Notes on The Primacy of Behavioral Research for Understanding the Brain

Abstract
The core thesis is that understanding what the brain does (its cognitive functions) is best advanced by behavioral research, often more so than by neural measurements alone. Yael Niv argues that behavior provides sharper constraints on computational and mechanistic models of cognition than expensive neural data in many cases, and that even questions about neural mechanisms can benefit from carefully crafted behavioral paradigms. He contends that purely neural data often fall short of explaining brain function, and that the field has suffered from “neuroscience chauvinism” that overvalues neural measurements at the expense of behavior. The overarching claim is that behavioral research should be restored to its primacy in neuroscience, not discarded as subsidiary to neural data. The abstract foreshadows a program in which behavior guides theory, models, and interpretation of neural data, and where behavioral paradigms deliver strong causal inferences and practical constraints for understanding cognition.

Introduction
Niv critiques the contemporary trend toward ever more precise neural measurements driven by large-scale projects like the Brain Initiative and the Human Connectome Project. He notes a cultural shift that demotes animal and human behavior as merely “behavior” rather than the primary source of insight into brain function. The language of neuroscience often frames decision making as a neural process, rather than as a behaviorally grounded phenomenon, which he argues misrepresents the relationship between brain and mind. He warns that attempting to explain decision making solely through single neurons or neural ensembles is a misstep, likening it to drawing inferences about a country from its DNA rather than from its laws and culture.

Neural firing patterns are an inappropriate or incomplete level of analysis for many questions about the brain. Even with perfect neural data, without incisive behavioral paradigms, our understanding remains limited. The author cites Caenorhabditis elegans as a sobering example: a fully characterized nervous system and 302 neurons do not yield complete predictive understanding of the organism’s behavior. This illustrates that a circuit’s full knowledge does not guarantee understanding of its function, and that behavior must be integrated with neural inquiry. The piece advocates reversing the current hierarchy that places neural measurements at the base and behavior as optional, arguing for a historical primacy of behavioral inquiry.

Niv also reflects on methodological progress: cutting-edge tools for recording and manipulating neural activity have largely corroborated ideas derived from behavior rather than delivering entirely novel insights. While neural measurements are indispensable for localization and for testing theories about brain functions, he argues that many foundational insights about cognition originated from behavioral research and computational modeling. He emphasizes that a purely neural approach, even when complemented by perturbations, will rarely answer questions about high-level cognition without well-designed behavioral tasks. The message is that science advances best through a synthesis of behavior, computation, and neural data, with behavior serving as a robust constraints-driven framework for interpreting neural mechanisms.

What has behavior taught us about the brain?
Niv surveys historic demonstrations where behavioral inquiry revealed latent cognitive processes long before modern neural tools became available. In color vision, psychophysics inferred the existence of three cone types and their wavelength sensitivities before physiological patch-clamp measurements confirmed properties of cones. In attentional control and cognitive processing, simple behavioral paradigms such as visual search and studies of practice generalization yielded insights into the content and structure of attention and learning that guided later neural theories. The section argues that behavioral experiments can illuminate neural processes beyond what would be achievable with current or near-future neural measurements alone.

The author then presents a vivid behavioral example: a rat trained to run a T-maze base to the right arm to obtain food. The rat’s strategy could be egocentric (turning right relative to its body) or allocentric (relying on external room cues). The question is how to determine which strategy the rat uses during decision making. Merely recording brain activity would not straightforwardly reveal the strategy in use, and perturbing a brain region could not reveal the decision process in real time. Packard and McGaugh (1996) offered a clever behavioral manipulation: they rotated the maze, changing the alignment between egocentric and allocentric cues. If the rat consistently turned right, the strategy was egocentric; if it turned left, the rat relied on external, room-based cues. Most rats shifted to allocentric strategies after training, revealing that strategy selection can depend on training duration and context. This manipulation informed subsequent computational models (Daw, Niv, & Dayan, 2005) about how data must be learned and how strategies transition under different conditions. Lesion studies continued to reveal computational principles, but the initial discovery was behavioral, demonstrating how changing task structure can uncover neural representations and guide subsequent experiments. The discussion then traces a long line of foundational behavioral insights—most fundamental of learning is error-correction. The blocking phenomenon (Kamin, 1968) shows that prediction error, not merely co-occurrence, drives learning; if a light already predicts food, adding a tone to light does not yield learning about the tone. This insight led to the Rescorla-Wagner computational model of learning based on prediction error (Rescorla & Wagner, 1972), which has deeply influenced reinforcement learning theory and neural data interpretation (e.g., dopamine prediction error signals). The narrative connects behavioral findings to computational formalisms (e.g., Barto, Sutton, & Anderson, 1983; Sutton, 1988) and early neural recordings of dopamine neurons during cue-outcome learning (Ljungberg, Apicella, & Schultz, 1992).

