Closed-Loop tACS Delivered during Slow-Wave Sleep Reduces Retroactive Interference on a Paired-Associates Learning Task.

Previous studies have found a benefit of closed-loop transcranial alternating current stimulation (CL-tACS) matched to ongoing slow-wave oscillations (SWO) during sleep on memory consolidation for words in a paired associates task (PAT). Here, we examined the effects of CL-tACS in a retroactive interference PAT (ri-PAT) paradigm, where additional stimuli were presented to increase interference and reduce memory performance. Thirty-one participants were tested on a PAT before sleep, and CL-tACS was applied over the right and left DLPFC (F3 and F4) vs. mastoids for five cycles after detection of the onset of each discrete event of SWO during sleep. Participants were awoken the following morning, learned a new PAT list, and then were tested on the original list. There was a significant effect of stimulation condition (p = 0.04297; Cohen's d = 0.768), where verum stimulation resulted in reduced retroactive interference compared with sham and a significant interaction of encoding strength and stimulation condition (p = 0.03591). Planned simple effects testing within levels of encoding revealed a significant effect of stimulation only for low-encoders (p = 0.0066; Cohen's d = 1.075) but not high-encoders. We demonstrate here for the first time that CL-tACS during sleep can enhance the protective benefits on retroactive interference in participants who have lower encoding aptitude.

A domain-agnostic approach for characterization of lifelong learning systems.

Despite the advancement of machine learning techniques in recent years, state-of-the-art systems lack robustness to "real world" events, where the input distributions and tasks encountered by the deployed systems will not be limited to the original training context, and systems will instead need to adapt to novel distributions and tasks while deployed. This critical gap may be addressed through the development of "Lifelong Learning" systems that are capable of (1) Continuous Learning, (2) Transfer and Adaptation, and (3) Scalability. Unfortunately, efforts to improve these capabilities are typically treated as distinct areas of research that are assessed independently, without regard to the impact of each separate capability on other aspects of the system. We instead propose a holistic approach, using a suite of metrics and an evaluation framework to assess Lifelong Learning in a principled way that is agnostic to specific domains or system techniques. Through five case studies, we show that this suite of metrics can inform the development of varied and complex Lifelong Learning systems. We highlight how the proposed suite of metrics quantifies performance trade-offs present during Lifelong Learning system development - both the widely discussed Stability-Plasticity dilemma and the newly proposed relationship between Sample Efficient and Robust Learning. Further, we make recommendations for the formulation and use of metrics to guide the continuing development of Lifelong Learning systems and assess their progress in the future.

A model of bi-directional interactions between complementary learning systems for memory consolidation of sequential experiences.

The standard theory of memory consolidation posits a dual-store memory system: a fast-learning fast-decaying hippocampus that transfers memories to slow-learning long-term cortical storage. Hippocampal lesions interrupt this transfer, so recent memories are more likely to be lost than more remote memories. Existing models of memory consolidation that simulate this temporally graded retrograde amnesia operate only on static patterns or unitary variables as memories and study only one-way interaction from the hippocampus to the cortex. However, the mechanisms underlying the consolidation of episodes, which are sequential in nature and comprise multiple events, are not well-understood. The representation of learning for sequential experiences in the cortical-hippocampal network as a self-consistent dynamical system is not sufficiently addressed in prior models. Further, there is evidence for a bi-directional interaction between the two memory systems during offline periods, whereby the reactivation of waking neural patterns originating in the cortex triggers time-compressed sequential replays in the hippocampus, which in turn drive the consolidation of the pertinent sequence in the cortex. We have developed a computational model of memory encoding, consolidation, and recall for storing temporal sequences that explores the dynamics of this bi-directional interaction and time-compressed replays in four simulation experiments, providing novel insights into whether hippocampal learning needs to be suppressed for stable memory consolidation and into how new and old memories compete for limited replay opportunities during offline periods. The salience of experienced events, based on factors such as recency and frequency of use, is shown to have considerable impact on memory consolidation because it biases the relative probability that a particular event will be cued in the cortex during offline periods. In the presence of hippocampal learning during sleep, our model predicts that the fast-forgetting hippocampus can continually refresh the memory traces of a given episodic sequence if there are no competing experiences to be replayed.

