Stephen José Hanson

Deep learning (DL), a variant of the neural network algorithms originally proposed in the early twentieth century, has resulted in a renaissance of artificial intelligence. Despite the growing dominance of DL networks, little is understood about the learning mechanism that makes these networks so effective across such a wide range of applications. Drawing on a century of psychological learning theory (e.g., LL Thurstone), an account is offered of the learning mechanism that may enable DL networks to perform so successfully across so many different tasks. Specifically, evidence is provided that learning in DL networks is fit best by a hyperbolic function. This function is independent of hyper/meta parameters–not a scaling function but a learning curve to a specific equilibrium, a function that also entails an autocatalytic mechanism through which complex structures, abstracted from sensory features, can in principle create and support cognitive function in both biological and artificial systems. Keywords: Learning theory, Deep Learning, hyperbolic, autocatalytic, AI, Thurstone

Deep learning (DL), a variant of the neural network algorithms originally proposed in the 1980s (Rumelhart et al., 1986), has made surprising progress in artificial intelligence (AI), ranging from language translation, protein folding (Jumper et al., 2021), autonomous cars, and, more recently, human-like language models (chatbots). All that seemed intractable until very recently. Despite the growing use of DL networks, little is understood about the learning mechanisms and representations that make these networks effective across such a diverse range of applications. Part of the answer must be the huge scale of the architecture and, of course, the large scale of the data, since not much has changed since 1986. But the nature of deep learned representations remains largely unknown. Unfortunately, training sets with millions or billions of tokens have unknown combinatorics, and networks with millions or billions of hidden units can't easily be visualized and their mechanisms can't be easily revealed. In this letter, we explore these challenges with a large (1.24 million weights VGG) DL in a novel high-density sample task (five unique tokens with more than 500 exemplars per token), which allows us to more carefully follow the emergence of category structure and feature construction. We use various visualization methods for following the emergence of the classification and the development of the coupling of feature detectors and structures that provide a type of graphical bootstrapping. From these results, we harvest some basic observations of the learning dynamics of DL and propose a new theory of complex feature construction based on our results.

Editorial: Brain connectivity, dynamics, and complexity

Functional connectivity (FC) studies have predominantly focused on resting state, where ongoing dynamics are thought to reflect the brain’s intrinsic network architecture, which is thought to be broadly relevant because it persists across brain states (i.e., is state-general). However, it is unknown whether resting state is the optimal state for measuring intrinsic FC. We propose that latent FC, reflecting shared connectivity patterns across many brain states, better captures state-general intrinsic FC relative to measures derived from resting state alone. We estimated latent FC independently for each connection using leave-one-task-out factor analysis in seven highly distinct task states (24 conditions) and resting state using fMRI data from the Human Connectome Project. Compared with resting-state connectivity, latent FC improves generalization to held-out brain states, better explaining patterns of connectivity and task-evoked activation. We also found that latent connectivity improved prediction of behavior outside the scanner, indexed by the general intelligence factor (g). Our results suggest that FC patterns shared across many brain states, rather than just resting state, better reflect state-general connectivity. This affirms the notion of “intrinsic” brain network architecture as a set of connectivity properties persistent across brain states, providing an updated conceptual and mathematical framework of intrinsic connectivity as a latent factor.

The Failure of Blobology: fMRI Misinterpretation, Maleficience and Muddle

Brain activity flow models estimate the movement of task-evoked activity over brain connections to help explain the emergence of task-related functionality. Activity flow estimates have been shown to accurately predict task-evoked brain activations across a wide variety of brain regions and task conditions. However, these predictions have had limited explanatory power, given known issues with causal interpretations of the standard functional connectivity measures used to parameterize activity flow models. We show here that functional/effective connectivity (FC) measures grounded in causal principles facilitate mechanistic interpretation of activity flow models. Starting from Pearson correlation (the current field standard), we progress from FC measures with poor to excellent causal grounding, demonstrating a continuum of causal validity using simulations and empirical fMRI data. Finally, we apply a causal FC method to a dorsolateral prefrontal cortex region, demonstrating causal network mechanisms contributing to its strong activation during a 2-back (relative to a 0-back) working memory task. Together, these results reveal the promise of parameterizing activity flow models using causal FC methods to identify network mechanisms underlying cognitive computations in the human brain. Highlights - Activity flow models provide insight into how cognitive neural effects emerge from brain network interactions. - Functional connectivity methods grounded in causal principles facilitate mechanistic interpretations of task activity flow models. - Mechanistic activity flow models accurately predict task-evoked neural effects across a wide variety of brain regions and cognitive tasks.

