Source-linked AI summary
The neuroconnectionist research programme
Adrien Doerig, Rowan Sommers, Katja Seeliger, Blake Richards, Jenann Ismael, Grace Lindsay, Konrad Kording, Talia Konkle, Marcel A. J. Van Gerven, Nikolaus Kriegeskorte, Tim C. Kietzmann
TL;DR
The paper addresses how to assess ANNs as brain models despite their behavioral and biological shortcomings. It frames neuroconnectionism as a Lakatosian research programme using ANN-based hypothesis testing, and concludes that its evolving models have generated novel insights and corroborated predictions while retaining important limitations.
Problem
Current ANNs differ from biology and behavior, and their shortcomings raise questions about whether they can support useful understanding of brain function.
Method
The paper characterizes neuroconnectionism as a Lakatosian programme that uses ANN architectures, data, objectives, and learning rules to test hypotheses across neural and behavioral levels.
Results
Neuroconnectionist projects have produced state-of-the-art models, novel theoretical insights and predictions, and predictions corroborated by in vivo experiments.
Takeaways & Limitations
The programme should be evaluated by its longitudinal progress, including how it generates insights, corroborates hypotheses, and productively addresses criticism.
Takeaways & Limitations
The approach includes edge cases: some models use high-level conceptual inputs, omit behavioral objectives, fit neural data directly, or target specific cognitive components.
Abstract
from arXiv · showhide
Artificial Neural Networks (ANNs) inspired by biology are beginning to be widely used to model behavioral and neural data, an approach we call neuroconnectionism. ANNs have been lauded as the current best models of information processing in the brain, but also criticized for failing to account for basic cognitive functions. We propose that arguing about the successes and failures of a restricted set of current ANNs is the wrong approach to assess the promise of neuroconnectionism. Instead, we take inspiration from the philosophy of science, and in particular from Lakatos, who showed that the core of scientific research programmes is often not directly falsifiable, but should be assessed by its capacity to generate novel insights. Following this view, we present neuroconnectionism as a cohesive large-scale research programme centered around ANNs as a computational language for expressing falsifiable theories about brain computation. We describe the core of the programme, the underlying computational framework and its tools for testing specific neuroscientific hypotheses. Taking a longitudinal view, we review past and present neuroconnectionist projects and their responses to challenges, and argue that the research programme is highly progressive, generating new and otherwise unreachable insights into the workings of the brain.
Neuroconnectionism as a Lakatosian research programme
Neuroconnectionism is framed as a Lakatosian research programme with a stable ANN-centered core and a revisable belt of auxiliary hypotheses. Its promise is assessed longitudinally by whether it generates insights and corroborated predictions rather than by current experimental success alone.
- The core contains background assumptions that are not typically challenged internally, while belt elements are experimentally tested and changeable.
- A research programme is progressive when it generates new theoretical insights and novel predictions that receive empirical corroboration.
- The programme’s value depends on longitudinal progress, not only its current experimental success relative to competing programmes.
- Neuroconnectionism comprises auxiliary hypotheses organized in a belt around a unifying core, rather than one falsifiable hypothesis.
BOX I: THEORY SELECTION & PHILOSOPHY OF SCIENCE
The paper contrasts Popperian rejection with a Lakatosian approach that tests theories through auxiliary assumptions and revises the surrounding belt while preserving a productive core. This reflects the holistic structure of scientific confirmation.
- Popperian testing treats failed predictions as grounds for rejecting the theory once experimental error is excluded.
- Predictions typically depend on a target hypothesis plus auxiliary beliefs, so a failed observation does not identify which component is responsible.
- Lakatos’ core-and-belt distinction captures this holistic confirmation structure by insulating guiding principles from direct testing through intermediary hypotheses.
- Lakatosian testing preserves core principles while exploring alternative auxiliary hypotheses, rejecting the core only when preserving it becomes unproductive.
The neuroconnectionist core
The neuroconnectionist core holds that ANN architectures, input statistics, objectives, and learning rules are the right modeling language for multilevel, sensory-grounded, iterative brain computations. The programme remains cohesive but flexible, using hypothesis-driven models and revisable design choices to connect mechanisms with cognition.
- The neuroconnectionist core: ANNs model cognition as distributed computations emerging from simple interconnected units rather than from individually cognitive components.
- The neuroconnectionist core: Architectures constrain connectivity and encode inductive biases, while convolutional networks share feature selectivity across locations in vision.
- The neuroconnectionist core: ANN architecture, training data, objectives, and learning rules map onto questions about structure, experience, neural selectivity, and representational change.
- The neuroconnectionist core: Iterative processing is treated as necessary because neural-network learning and dynamics cannot currently be replaced by analytical shortcuts.
- The neuroconnectionist core: The core proposes that brain computations, representations, learning mechanisms, and inductive biases are best understood through ANN models defined by architecture, input statistics, objectives, and learning rules.
- The neuroconnectionist core: The programme has no rigid definition: models may use conceptual inputs, omit behavioral objectives, fit neural data directly, or focus on specific cognitive components.
The neuroconnectionist toolbox
The neuroconnectionist toolbox provides methods for instantiating ANN belt hypotheses and testing their alignment with neural data and behavior. Failures are treated as evidence for improving models and understanding.
- The toolbox includes tools for testing how closely ANN instances align with neural data and behavior.
- Findings that reveal ANN shortcomings supply evidence about where models and understanding need improvement.
Network design and training
Neuroconnectionist models are designed and trained by selecting architectures, datasets, objectives, and learning rules that support testing biological and behavioral hypotheses. This framework enables researchers to examine which biological aspects are necessary to reproduce brain function and behavior.
- Researchers select among architectures and unit types ranging from simple ReLU units to memory-oriented LSTMs, increasingly incorporating biological features.