Context meta-reinforcement learning via neuromodulation.

Meta-reinforcement learning (meta-RL) algorithms enable agents to adapt quickly to tasks from few samples in dynamic environments. Such a feat is achieved through dynamic representations in an agent's policy network (obtained via reasoning about task context, model parameter updates, or both). However, obtaining rich dynamic representations for fast adaptation beyond simple benchmark problems is challenging due to the burden placed on the policy network to accommodate different policies. This paper addresses the challenge by introducing neuromodulation as a modular component to augment a standard policy network that regulates neuronal activities in order to produce efficient dynamic representations for task adaptation. The proposed extension to the policy network is evaluated across multiple discrete and continuous control environments of increasing complexity. To prove the generality and benefits of the extension in meta-RL, the neuromodulated network was applied to two state-of-the-art meta-RL algorithms (CAVIA and PEARL). The result demonstrates that meta-RL augmented with neuromodulation produces significantly better result and richer dynamic representations in comparison to the baselines.

Deep Reinforcement Learning With Modulated Hebbian Plus Q-Network Architecture.

In this article, we consider a subclass of partially observable Markov decision process (POMDP) problems which we termed confounding POMDPs. In these types of POMDPs, temporal difference (TD)-based reinforcement learning (RL) algorithms struggle, as TD error cannot be easily derived from observations. We solve these types of problems using a new bio-inspired neural architecture that combines a modulated Hebbian network (MOHN) with deep Q-network (DQN), which we call modulated Hebbian plus Q-network architecture (MOHQA). The key idea is to use a Hebbian network with rarely correlated bio-inspired neural traces to bridge temporal delays between actions and rewards when confounding observations and sparse rewards result in inaccurate TD errors. In MOHQA, DQN learns low-level features and control, while the MOHN contributes to high-level decisions by associating rewards with past states and actions. Thus, the proposed architecture combines two modules with significantly different learning algorithms, a Hebbian associative network and a classical DQN pipeline, exploiting the advantages of both. Simulations on a set of POMDPs and on the Malmo environment show that the proposed algorithm improved DQN's results and even outperformed control tests with advantage-actor critic (A2C), quantile regression DQN with long short-term memory (QRDQN + LSTM), Monte Carlo policy gradient (REINFORCE), and aggregated memory for reinforcement learning (AMRL) algorithms on most difficult POMDPs with confounding stimuli and sparse rewards.

Brain connectivity alterations during sleep by closed-loop transcranial neurostimulation predict metamemory sensitivity.

Metamemory involves the ability to correctly judge the accuracy of our memories. The retrieval of memories can be improved using transcranial electrical stimulation (tES) during sleep, but evidence for improvements to metamemory sensitivity is limited. Applying tES can enhance sleep-dependent memory consolidation, which along with metamemory requires the coordination of activity across distributed neural systems, suggesting that examining functional connectivity is important for understanding these processes. Nevertheless, little research has examined how functional connectivity modulations relate to overnight changes in metamemory sensitivity. Here, we developed a closed-loop short-duration tES method, time-locked to up-states of ongoing slow-wave oscillations, to cue specific memory replays in humans. We measured electroencephalographic (EEG) coherence changes following stimulation pulses, and characterized network alterations with graph theoretic metrics. Using machine learning techniques, we show that pulsed tES elicited network changes in multiple frequency bands, including increased connectivity in the theta band and increased efficiency in the spindle band. Additionally, stimulation-induced changes in beta-band path length were predictive of overnight changes in metamemory sensitivity. These findings add new insights into the growing literature investigating increases in memory performance through brain stimulation during sleep, and highlight the importance of examining functional connectivity to explain its effects.