Functional connectivity (FC) studies have predominantly focused on resting state, where ongoing dynamics are thought to reflect the brain’s intrinsic network architecture, which is thought to be broadly relevant because it persists across brain states (i.e., is state-general). However, it is unknown whether resting state is the optimal state for measuring intrinsic FC. We propose that latent FC, reflecting shared connectivity patterns across many brain states, better captures state-general intrinsic FC relative to measures derived from resting state alone. We estimated latent FC independently for each connection using leave-one-task-out factor analysis in 7 highly distinct task states (24 conditions) and resting state using fMRI data from the Human Connectome Project. Compared to resting-state connectivity, latent FC improves generalization to held-out brain states, better explaining patterns of connectivity and task-evoked activation. We also found that latent connectivity improved prediction of behavior outside the scanner, indexed by the general intelligence factor ( g ). Our results suggest that FC patterns shared across many brain states, rather than just resting state, better reflects state-general connectivity. This affirms the notion of “intrinsic” brain network architecture as a set of connectivity properties persistent across brain states, providing an updated conceptual and mathematical framework of intrinsic connectivity as a latent factor.

Background: Mentally simulating counterfactuals (scenarios that have not actually occurred) is a sophisticated human cognitive function underlying creativity, planning, and daydreaming. One example is the “would you rather” game, in which forced choices are made between outlandish negative counterfactuals. Methods: We measured behavioral and neural correlates while participants made “would you rather” choices framed as approaching or avoiding aversive counterfactual scenarios (e.g. illnesses, car accidents). Results: We found in two independent cohorts that participants were highly susceptible to framing effects when making these decisions, taking significantly longer to respond to approach frames compared to avoidance. Brain imaging showed that choices to approach and avoid resulted in a pattern of activation consistent with a network associated with responding to aversive stimuli, identified via a coordinate-based meta-analysis of 238 studies. Bayesian graph connectivity analysis showed that network connectivity differed by choice frame, with significantly stronger connectivity for approach choices compared to avoidance choices among primarily limbic nodes (putamen, insula, caudate, and amygdala). Computational modeling of behavior revealed that approach frames led to significantly longer non-decision times, increased evidence required to make decisions, and faster evidence accumulation than avoidance frames. Stronger network connectivity between corticostriatal and limbic regions was associated with rate of evidence accumulation and length of non-decision time during approach choices. For avoidance choices, prefrontal connectivity was related to non-decision time. Conclusions: These results suggest that “would you rather” decisions about aversive counterfactuals differentially recruit limbic circuit connectivity based on choice frame.

We introduce a new method for decoding neural data from fMRI. It is based on two assumptions, first that neural representation is distributed over networks of neurons embeded in voxel noise and second that the stimuli can be decoded as learned relations from sets of categorical stimuli. We illustrate these principles with two types of stimuli, color (wavelength) and letters (visual shape), both of which have early visual system response, but at the same time must be learned within a given function or category (color contrast, alphabet). Key to the decoding method is reducing the stimulus cross-correlation by a matched noise voxel sample by normalizing the stimulus voxel matrix thus unmasking a highly discriminative neural profile per stimulus. Projection of this new voxel space (ROI) to a smaller set of dimensions (with e.g., non-metric Multidimensional scaling), the relational information takes a unique geometric form revealing functional relationships between sets of stimuli, defined by R. Shepard, as second-order isomorphisms (SOI) . In the case of colors the SOI appears as a nearly equally spaced set of wavelengths arranged in a color wheel, with a gap between the “purples” and “reds” (consistent with the gap in the original Ekman’s color set). In the case of letters, a cluster space resulted from the decorrelated voxel neural profiles, which matched the phrase structure of the mnemonic used for more than 100 years to teach children the alphabet (across multiple languages), The Alphabet Song .