical and experimental progress in neuroscience. We contend that this principled perspective on systems neuroscience will help to generate more rapid progress.

In recent years, deep learning has unlocked unprecedented success in various domains, especially in image, text, and speech processing. These breakthroughs may hold promise for neuroscience and especially for brain-imaging investigators who start to analyze thousands of participants. However, deep learning is only beneficial if the data have nonlinear relationships and if they are exploitable at currently available sample sizes.We systematically profiled the performance of deep models, kernel models, and linear models as a function of sample size on UK Biobank brain images against established machine learning references. On MNIST and Zalando Fashion, prediction accuracy consistently improved when escalating from linear models to shallow-nonlinear models, and further improved when switching to deep-nonlinear models. The more observations were available for model training, the greater the performance gain we saw. In contrast, using structural or functional brain scans, simple linear models performed on par with more complex, highly parameterized models in age/sex prediction across increasing samplesizes. In fact, linear models kept improving as the sample size approached ~10,000 participants. Our results indicate that the increase in performance of linear models with additional data does not saturate at the limit of current feasibility. Yet, nonlinearities of common brain scans remain largely inaccessible to both kernel and deep learning methods at any axemined scale

Synchronization of precise spike-times across multiple neurons carries information about sensory stimuli. Inhibitory interneurons are suggested to promote this synchronization, but it is unclear whether distinct interneuron subtypes provide different contributions. To test this, we examined single-unit recordings from barrel cortex in vivo and used optogenetics to determine the contribution of two classes of inhibitory interneurons: parvalbumin (PV)- and somatostatin (SST)-positive interneurons to spike-timing synchronization across cortical layers. We found that PV interneurons preferentially promote the synchronization of spike-times when instantaneous firing-rates are low (12 Hz). Furthermore, using a computational model, we demonstrate that these effects can be explained by PV and SST interneurons having preferential contribution to feedforward and feedback inhibition, respectively. Our findings demonstrate that distinct subtypes of inhibitory interneurons have frequency-selective roles in spatio-temporal synchronization of precise spike-times.

Anxiety is an adaptive response to potentially threatening situations. Exaggerated and uncontrolled anxiety responses become maladaptive and lead to anxiety disorders. Anxiety is shaped by a network of forebrain structures, including the hippocampus, septum, and prefrontal cortex. In particular, neural inputs arising from the ventral hippocampus (vHPC) to the lateral septum (LS) and medial prefrontal cortex (mPFC) are thought to serve as principal components of the anxiety circuit. However, the role of vHPC-to-LS and vHPC-to-mPFC signals in anxiety is unclear, as no study has directly compared their behavioral contribution at circuit level. We targeted LS-projecting vHPC cells and mPFC-projecting vHPC cells by injecting the retrogradely propagating canine adenovirus encoding Cre recombinase into the LS or mPFC, and injecting a Cre-responsive AAV (AAV8-hSyn-FLEX-hM3D or hM4D) into the vHPC. Consequences of manipulating these neurons were examined in well-established tests of anxiety. Chemogenetic manipulation of LS-projecting vHPC cells led to bidirectional changes in anxiety: activation of LS-projecting vHPC cells decreased anxiety whereas inhibition of these cells produced opposite anxiety-promoting effects. The observed anxiety-reducing function of LS-projecting cells was in contrast with the function of mPFC-projecting cells, which promoted anxiety. In addition, double retrograde tracing demonstrated that LS- and mPFC-projecting cells represent two largely anatomically distinct cell groups. Altogether, our findings suggest that the vHPC houses discrete populations of cells that either promote or suppress anxiety through differences in their projection targets. Disruption of the intricate balance in the activity of these two neuron populations may drive inappropriate behavioral responses seen in anxiety disorders.