Neuromodulated attention and goal-driven perception in uncertain domains.

In uncertain domains, the goals are often unknown and need to be predicted by the organism or system. In this paper, contrastive Excitation Backprop (c-EB) was used in two goal-driven perception tasks - one with pairs of noisy MNIST digits and the other with a robot in an action-based attention scenario. The first task included attending to even, odd, low, and high digits, whereas the second task included action goals, such as "eat", "work-on-computer", "read", and "say-hi" that led to attention to objects associated with those actions. The system needed to increase attention to target items and decrease attention to distractor items and background noise. Because the valid goal was unknown, an online learning model based on the cholinergic and noradrenergic neuromodulatory systems was used to predict a noisy goal (expected uncertainty) and re-adapt when the goal changed (unexpected uncertainty). This neurobiologically plausible model demonstrates how neuromodulatory systems can predict goals in uncertain domains and how attentional mechanisms can enhance the perception for that goal.

Detecting Changes and Avoiding Catastrophic Forgetting in Dynamic Partially Observable Environments.

The ability of an agent to detect changes in an environment is key to successful adaptation. This ability involves at least two phases: learning a model of an environment, and detecting that a change is likely to have occurred when this model is no longer accurate. This task is particularly challenging in partially observable environments, such as those modeled with partially observable Markov decision processes (POMDPs). Some predictive learners are able to infer the state from observations and thus perform better with partial observability. Predictive state representations (PSRs) and neural networks are two such tools that can be trained to predict the probabilities of future observations. However, most such existing methods focus primarily on static problems in which only one environment is learned. In this paper, we propose an algorithm that uses statistical tests to estimate the probability of different predictive models to fit the current environment. We exploit the underlying probability distributions of predictive models to provide a fast and explainable method to assess and justify the model's beliefs about the current environment. Crucially, by doing so, the method can label incoming data as fitting different models, and thus can continuously train separate models in different environments. This new method is shown to prevent catastrophic forgetting when new environments, or tasks, are encountered. The method can also be of use when AI-informed decisions require justifications because its beliefs are based on statistical evidence from observations. We empirically demonstrate the benefit of the novel method with simulations in a set of POMDP environments.

Transcranial alternating current stimulation entrains single-neuron activity in the primate brain.

Spike timing is thought to play a critical role in neural computation and communication. Methods for adjusting spike timing are therefore of great interest to researchers and clinicians alike. Transcranial electrical stimulation (tES) is a noninvasive technique that uses weak electric fields to manipulate brain activity. Early results have suggested that this technique can improve subjects' behavioral performance on a wide range of tasks and ameliorate some clinical conditions. Nevertheless, considerable skepticism remains about its efficacy, especially because the electric fields reaching the brain during tES are small, whereas the likelihood of indirect effects is large. Our understanding of its effects in humans is largely based on extrapolations from simple model systems and indirect measures of neural activity. As a result, fundamental questions remain about whether and how tES can influence neuronal activity in the human brain. Here, we demonstrate that tES, as typically applied to humans, affects the firing patterns of individual neurons in alert nonhuman primates, which are the best available animal model for the human brain. Specifically, tES consistently influences the timing, but not the rate, of spiking activity within the targeted brain region. Such effects are frequency- and location-specific and can reach deep brain structures; control experiments show that they cannot be explained by sensory stimulation or other indirect influences. These data thus provide a strong mechanistic rationale for the use of tES in humans and will help guide the development of future tES applications.

Transcranial Current Stimulation During Sleep Facilitates Insight into Temporal Rules, but does not Consolidate Memories of Individual Sequential Experiences.