This behavioral lead to a normative computational framework—learning driven by prediction errors—that provided precise, testable predictions about both behavior and neural signals. The author places this story within a broader view: behavioral findings anchored by computational models can constrain neural hypotheses, offering a powerful example of how behavioral data and theory can guide neural interpretation. The cascade from behavioral observations to formal models to neural predictions—often described as Marr’s computational level, algorithmic level, and neural implementation—illustrates the central thesis: behavior can lead, while neural data should constrain, not supplant, our understanding of cognition. The section closes by noting that the broader implications of prediction-error theory extend to mental health research, illustrating how impairments in blocking correlate with schizophrenia and how dopamine-related prediction errors relate to symptomatology and pharmacotherapy.

Behavioral experiments, attention, and memory illustrate how rich behavioral data reveal patterns relevant to neural mechanisms without requiring invasive neural measurements. For memory, the retrieval-induced forgetting phenomenon (Anderson, 2003; Anderson, Bjork, & Bjork, 1994; Anderson & Spellman, 1995) shows that competition among memory traces during retrieval can weaken competing memories, with item-specific boundaries and dependence on retrieval competition. This behavioral constraint informs theories of cortical inhibition and hippocampal attractor networks, suggesting oscillatory inhibition as a mechanism to strengthen weak memories while punishing competitors (Norman, Newman, Detre, & Polyn, 2006; Norman, Newman, & Detre, 2007). In attention research, the idea of an attentional spotlight that “blinks” at about 8 Hz (Fiebelkorn, Saalmann, & Kastner, 2013) suggests a rhythm for environmental sampling and constrains models of frontoparietal attention networks. Importantly, the 8 Hz rhythm emerged from behavioral data, not neural measurements in the initial discovery, underscoring the central claim that behavior can reveal principled constraints for neural theories.

In summary, the author argues that while neural measurements illuminate implementation details and localization, the most enduring and impactful insights into cognition have often emerged from behavioral studies. He invites readers to consider what the brain has learned from neural data that behavior had not already taught, concluding that the catalog of such insights is disappointingly short. The broader message is a call for a true merger of psychology and neuroscience, recognizing that computational modeling, behavior, and neural data each contribute to understanding cognition, and that a pure focus on neural data risks neglecting the very questions that behavioral studies uniquely illuminate. The author emphasizes that behavioral experiments, when paired with carefully chosen neural measurements and rigorous computational models, can yield sophisticated causal inferences about brain function without sacrificing depth of understanding. This perspective also highlights the importance of designing task demands that selectively engage specific cognitive processes, thereby turning behavioral measurements into effective causal probes of neural mechanisms.

Clever behavioral experiments allow causal conclusions despite correlative measures
A central claim is that clever behavioral designs enable strong causal inferences even when neural measurements are correlational by nature. The author argues that behavior can reveal the structure and content of cognitive processes through task demands and manipulations that selectively engage those processes. In a key example, the N-back working memory task is used to illustrate how task design can effectively turn “on” or “off” particular cognitive mechanisms. By varying the N parameter (e.g., N = 1 versus N = 2), researchers modulate the maintenance and updating demands of working memory, thereby creating a causal manipulation of the cognitive process. This approach demonstrates how changes in task demands can constrain interpretations about neural substrates and computational mechanisms, while avoiding simplistic one-to-one mappings between brain regions and cognitive functions. The paragraph highlights a broader methodological principle: behavioral perturbations—whether through task structure, stimulus demands, or manipulation of attentional load—can produce clear, interpretable causal effects on behavior that in turn constrain neural theories.