Forgetting involves the loss of information over time; however, we know little about what form this information loss takes. Do memories become less precise over time, or do they instead become less accessible? Here we assessed memory for word-location associations across four days, testing whether forgetting involves losses in precision versus accessibility and whether such losses are modulated by learning a generalizable pattern. We show that forgetting involves losses in memory accessibility with no changes in memory precision. When participants learned a set of related word-location associations that conformed to a general pattern, we saw a strong trade-off; accessibility was enhanced, whereas precision was reduced. However, this trade-off did not appear to be modulated by time or confer a long-term increase in the total amount of information maintained in memory. Our results place theoretical constraints on how models of forgetting and generalization account for time-dependent memory processes. PROTOCOL REGISTRATION: The stage 1 protocol for this Registered Report was accepted in principle on 4 June 2019. The protocol, as accepted by the journal, can be found at https://doi.org/10.6084/m9.figshare.c.4368464.v1 .

Classic theories of reinforcement learning and neuromodulation rely on reward prediction errors. A new machine learning technique relies on neuromodulatory signals that are optimized for specific tasks, which may lead to better AI and better explanations of neuroscience data.

SummaryThe ability to generate action potentials (spikes) in response to synaptic input determines whether a neuron participates in information processing. How a developing neuron becomes an active participant in a circuit or whether this process is activity dependent is not known, especially as spike-dependent plasticity mechanisms would not be available to non-spiking neurons. Here we use the optic tectum of awake Xenopus laevis tadpoles to determine how a neuron becomes able to generate sensory-driven spikes in vivo. At the onset of vision, many tectal neurons do not exhibit visual spiking behavior, despite being intrinsically excitable and receiving visuotopically organized synaptic inputs. However, a brief period of visual stimulation can drive these neurons to start generating stimulus-driven spikes. This conversion relies upon a selective increase in glutamatergic input and requires depolarizing GABAergic transmission and NMDA receptor activation. This permissive form of experience-dependent plasticity enables a neuron to start contributing to circuit function.

Following learning, increased coupling between spindle oscillations in the medial prefrontal cortex (mPFC) and ripple oscillations in the hippocampus is thought to underlie memory consolidation. However, whether learning-induced increases in ripple-spindle coupling are necessary for successful memory consolidation has not been tested directly. In order to decouple ripple-spindle oscillations, here we chemogenetically inhibited parvalbumin-positive (PV+) interneurons, since their activity is important for regulating the timing of spiking activity during oscillations. We found that contextual fear conditioning increased ripple-spindle coupling in mice. However, inhibition of PV+ cells in either CA1 or mPFC eliminated this learning-induced increase in ripple-spindle coupling without affecting ripple or spindle incidence. Consistent with the hypothesized importance of ripple-spindle coupling in memory consolidation, post-training inhibition of PV+ cells disrupted contextual fear memory consolidation. These results indicate that successful memory consolidation requires coherent hippocampal-neocortical communication mediated by PV+ cells., eLife digest Sleep contributes to the strengthening of memories. During non-dreaming sleep, neurons in regions of the brain that form and store memories – such as the hippocampus and prefrontal cortex – fire in rhythmic waves. The neurons in the hippocampus tend to fire during a wave that repeats up to 250 times per second, called sharp-wave ripples. Meanwhile, in the prefrontal cortex, the neurons tend to fire during a lower frequency wave that repeats 12 to 15 times per second, called spindles. During sleep and quiet wakefulness, hippocampal ripples often synchronize with prefrontal spindles; that is, both waves tend to occur at approximately the same time. Many neuroscientists think this allows the brain regions to better communicate with one another, which in turn should help the brain to strengthen memories. Consistent with this possibility, rodents that learn a new task show more synchrony between ripples and spindles afterwards. But no one had actually tested whether this increase in ripple-spindle synchrony does strengthen the rodent’s memory of the task. It was also unclear how the brain achieves such an increase. Xia et al. suspected that this process involved a group of inhibitory brain cells called parvalbumin-positive interneurons. These cells act like timekeepers, and help to synchronize the firing of groups of neurons. Xia et al. now show that training mice to associate an environment with a mild electric shock made it more likely that the animals would show ripple-spindle synchrony. Yet, inhibiting the activity of parvalbumin-positive interneurons in either the hippocampus or prefrontal cortex blocked this effect. It also prevented sleep from strengthening the animals’ memory of the link between the environment and the shock. Patients with Alzheimer’s disease have fewer parvalbumin-positive interneurons. By showing that these neurons help strengthen new memories, these findings may explain why losing them can impair memory. Restoring or replacing interneuron activity could be a promising therapeutic avenue to explore.