Slow-wave sleep (SWS) is known to contribute to memory consolidation, likely through the reactivation of previously encoded waking experiences. Contemporary studies demonstrate that when auditory or olfactory stimulation is administered during memory encoding and then reapplied during SWS, memory consolidation can be enhanced, an effect that is believed to rely on targeted memory reactivation (TMR) induced by the sensory stimulation. Here, we show that transcranial current stimulations (tCS) during sleep can also be used to induce TMR, resulting in the facilitation of high-level cognitive processes. Participants were exposed to repeating sequences in a realistic 3D immersive environment while being stimulated with particular tCS patterns. A subset of these tCS patterns was then reapplied during sleep stages N2 and SWS coupled to slow oscillations in a closed-loop manner. We found that in contrast to our initial hypothesis, performance for the sequences corresponding to the reapplied tCS patterns was no better than for other sequences that received stimulations only during wake or not at all. In contrast, we found that the more stimulations participants received overnight, the more likely they were to detect temporal regularities governing the learned sequences the following morning, with tCS-induced beta power modulations during sleep mediating this effect.

One-Shot Tagging During Wake and Cueing During Sleep With Spatiotemporal Patterns of Transcranial Electrical Stimulation Can Boost Long-Term Metamemory of Individual Episodes in Humans.

Targeted memory reactivation (TMR) during slow-wave oscillations (SWOs) in sleep has been demonstrated with sensory cues to achieve about 5-12% improvement in post-nap memory performance on simple laboratory tasks. But prior work has not yet addressed the one-shot aspect of episodic memory acquisition, or dealt with the presence of interference from ambient environmental cues in real-world settings. Further, TMR with sensory cues may not be scalable to the multitude of experiences over one's lifetime. We designed a novel non-invasive non-sensory paradigm that tags one-shot experiences of minute-long naturalistic episodes in immersive virtual reality (VR) with unique spatiotemporal amplitude-modulated patterns (STAMPs) of transcranial electrical stimulation (tES). In particular, we demonstrated that these STAMPs can be re-applied as brief pulses during SWOs in sleep to achieve about 10-20% improvement in the metamemory of targeted episodes compared to the control episodes at 48 hours after initial viewing. We found that STAMPs can not only facilitate but also impair metamemory for the targeted episodes based on an interaction between pre-sleep metamemory and the number of STAMP applications during sleep. Overnight metamemory improvements were mediated by spectral power increases following the offset of STAMPs in the slow-spindle band (8-12 Hz) for left temporal areas in the scalp electroencephalography (EEG) during sleep. These results prescribe an optimal strategy to leverage STAMPs for boosting metamemory and suggest that real-world episodic memories can be modulated in a targeted manner even with coarser, non-invasive spatiotemporal stimulation.

Dose-Dependent Effects of Closed-Loop tACS Delivered During Slow-Wave Oscillations on Memory Consolidation.

Sleep is critically important to consolidate information learned throughout the day. Slow-wave sleep (SWS) serves to consolidate declarative memories, a process previously modulated with open-loop non-invasive electrical stimulation, though not always effectively. These failures to replicate could be explained by the fact that stimulation has only been performed in open-loop, as opposed to closed-loop where phase and frequency of the endogenous slow-wave oscillations (SWOs) are matched for optimal timing. The current study investigated the effects of closed-loop transcranial Alternating Current Stimulation (tACS) targeting SWOs during sleep on memory consolidation. 21 participants took part in a three-night, counterbalanced, randomized, single-blind, within-subjects study, investigating performance changes (correct rate and F1 score) on images in a target detection task over 24 h. During sleep, 1.5 mA closed-loop tACS was delivered in phase over electrodes at F3 and F4 and 180° out of phase over electrodes at bilateral mastoids at the frequency (range 0.5-1.2 Hz) and phase of ongoing SWOs for a duration of 5 cycles in each discrete event throughout the night. Data were analyzed in a repeated measures ANOVA framework, and results show that verum stimulation improved post-sleep performance specifically on generalized versions of images used in training at both morning and afternoon tests compared to sham, suggesting the facilitation of schematization of information, but not of rote, veridical recall. We also found a surprising inverted U-shaped dose effect of sleep tACS, which is interpreted in terms of tACS-induced faciliatory and subsequent refractory dynamics of SWO power in scalp EEG. This is the first study showing a selective modulation of long-term memory generalization using a novel closed-loop tACS approach, which holds great potential for both healthy and neuropsychiatric populations.