The discussion emphasizes that behavior can be used to isolate cognitive processes in ways that neural perturbations alone may not achieve. For example, to address questions about whether visual and tactile working memory share substrates or rely on distinct neural resources, researchers employ dual-task manipulations and modality-specific load to infer independence or overlap in memory representations. Katus and Eimer (2018) showed that increasing memory load in one modality did not degrade performance in the other, suggesting separate memory stores for visual and tactile information. Such behavioral results guide neuroimaging and neurostimulation studies by identifying whether shared or distinct neural networks should be expected. The author also notes that high-level cognitive questions—such as how attention is deployed or how memory representations are organized—can often be addressed more efficiently and with clearer causal interpretation via behavioral designs than via neural perturbations alone.

The author cautions that neuroscience remains valuable and necessary for certain questions, particularly those about neural implementation, localization, and mechanisms that require direct neural manipulation. However, he argues that the default assumption that neural data alone will yield understanding is misguided. He endorses a collaborative approach, where behavioral paradigms inform theory and computational models, which in turn are tested and refined by neural data. The overarching message is that clever, carefully controlled behavioral experiments can reveal causal structure and computational principles that neural measurements alone may fail to uncover, especially when the goal is to understand cognition in the real world. The N-back example illustrates how task design can recruit specific cognitive processes, enabling causal conclusions about their role in behavior and their neural underpinnings. The broader implication is that the field should value behavioral designs not as a fallback, but as a primary instrument for uncovering cognitive architecture and guiding neural inquiry.

Why do we devalue behavioral work and what should change?
Niv addresses why behavior has been undervalued in neuroscience funding and publication, arguing that there is a pervasive perception that neural data are more objective and theory-driven than behavioral data, which is viewed as subjective or less rigorous. He identifies several misconceptions fueling this bias: neural data are presumed to be inherently more informative about cognition, behavior is thought to be fully solvable or less interesting, and neuroscience is viewed as the sole road to understanding the mind. He notes that this bias parallels the unidirectional influence of machine learning and computational modeling on neuroscience while the reverse influence remains limited. The author cites practical evidence of funding and review patterns—abstracts and grants without neural components may be rejected or deemed less relevant—despite the historical record of behavioral findings driving robust theories about perception, attention, memory, and learning.

Niv argues that the valuation of disciplines should be reoriented toward prioritizing questions rather than techniques. He emphasizes the necessity of a joint, complementary approach in which behavioral paradigms, neural measurements, and computational models are integrated. He underscores that certain domains are best understood through behavioral data alone or primarily through cleverly designed behavioral experiments that yield strong causal inferences about cognitive processes. The argument is not that neuroscientific methods are useless, but that they must be integrated with behavioral approaches to yield meaningful progress. For instance, he notes that while functional neuroimaging can inform questions at a localization level or support computational accounts, many insights about cognition—especially higher-level functions—have historically been achieved through behavioral studies long before, or independently of, neural measurements. The section thus reframes the prestige and utility of behavioral research as foundational to any comprehensive theory of brain function.

The author also explores concrete examples of behavioral phenomena that have guided neural theories and clinical understanding. He discusses face perception and theory of mind as domains in which behavioral work laid the groundwork for later neural studies, while acknowledging that neuroscience can refine our understanding but often cannot replace behavioral insight. He points to memory research and the differentiation of memory systems through dissociations revealed by behavioral paradigms combined with lesion studies, illustrating how behavioral manipulations reveal functional architecture that neuroscience alone cannot fully capture. The central claim is that there is a meaningful asymmetry: behavior contributes more to understanding the mind than neural data contributes to understanding behavior. The call is for a true merger, with behavior serving as the architectural base for cognitive science and neuroscience, rather than as a peripheral component.

Practical and ethical implications
The discussion carries practical and ethical implications for research design, funding priorities, and clinical translation. If behavioral work yields faster, cheaper, and more generalizable insights into cognitive dysfunctions (e.g., bipolar disorder or schizophrenia) than expensive neuroimaging studies, then funding agencies should prioritize well-designed behavioral experiments coupled with computational models before resorting to costly neural measurements. This stance is framed as a moral obligation to optimize resource use for the benefit of subjects, taxpayers, and patients awaiting cures. The author emphasizes that behavior should be leveraged to maximize scientific return and clinical relevance, with neural data providing complementary constraints rather than serving as the sole path to understanding the mind. He also argues for methodological pluralism: combining correlational data with targeted causal perturbations (behaviorally or neurally) to illuminate underlying cognitive structures, while recognizing that some questions may be most efficiently answered through behavior-first approaches.