Symptoms of schizophrenia may arise from a failure of cortical circuits to filter-out irrelevant inputs. Schizophrenia has also been linked to disruptions in cortical inhibitory interneurons, consistent with the possibility that in the normally functioning brain, these cells are in some part responsible for determining which sensory inputs are relevant versus irrelevant. Here, we develop a neural network model that demonstrates how the cortex may learn to ignore irrelevant inputs through plasticity processes affecting inhibition. The model is based on the proposal that the amount of excitatory output from a cortical circuit encodes the expected magnitude of reward or punishment (“relevance”), which can be trained using a temporal difference learning mechanism acting on feedforward inputs to inhibitory interneurons. In the model, irrelevant and blocked stimuli drive lower levels of excitatory activity compared with novel and relevant stimuli, and this difference in activity levels is lost following disruptions to inhibitory units. When excitatory units are connected to a competitive-learning output layer with a threshold, the relevance code can be shown to “gate” both learning and behavioral responses to irrelevant stimuli. Accordingly, the combined network is capable of recapitulating published experimental data linking inhibition in frontal cortex with fear learning and expression. Finally, the model demonstrates how relevance learning can take place in parallel with other types of learning, through plasticity rules involving inhibitory and excitatory components, respectively. Altogether, this work offers a theory of how the cortex learns to selectively inhibit inputs, providing insight into how relevance-assignment problems may emerge in schizophrenia., Author summary Individuals with schizophrenia have difficulty ignoring ideas and experiences that most people would treat as unimportant. There is evidence that this may be due to changes in neuronal inhibition, suggesting that inhibitory neurons may be involved in learning to ignore irrelevant inputs. By developing a computational model that learns relevance and irrelevance through changes in the strength of feedforward inhibition, we are able to simulate many specific effects of inhibitory neuron dysfunction on behavior. We also show two computational advantages to this mechanism: (1) if relevance is signaled by the level of excitatory activity, then downstream circuits can easily avoid learning from irrelevant stimuli, (2) relevance learning can occur simultaneously with other types of learning. The model therefore offers insight into the relationships between neural inhibition and behavior, including symptoms of schizophrenia.

Symptoms of schizophrenia may arise from a failure of cortical circuits to filter-out irrelevant inputs. Schizophrenia has also been linked to disruptions to cortical inhibitory interneurons, consistent with the possibility that in the normally functioning brain, these cells are in some part responsible for determining which inputs are relevant and which irrelevant. Here, we develop an abstract but biologically plausible neural network model that demonstrates how the cortex may learn to ignore irrelevant inputs through plasticity processes affecting inhibition. The model is based on the proposal that the amount of excitatory output from a cortical circuit encodes expected magnitude of reward or punishment (”relevance”), which can be trained using a temporal difference learning mechanism acting on feed-forward inputs to inhibitory interneurons. The model exhibits learned irrelevance and blocking, which become impaired following disruptions to inhibitory units. When excitatory units are connected to a competitive-learning output layer, the relevance code is capable of modulating learning and activity. Accordingly, the combined network is capable of recapitulating published experimental data linking inhibition in frontal cortex with fear learning and expression. Finally, the model demonstrates how relevance learning can take place in parallel with other types of learning, through plasticity rules involving inhibitory and excitatory components respectively. Altogether, this work offers a theory of how the cortex learns to selectively inhibit inputs, providing insight into how relevance-assignment problems may emerge in schizophrenia.

