Flexible computation is a hallmark of intelligent behavior. Yet, little is known about how neural networks contextually reconfigure for different computations. Humans are able to perform a new task without extensive training, presumably through the composition of elementary processes that were previously learned. Cognitive scientists have long hypothesized the possibility of a compositional neural code, where complex neural computations are made up of constituent components; however, the neural substrate underlying this structure remains elusive in biological and artificial neural networks. Here we identified an algorithmic neural substrate for compositional computation through the study of multitasking artificial recurrent neural networks. Dynamical systems analyses of networks revealed learned computational strategies that mirrored the modular subtask structure of the task-set used for training. Dynamical motifs such as attractors, decision boundaries and rotations were reused across different task computations. For example, tasks that required memory of a continuous circular variable repurposed the same ring attractor. We show that dynamical motifs are implemented by clusters of units and are reused across different contexts, allowing for flexibility and generalization of previously learned computation. Lesioning these clusters resulted in modular effects on network performance: a lesion that destroyed one dynamical motif only minimally perturbed the structure of other dynamical motifs. Finally, modular dynamical motifs could be reconfigured for fast transfer learning. After slow initial learning of dynamical motifs, a subsequent faster stage of learning reconfigured motifs to perform novel tasks. This work contributes to a more fundamental understanding of compositional computation underlying flexible general intelligence in neural systems. We present a conceptual framework that establishes dynamical motifs as a fundamental unit of computation, intermediate between the neuron and the network. As more whole brain imaging studies record neural activity from multiple specialized systems simultaneously, the framework of dynamical motifs will guide questions about specialization and generalization across brain regions.

A major neuroscientific challenge is to identify the neural mechanisms that support consciousness. Though experimental studies have accumulated evidence about the location of the neural substrate of consciousness, we still lack a full understanding of why certain brain areas, but not others, can support consciousness. In this talk, I will give an overview of our approach, taking advantage of the theoretical framework provided by Integrated Information Theory (IIT). First, I will introduce results showing that a maximum of integrated information within the human brain matches our best evidence concerning the location of the NSC, supporting the IIT’s prediction. Furthermore, I will discuss the possibility that the NSC can change its location and even split into two depending on the task demand. Finally, based on some graph-theoretical analyses, I will argue that the ability of different brain regions to contribute or not to consciousness depends on specific properties of their anatomical connectivity, which determines their ability to support high integrated information.

Integrated information theory (IIT) takes as its starting point phenomenology, rather than behavioral, functional, or neural correlates of consciousness. The theory characterizes the essential properties of phenomenal existence—which is immediate and indubitable. These are translated into physical properties, expressed operationally as cause-effect power, which must be satisfied by the neural substrate of consciousness. On this basis, the theory can account for clinical and experimental data about the presence and absence of consciousness. Current work aims at accounting for specific qualities of different experiences, such as spatial extendedness and the flow of time. Several implications of IIT have ethical relevance. One is that functional equivalence does not imply phenomenal equivalence—computers may one day be able to do everything we do, but they will not experience anything. Another is that we do have free will in the fundamental, metaphysical sense—we have true alternatives and we, not our neurons, are the true cause of our willed actions.

New memories are initially labile and have to be consolidated into stable long-term representations. Current theories assume that this is supported by a shift in the neural substrate that supports the memory, away from rapidly plastic hippocampal networks towards more stable representations in the neocortex. Rehearsal, i.e. repeated activation of the neural circuits that store a memory, is thought to crucially contribute to the formation of neocortical long-term memory representations. This may either be achieved by repeated study during wakefulness or by a covert reactivation of memory traces during offline periods, such as quiet rest or sleep. My research investigates memory consolidation in the human brain with multivariate decoding of neural processing and non-invasive in-vivo imaging of microstructural plasticity. Using pattern classification on recordings of electrical brain activity, I show that we spontaneously reprocess memories during offline periods in both sleep and wakefulness, and that this reactivation benefits memory retention. In related work, we demonstrate that active rehearsal of learning material during wakefulness can facilitate rapid systems consolidation, leading to an immediate formation of lasting memory engrams in the neocortex. These representations satisfy general mnemonic criteria and cannot only be imaged with fMRI while memories are actively processed but can also be observed with diffusion-weighted imaging when the traces lie dormant. Importantly, sleep seems to hold a crucial role in stabilizing the changes in the contribution of memory systems initiated by rehearsal during wakefulness, indicating that online and offline reactivation might jointly contribute to forming long-term memories. Characterizing the covert processes that decide whether, and in which ways, our brains store new information is crucial to our understanding of memory formation. Directly imaging consolidation thus opens great opportunities for memory research.

How do people learn in real-world environments where the space of possible actions can be vast or even infinite? The study of human learning has made rapid progress in past decades, from discovering the neural substrate of reward prediction errors, to building AI capable of mastering the game of Go. Yet this line of research has primarily focused on learning through repeated interactions with the same stimuli. How are humans able to rapidly adapt to novel situations and learn from such sparse examples? I propose a theory of how generalization guides human learning, by making predictions about which unobserved options are most promising to explore. Inspired by Roger Shepard’s law of generalization, I show how a Bayesian function learning model provides a mechanism for generalizing limited experiences to a wide set of novel possibilities, based on the simple principle that similar actions produce similar outcomes. This model of generalization generates predictions about the expected reward and underlying uncertainty of unexplored options, where both are vital components in how people actively explore the world. This model allows us to explain developmental differences in the explorative behavior of children, and suggests a general principle of learning across spatial, conceptual, and structured domains.

In many ways, cognitive neuroscience is the attempt to use physiological observation to clarify the mechanisms that shape behavior. Over the past 25 years, fMRI has provided a system-wide and yet somewhat spatially precise view of the response in human cortex evoked by a wide variety of stimuli and task contexts. The current talk focuses on the other direction of inference; the implications of this observed functional organization for behavior. To begin, we must interrogate the methodological and empirical frameworks underlying our derivation of this organization, partially by exploring its relationship to and predictability from gross neuroanatomy. Next, across a series of studies, the implications of two properties of functional organization for behavior will be explored: 1) the co-localization of visual working memory and perceptual processing and 2) implicit learning in the context of distributed responses. In sum, these results highlight the limitations of our current approach and hint at a new general mechanism for explaining observed behavior in context with the neural substrate.