The relational bottleneck as an inductive bias for efficient abstraction.

A central challenge for cognitive science is to explain how abstract concepts are acquired from limited experience. This has often been framed in terms of a dichotomy between connectionist and symbolic cognitive models. Here, we highlight a recently emerging line of work that suggests a novel reconciliation of these approaches, by exploiting an inductive bias that we term the relational bottleneck. In that approach, neural networks are constrained via their architecture to focus on relations between perceptual inputs, rather than the attributes of individual inputs. We review a family of models that employ this approach to induce abstractions in a data-efficient manner, emphasizing their potential as candidate models for the acquisition of abstract concepts in the human mind and brain.

Human Curriculum Effects Emerge with In-Context Learning in Neural Networks.

Human learning is sensitive to rule-like structure and the curriculum of examples used for training. In tasks governed by succinct rules, learning is more robust when related examples are blocked across trials, but in the absence of such rules, interleaving is more effective. To date, no neural model has simultaneously captured these seemingly contradictory effects. Here we show that this same tradeoff spontaneously emerges with "in-context learning" (ICL) both in neural networks trained with metalearning and in large language models (LLMs). ICL is the ability to learn new tasks "in context" - without weight changes - via an inner-loop algorithm implemented in activation dynamics. Experiments with pretrained LLMs and metalearning transformers show that ICL exhibits the blocking advantage demonstrated in humans on a task involving rule-like structure, and conversely, that concurrent in-weight learning reproduces the interleaving advantage observed in humans on tasks lacking such structure.

Properties of LoTs: The footprints or the bear itself?

There are two ways to understand any proposed properties of language-of-thoughts (LoTs): As diagnostic or constitutive. We argue that this choice is critical. If candidate properties are diagnostic, their homeostatic clustering requires explanation via an underlying homeostatic mechanism. If constitutive, there is no clustering, only the properties themselves. Whether deep neural networks (DNNs) are alternatives to LoTs or potential implementations turn on this choice.

The Structure of Systematicity in the Brain.

A hallmark of human intelligence is the ability to adapt to new situations, by applying learned rules to new content (systematicity) and thereby enabling an open-ended number of inferences and actions (generativity). Here, we propose that the human brain accomplishes these feats through pathways in the parietal cortex that encode the abstract structure of space, events, and tasks, and pathways in the temporal cortex that encode information about specific people, places, and things (content). Recent neural network models show how the separation of structure and content might emerge through a combination of architectural biases and learning, and these networks show dramatic improvements in the ability to capture systematic, generative behavior. We close by considering how the hippocampal formation may form integrative memories that enable rapid learning of new structure and content representations.

A Neural Network Model of Continual Learning with Cognitive Control.

Neural networks struggle in continual learning settings from catastrophic forgetting: when trials are blocked, new learning can overwrite the learning from previous blocks. Humans learn effectively in these settings, in some cases even showing an advantage of blocking, suggesting the brain contains mechanisms to overcome this problem. Here, we build on previous work and show that neural networks equipped with a mechanism for cognitive control do not exhibit catastrophic forgetting when trials are blocked. We further show an advantage of blocking over interleaving when there is a bias for active maintenance in the control signal, implying a tradeoff between maintenance and the strength of control. Analyses of map-like representations learned by the networks provided additional insights into these mechanisms. Our work highlights the potential of cognitive control to aid continual learning in neural networks, and offers an explanation for the advantage of blocking that has been observed in humans.

Complementary Structure-Learning Neural Networks for Relational Reasoning.

The neural mechanisms supporting flexible relational inferences, especially in novel situations, are a major focus of current research. In the complementary learning systems framework, pattern separation in the hippocampus allows rapid learning in novel environments, while slower learning in neocortex accumulates small weight changes to extract systematic structure from well-learned environments. In this work, we adapt this framework to a task from a recent fMRI experiment where novel transitive inferences must be made according to implicit relational structure. We show that computational models capturing the basic cognitive properties of these two systems can explain relational transitive inferences in both familiar and novel environments, and reproduce key phenomena observed in the fMRI experiment.

Compositional Processing Emerges in Neural Networks Solving Math Problems.

A longstanding question in cognitive science concerns the learning mechanisms underlying compositionality in human cognition. Humans can infer the structured relationships (e.g., grammatical rules) implicit in their sensory observations (e.g., auditory speech), and use this knowledge to guide the composition of simpler meanings into complex wholes. Recent progress in artificial neural networks has shown that when large models are trained on enough linguistic data, grammatical structure emerges in their representations. We extend this work to the domain of mathematical reasoning, where it is possible to formulate precise hypotheses about how meanings (e.g., the quantities corresponding to numerals) should be composed according to structured rules (e.g., order of operations). Our work shows that neural networks are not only able to infer something about the structured relationships implicit in their training data, but can also deploy this knowledge to guide the composition of individual meanings into composite wholes.