During the development of sensory systems, receptive fields are modified by stimuli in the environment. This is thought to rely on learning algorithms that are sensitive to correlations in spike timing between cells, but the manner in which developing circuits selectively exploit correlations that are related to sensory inputs is unknown. We recorded from neurons in the developing optic tectum of Xenopus laevis and found that repeated presentation of moving visual stimuli induced receptive field changes that reflected the properties of the stimuli and that this form of learning was disrupted when GABAergic transmission was blocked. Consistent with a role for spike timing-dependent mechanisms, GABA blockade altered spike-timing patterns in the tectum and increased correlations between cells that would affect plasticity at intratectal synapses. This is a previously unknown role for GABAergic signals in development and highlights the importance of regulating the statistics of spiking activity for learning.

Memory stabilization following encoding (synaptic consolidation) or memory reactivation (reconsolidation) requires gene expression and protein synthesis (Dudai and Eisenberg, 2004; Tronson and Taylor, 2007; Nader and Einarsson, 2010; Alberini, 2011). Although consolidation and reconsolidation may be mediated by distinct molecular mechanisms (Lee et al., 2004), disrupting the function of the transcription factor CREB impairs both processes (Kida et al., 2002; Mamiya et al., 2009). Phosphorylation of CREB at Ser133 recruits CREB binding protein (CBP)/p300 coactivators to activate transcription (Chrivia et al., 1993; Parker et al., 1996). In addition to this well known mechanism, CREB regulated transcription coactivators (CRTCs), previously called transducers of regulated CREB (TORC) activity, stimulate CREB-mediated transcription, even in the absence of CREB phosphorylation. Recently, CRTC1 has been shown to undergo activity-dependent trafficking from synapses and dendrites to the nucleus in excitatory hippocampal neurons (Ch'ng et al., 2012). Despite being a powerful and specific coactivator of CREB, the role of CRTC in memory is virtually unexplored. To examine the effects of increasing CRTC levels, we used viral vectors to locally and acutely increase CRTC1 in the dorsal hippocampus dentate gyrus region of mice before training or memory reactivation in context fear conditioning. Overexpressing CRTC1 enhanced both memory consolidation and reconsolidation; CRTC1-mediated memory facilitation was context specific (did not generalize to nontrained context) and long lasting (observed after virally expressed CRTC1 dissipated). CREB overexpression produced strikingly similar effects. Therefore, increasing CRTC1 or CREB function is sufficient to enhance the strength of new, as well as established reactivated, memories without compromising memory quality.

The retinotectal pathway of Xenopus laevis is a well-established experimental model for studying activity-dependent processes during visual system development. Such processes can be guided by stimulus-evoked activity patterns, which depend on the refractive characteristics of the eye. Previous work has shown that many animals are hyperopic at early developmental stages due to immature refractive properties. Whether this is also the case for Xenopus laevis is unknown. Here, we measure the focal length of the lens and the size of the eye of embryos at different stages and find that Xenopus laevis exhibits a similar shift from hyperopia to emmetropia. At early stages, immediately after innervation of the tectum by the optic nerve, Xenopus embryos are hyperopic. Soon afterwards the focal length of the lens decreases and the eye converges to a state of emmetropia. Despite being hyperopic we find that some visuospatial information is available to the young circuit. Calculations based on the optical properties of the eye show that even when the animals are hyperopic the blurred retinal image provides a crude spatial resolution. Furthermore, using whole-cell recordings in the optic tectum combined with visual stimulation through the intact eye, we show that tectal neurons in hyperopic embryos have spatially restricted glutamatergic receptive fields. Our data demonstrate that Xenopus laevis eyes undergo a process of developmental emmetropization, and suggest that despite an initial stage of suboptimal image formation there is potentially enough information to guide activity-dependent refinements of the retinotectal pathway from the onset of vision. © 2011 Wiley Periodicals, Inc. Develop Neurobiol 72: 507–519, 2012