Modeling Contextual Modulation of Memory Associations in the Hippocampus.

We present a computational model of how memories can be contextually acquired and recalled in the hippocampus. Our adaptive contextual memory model comprises the lateral entorhinal cortex (LEC), the dentate gyrus (DG) and areas CA3 and CA1 in the hippocampus, and assumes external inputs about context that originate in the prefrontal cortex (PFC). Specifically, we propose that there is a top-down bias on the excitability of cells in the DG of the hippocampus that recruits a sub-population of cells to differentiate contexts, independent of experienced stimuli, expanding the "pattern separation" role typically attributed to the DG. It has been demonstrated in rats that if PFC is inactivated, both acquisition and recall of memory associations are impaired. However, PFC inactivation during acquisition of one set of memory associations surprisingly leads to subsequent facilitation of the acquisition of a conflicting set of memory associations in the same context under normal PFC operation. We provide here the first computational and algorithmic account of how the absence or presence of the top-down contextual biases on the excitability of DG cells during different learning phases of these experiments explains these data. Our model simulates PFC inactivation as the loss of inhibitory control on DG, which leads to full or partial activation of DG cells related to conflicting memory associations previously acquired in different contexts. This causes context-inappropriate memory traces to become active in the CA3 recurrent network and thereby the output CA1 area within the hippocampus. We show that these incongruous memory patterns proactively interfere with and slow the acquisition of new memory associations. Further, we demonstrate that pattern completion within CA3 in response to a partial cue for the recall of previously acquired memories is also impaired by PFC inactivation for the same reason. Pre-training the model with interfering memories in contexts different from those used in the experiments, simulating a lifetime of experiences, was crucial to reproduce the rat behavioral data. Finally, we made several testable predictions based on the model that suggest future experiments to deepen our understanding of brain-wide memory processes.

The Benefits of Closed-Loop Transcranial Alternating Current Stimulation on Subjective Sleep Quality.

BACKGROUND: Poor sleep quality is a common complaint, affecting over one third of people in the United States. While sleep quality is thought to be related to slow-wave sleep (SWS), there has been little investigation to address whether modulating slow-wave oscillations (SWOs) that characterize SWS could impact sleep quality. Here we examined whether closed-loop transcranial alternating current stimulation (CL-tACS) applied during sleep impacts sleep quality and efficiency. METHODS: CL-tACS was used in 21 participants delivered at the same frequency and in phase with endogenous SWOs during sleep. Sleep quality was assessed in the morning following either verum or sham control stimulation during sleep, with order counterbalanced within participants. RESULTS: Higher sleep quality and efficiency were found after verum stimulation nights compared to control. The largest effects on sleep quality were found immediately following an adaptation night in the laboratory for which sleep quality was reduced. CONCLUSIONS: Applying CL-tACS at the same frequency and phase as endogenous SWOs may offer a novel method to improve subjective sleep quality after a night with poor quality sleep. CL-tACS might be helpful for increasing sleep quality and efficiency in otherwise healthy people, and in patients with clinical disorders that involve sleep deficits.

Closed-Loop Slow-Wave tACS Improves Sleep-Dependent Long-Term Memory Generalization by Modulating Endogenous Oscillations.