Deep Predictive Learning in Neocortex and Pulvinar.

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

How Sequential Interactive Processing Within Frontostriatal Loops Supports a Continuum of Habitual to Controlled Processing.

We address the distinction between habitual/automatic vs. goal-directed/controlled behavior, from the perspective of a computational model of the frontostriatal loops. The model exhibits a continuum of behavior between these poles, as a function of the interactive dynamics among different functionally-specialized brain areas, operating iteratively over multiple sequential steps, and having multiple nested loops of similar decision making circuits. This framework blurs the lines between these traditional distinctions in many ways. For example, although habitual actions have traditionally been considered purely automatic, the outer loop must first decide to allow such habitual actions to proceed. Furthermore, because the part of the brain that generates proposed action plans is common across habitual and controlled/goal-directed behavior, the key differences are instead in how many iterations of sequential decision-making are taken, and to what extent various forms of predictive (model-based) processes are engaged. At the core of every iterative step in our model, the basal ganglia provides a "model-free" dopamine-trained Go/NoGo evaluation of the entire distributed plan/goal/evaluation/prediction state. This evaluation serves as the fulcrum of serializing otherwise parallel neural processing. Goal-based inputs to the nominally model-free basal ganglia system are among several ways in which the popular model-based vs. model-free framework may not capture the most behaviorally and neurally relevant distinctions in this area.

Computational models of motivated frontal function.

Computational models of frontal function have made important contributions to understanding how the frontal lobes support a wide range of important functions, in their interactions with other brain areas including, critically, the basal ganglia (BG). We focus here on the specific case of how different frontal areas support goal-directed, motivated decision-making, by representing three essential types of information: possible plans of action (in more dorsal and lateral frontal areas), affectively significant outcomes of those action plans (in ventral, medial frontal areas including the orbital frontal cortex), and the overall utility of a given plan compared to other possible courses of action (in anterior cingulate cortex). Computational models of goal-directed action selection at multiple different levels of analysis provide insight into the nature of learning and processing in these areas and the relative contributions of the frontal cortex versus the BG. The most common neurologic disorders implicate these areas, and understanding their precise function and modes of dysfunction can contribute to the new field of computational psychiatry, within the broader field of computational neuroscience.

Effects of the presence and absence of amino acids on translation, signaling, and long-term depression in hippocampal slices from Fmr1 knockout mice.

Fragile X syndrome (FXS) is caused by silencing of the FMR1 gene and consequent absence of its protein product, fragile X mental retardation protein (FMRP). FMRP is an RNA-binding protein that can suppress translation. The absence of FMRP leads to symptoms of FXS including intellectual disability and has been proposed to lead to abnormalities in synaptic plasticity. Synaptic plasticity, protein synthesis, and cellular growth pathways have been studied extensively in hippocampal slices from a mouse model of FXS (Fmr1 KO). Enhanced metabotropic glutamate receptor 5 (mGluR5)-dependent long-term depression (LTD), increased rates of protein synthesis, and effects on signaling molecules have been reported. These phenotypes were found under amino acid starvation, a condition that has widespread, powerful effects on activation and translation of proteins involved in regulating protein synthesis. We asked if this non-physiological condition could have effects on Fmr1 KO phenotypes reported in hippocampal slices. We performed hippocampal slice experiments in the presence and absence of amino acids. We measured rates of incorporation of a radiolabeled amino acid into protein to determine protein synthesis rates. By means of western blots, we assessed relative levels of total and phosphorylated forms of proteins involved in signaling pathways regulating translation. We measured evoked field potentials in area CA1 to assess the strength of the long-term depression response to mGluR activation. In the absence of amino acids, we replicate many of the reported findings in Fmr1 KO hippocampal slices, but in the more physiological condition of inclusion of amino acids in the medium, we did not find evidence of enhanced mGluR5-dependent LTD. Activation of mGluR5 increased protein synthesis in both wild type and Fmr1 KO. Moreover, mGluR5 activation increased eIF2α phosphorylation and decreased phosphorylation of p70S6k in slices from Fmr1 KO. We propose that the eIF2α response is a cellular attempt to compensate for the lack of regulation of translation by FMRP. Our findings call for a re-examination of the mGluR theory of FXS.