Over the course of systems consolidation, there is a switch from a reliance on detailed episodic memories to generalized schematic memories. This switch is sometimes referred to as “memory transformation.” Here we demonstrate a previously unappreciated benefit of memory transformation, namely, its ability to enhance reinforcement learning in a dynamic environment. We developed a neural network that is trained to find rewards in a foraging task where reward locations are continuously changing. The network can use memories for specific locations (episodic memories) and statistical patterns of locations (schematic memories) to guide its search. We find that switching from an episodic to a schematic strategy over time leads to enhanced performance due to the tendency for the reward location to be highly correlated with itself in the short-term, but regress to a stable distribution in the long-term. We also show that the statistics of the environment determine the optimal utilization of both types of memory. Our work recasts the theoretical question of why memory transformation occurs, shifting the focus from the avoidance of memory interference toward the enhancement of reinforcement learning across multiple timescales.SIGNIFICANCE STATEMENTAs time passes, memories transform from a highly detailed state to a more gist-like state, in a process called “memory transformation.” Theories of memory transformation speak to its advantages in terms of reducing memory interference, increasing memory robustness, and building models of the environment. However, the role of memory transformation from the perspective of an agent that continuously acts and receives reward in its environment is not well explored. In this work, we demonstrate a view of memory transformation that defines it as a way of optimizing behavior across multiple timescales.

Deep learning has led to significant advances in artificial intelligence, in part, by adopting strategies motivated by neurophysiology. However, it is unclear whether deep learning could occur in the real brain. Here, we show that a deep learning algorithm that utilizes multi-compartment neurons might help us to understand how the brain optimizes cost functions. Like neocortical pyramidal neurons, neurons in our model receive sensory information and higher-order feedback in electrotonically segregated compartments. Thanks to this segregation, the neurons in different layers of the network can coordinate synaptic weight updates. As a result, the network can learn to categorize images better than a single layer network. Furthermore, we show that our algorithm takes advantage of multilayer architectures to identify useful representations---the hallmark of deep learning. This work demonstrates that deep learning can be achieved using segregated dendritic compartments, which may help to explain the dendritic morphology of neocortical pyramidal neurons., Comment: 41 pages, 11 figures

SummaryMemories are thought to be sparsely encoded in neuronal networks, but little is known about why a given neuron is recruited or allocated to a particular memory trace. Previous research shows that in the lateral amygdala (LA), neurons with increased CREB are selectively recruited to a fear memory trace. CREB is a ubiquitous transcription factor implicated in many cellular processes. Which process mediates neuronal memory allocation? One hypothesis is that CREB increases neuronal excitability to bias neuronal recruitment, although this has not been shown experimentally. Here we use several methods to increase neuronal excitability and show this both biases recruitment into the memory trace and enhances memory formation. Moreover, artificial activation of these neurons alone is a sufficient retrieval cue for fear memory expression, showing that these neurons are critical components of the memory trace. These results indicate that neuronal memory allocation is based on relative neuronal excitability immediately before training.Video Abstract

Spike-timing-dependent plasticity (STDP) is found in vivo in a variety of systems and species, but the first demonstrations of in vivo STDP were carried out in the optic tectum of Xenopus laevis embryos. Since then, the optic tectum has served as an excellent experimental model for studying STDP in sensory systems, allowing researchers to probe the developmental consequences of this form of synaptic plasticity during early development. In this review, we will describe what is known about the role of STDP in shaping feed-forward and recurrent circuits in the optic tectum with a focus on the functional implications for vision. We will discuss both the similarities and differences between the optic tectum and mammalian sensory systems that are relevant to STDP. Finally, we will highlight the unique properties of the embryonic tectum that make it an important system for researchers who are interested in how STDP contributes to activity-dependent development of sensory computations. © 2010 Richards, Aizenman and Akerman.

Children learn to name the objects they see by forming general associations between the words they hear and the images arriving at their retina. Discriminative neural network models can also be taught to classify objects, but to do so they require more information about how images pair with words (i.e. supervised data) than the brain seems to receive. We propose that the brain exploits unsupervised learning on raw sensory input to compensate for the scarcity of supervised data in its environment. Here we show that artificial neural networks which first develop a statistical model of the world in an unsupervised fashion are capable of learning good image-word pairings using dramatically less supervised data. This idea may help to explain how the brain learns sensorimotor problems for which there is little feedback available about the success of selected actions.