- Training choices include datasets, objectives, and learning rules, with networks sometimes trained directly to mirror brain activity.
- Objectives span supervised, unsupervised, predictive, generative, efficiency-based, and behavioral-reward learning, shaping what networks learn from input statistics.
- Backpropagation remains the dominant learning rule, while researchers also investigate biologically plausible alternatives such as Hebbian learning.
- Big-data and high-throughput frameworks let researchers test which biological aspects are necessary to reproduce brain function and behavior.
Model testing
Neuroconnectionism tests trained ANNs against behavior, neural activity, internal mechanisms, and developmental trajectories using complementary methods. This multilevel evaluation supports rigorous hypothesis testing while acknowledging that no single comparison method is sufficient.
- ANNs can be inspected through in-silico experiments because researchers have direct access to every unit’s activity and connectivity.These experiments can be conducted much faster than biological experiments and without classical experimentation’s ethical concerns.
- Because multiple methods are imperfect and structurally different systems can make successful predictions, complementary metrics are needed across behavioral and neural levels.
- Behavioral agreement is assessed with task performance, reaction times, error patterns, out-of-distribution tests, and psychophysical results.Overall accuracy is useful but increasingly insufficient as ANNs approach human performance on benchmark tasks.
- Neural agreement is commonly evaluated with RSA, which compares representational geometries, and encoding models, which predict biological-unit activity from ANN activations.Stimuli are presented identically to the brain and model before comparing their activity patterns.
- ANN-generated stimuli can strongly activate a target biological neuron or area, extending evaluation beyond correlational RSA and encoding approaches.
- Training trajectories can be compared with biological development from untrained to fully trained networks, although their relation to evolution versus lifetime learning remains unclear.
The neuroconnectionist belt
The neuroconnectionist belt is a changing set of hypotheses evaluated longitudinally rather than a fixed collection of models. Across vision and other domains, successive architectural, training, and objective changes have produced stronger fits to neural and behavioral data and new theoretical insights.
- The neuroconnectionist belt: A rejected belt hypothesis is treated as evidence for revising the programme rather than as a refutation of its core assumptions.
- The neuroconnectionist belt: The belt is assessed by whether it generates new insights and addresses challenges progressively rather than becoming blocked by them.
- Historical progression: In visual neuroscience, the neocognitron and HMAX illustrate how models were proposed, tested against neural and behavioral findings, and revised after mismatches.
- Historical progression: Convolutional networks trained on naturalistic photos matched neural activity patterns along the primate ventral stream, strengthening the neuroconnectionist belt.
- Progressive evolution: The field now tests diverse architectures, datasets, objectives, and learning techniques across settings to align models with brain and behavioral data.
- Progressive evolution: Recurrent networks improved matches to behavioral data and delayed neural activity, while later models incorporated additional architectural and objective choices.
- Progressive evolution: Self-supervised training produced as good or better matches to neural representations and animal behavior, and could account for distinctions between dorsal and ventral pathways.
- New insights: Neuroconnectionist models have yielded state-of-the-art fits and new perspectives on recurrence, attention, and the dorsal stream, with some predictions corroborated in vivo.
BOX II: PREDICTIONS & INSIGHTS GENERATED BY ANNs
Neuroconnectionism uses ANNs to generate novel, testable predictions and insights while treating behavioral, symbolic, and biological mismatches as signposts for model development. Its progressive research programme balances computational tractability with biologically grounded detail and repeatedly improves models, methods, and hypotheses.
- Predictions and insights: ANNs generate novel, testable predictions about neural selectivity, including stimuli that drive biological neurons more specifically than natural stimuli.Input optimization provides a causal extension to correlational model–brain alignment tests.
- Predictions and insights: ANN-derived object spaces unified macaque IT organization by aligning neural responses, including previously uncharacterized units, with low-dimensional object-space axes.The theory extended known selectivity profiles into previously uncharacterized IT regions.
- Predictions and insights: Neuroconnectionist models have produced theories of semantic learning, working-memory development, image memorability, and embodied learning in mouse visual cortex.These examples use ANN learning dynamics, pruning, or training conditions to derive computational insights into neural and behavioral phenomena.
- Progressive research programme: The research programme is progressive because model, theory, and methodology form a virtuous circle in which new results motivate improved methods and further discoveries.Researchers have altered architectures, training objectives, and datasets while acknowledging remaining methodological issues.
- Challenges as signposts: Behavioral mismatches, symbolic limitations, and biological simplifications are treated as challenges that motivate attention, recurrence, foveation, receptive-field, and neurosymbolic developments.The programme frames these shortcomings as signposts rather than roadblocks while continuing to test and improve models.
- Biological abstraction: Neuroconnectionism tests biological detail top-down across models that are abstract enough to be tractable and trainable yet detailed enough to map onto neural and behavioral data.The required level of complexity remains an empirical question, with biological features added through hypothesis testing rather than biological mirroring for its own sake.
Conclusions
The paper presents neuroconnectionism as a progressive research programme whose ANN-based models use challenges as signposts for further development. It also emphasizes that the field remains far from a complete explanation of cognition.
- Neuroconnectionism comprises auxiliary hypotheses and research directions sharing an ANN-based approach to cognitive phenomena through distributed neural communication.
- ANNs provide an abstract yet incrementally testable level of granularity, while their dimensionality supports sensory grounding and behavioral performance.Training hyperparameters serve as experimenter-controlled degrees of freedom for testing needed biological detail.
- The field remains far from explaining cognition and must address missing capacities including embodiment, symbolic reasoning, development, memory, and social learning.
- Current challenges can guide new research directions, while shortcomings in particular networks should not be treated as failures of the entire programme.