Benefits in long-term memory retention and generalization have been shown to be related to sleep-dependent processes, which correlate with neural oscillations as measured by changes in electric potential. The specificity and causal role of these oscillations, however, are still poorly understood. Here, we investigated the potential for augmenting endogenous slow-wave (SW) oscillations in humans with closed-loop transcranial alternating current stimulation (tACS) with an aim toward enhancing the consolidation of recent experiences into long-term memory. Sixteen (three female) participants were trained presleep on a target detection task identifying targets hidden in complex visual scenes. During post-training sleep, closed-loop SW detection and stimulation were used to deliver tACS matching the phase and frequency of the dominant oscillation in the range of 0.5-1.2 Hz. Changes in performance were assessed the following day using test images that were identical to the training ("repeated"), and images generated from training scenes but with novel viewpoints ("generalized"). Results showed that active SW tACS during sleep enhanced the postsleep versus presleep target detection accuracy for the generalized images compared with sham nights, while no significant change was found for repeated images. Using a frequency-agnostic clustering approach sensitive to stimulation-induced spectral power changes in scalp EEG, this behavioral enhancement significantly correlated with both a poststimulation increase and a subsequent decrease in measured spectral power within the SW band, which in turn showed increased coupling with spindle amplitude. These results suggest that augmenting endogenous SW oscillations can enhance consolidation by specifically improving generalization over recognition or cued recall.SIGNIFICANCE STATEMENT This human study demonstrates the use of a closed-loop noninvasive brain stimulation method to enhance endogenous neural oscillations during sleep with the effect of improving consolidation of recent experiences into long-term memory. Here we show that transient slow oscillatory transcranial alternating current stimulation (tACS) triggered by endogenous slow oscillations and matching their frequency and phase can increase slow-wave power and coupling with spindles. Further, this increase correlates with overnight improvements in generalization of recent training to facilitate performance in a target detection task. We also provide novel evidence for a tACS-induced refractory period following the tACS-induced increase. Here slow-wave power is temporarily reduced relative to sham stimulation, which nonetheless maintains a positive relationship with behavioral improvements.

Mental State Assessment and Validation Using Personalized Physiological Biometrics.

Mental state monitoring is a critical component of current and future human-machine interfaces, including semi-autonomous driving and flying, air traffic control, decision aids, training systems, and will soon be integrated into ubiquitous products like cell phones and laptops. Current mental state assessment approaches supply quantitative measures, but their only frame of reference is generic population-level ranges. What is needed are physiological biometrics that are validated in the context of task performance of individuals. Using curated intake experiments, we are able to generate personalized models of three key biometrics as useful indicators of mental state; namely, mental fatigue, stress, and attention. We demonstrate improvements to existing approaches through the introduction of new features. Furthermore, addressing the current limitations in assessing the efficacy of biometrics for individual subjects, we propose and employ a multi-level validation scheme for the biometric models by means of k-fold cross-validation for discrete classification and regression testing for continuous prediction. The paper not only provides a unified pipeline for extracting a comprehensive mental state evaluation from a parsimonious set of sensors (only EEG and ECG), but also demonstrates the use of validation techniques in the absence of empirical data. Furthermore, as an example of the application of these models to novel situations, we evaluate the significance of correlations of personalized biometrics to the dynamic fluctuations of accuracy and reaction time on an unrelated threat detection task using a permutation test. Our results provide a path toward integrating biometrics into augmented human-machine interfaces in a judicious way that can help to maximize task performance.

Transcranial Direct Current Stimulation Facilitates Associative Learning and Alters Functional Connectivity in the Primate Brain.

There has been growing interest in transcranial direct current stimulation (tDCS), a non-invasive technique purported to modulate neural activity via weak, externally applied electric fields. Although some promising preliminary data have been reported for applications ranging from stroke rehabilitation to cognitive enhancement, little is known about how tDCS affects the human brain, and some studies have concluded that it may have no effect at all. Here, we describe a macaque model of tDCS that allows us to simultaneously examine the effects of tDCS on brain activity and behavior. We find that applying tDCS to right prefrontal cortex improves monkeys' performance on an associative learning task. While firing rates do not change within the targeted area, tDCS does induce large low-frequency oscillations in the underlying tissue. These oscillations alter functional connectivity, both locally and between distant brain areas, and these long-range changes correlate with tDCS's effects on behavior. Together, these results are consistent with the idea that tDCS leads to widespread changes in brain activity and suggest that it may be a valuable method for cheaply and non-invasively altering functional connectivity in humans.