Bridging theory, behavior, and neuroscience
The concluding sections reiterate the incompatibility of a purely cognitive neuroscience or purely computational approach with robust progress in understanding the mind. The author contends that the field benefits from a synthesis of psychology (cognitive science), computational modeling, and neuroscience. He argues that computational explanations of behavior—often framed in terms of prediction errors, reinforcement learning, and memory dynamics—need to be tested against behavior first and then constrained by neural data. This perspective reframes the scientist’s toolkit: causal behavioral manipulations and well-designed tasks can reveal computational structure and neural constraints in ways that neural perturbations alone cannot replicate. The overarching recommendation is to restore behavior to its rightful place as the foundation for understanding cognition and neural implementation, using behavior as the primary driver of theory and as a rigorous testbed for neural hypotheses. In this view, the best progress emerges from a cooperative, iterative loop among behavioral experiments, computational theories, and neural measurements, rather than from an exclusive focus on any single level of analysis.

Conclusion
The conclusion reinforces the central argument: for advancing understanding of the mind and brain, behavioral research must occupy a central role. While neuroscience can constrain and illuminate neural implementation, it is not sufficient on its own to explain cognitive processes. The brain exhibits redundancy and multiple paths to the same functional outcome, and neural data alone often provide weak constraints on cognitive theories. Behavioral experiments—carefully crafted to isolate processes, manipulate task demands, and generate interpretable outcomes—have historically yielded more robust insight into cognition and should be prioritized to accelerate progress. The paper ends with a call to restore behavior as the solid base on which neuroscience and computational modeling can build, asserting that such a synthesis will better serve scientific understanding and societal needs.

Key concepts, theories, and examples

Behavioral primacy: Behavior often yields sharper constraints on cognitive theories and neural models than neural data alone.
Hierarchy of inquiry: The traditional emphasis on neural measurements can obscure behavioral contributions; reversing this hierarchy improves understanding.
Prediction error and learning: Blocking experiments highlighted a prediction-error-driven learning process, foundational to Rescorla–Wagner models and later reinforcement-learning formalisms. Canonical representations include dopamine-based prediction-error signaling in learning tasks.
Classic experiments and models cited: Kamin (1968) blocking; Rescorla & Wagner (1972) model; Barto, Sutton, & Anderson (1983); Sutton (1988); Ljungberg, Apicella, & Schultz (1992); Waelti, Dickinson, & Schultz (2001); Montague, Dayan, and Sejnowski (1996); Daw, Niv, & Dayan (2005).
Behavioral demonstrations of neural principles: The Packard & McGaugh (1996) T-maze rotation revealed shifts between egocentric and allocentric strategies, informing computational models of strategy selection and hippocampal representations.
Memory and attention: Retrieval-induced forgetting (Anderson et al., 1994; 1995) constrains memory network models; 8 Hz attentional sampling constrains frontoparietal networks; the role of inhibitory oscillations in training neural networks (Norman et al., 2006, 2007).
Working memory: Behavioral and computational work on visual working memory (Luck & Vogel, 1997; Bays & Husain, 2008; Brady, Konkle, & Alvarez, 2011; van den Berg et al., 2012, 2014) illustrates that behavior-driven questions about capacity and organization can yield strong neural hypotheses without invasive measures. Dual-task methods (Katus & Eimer, 2018) reveal modality-specific memory storage.
Theory of mind and domain specificity: Behavioral studies demonstrated early insights into theory of mind and face processing; neural work later identified modulatory regions but behavior often drove the formulation of questions.
Applications to mental health: Blocking deficits link to schizophrenia; differential prediction-error signals relate to learning and motor errors; implications for pharmacological interventions that modulate dopamine signaling. These links illustrate how behavioral principles inform understanding of psychopathology.
Ethical and methodological implications: Prioritizing behavioral research can reduce costs and accelerate discovery and translation, while neural data remain essential for testing implementation-level questions. The synthesis of behavior, computation, and neuroscience is presented as the most productive path forward.

Representative formulas and computational ideas

Temporal-difference prediction error (reinforcement learning):

\deltat = rt + \gamma V(s{t+1}) - V(st)

where r_t is the reward at time t, γ is the discount factor, and V(s) is the value of state s.
Classical Rescorla–Wagner learning rule (prediction error driven):
$<br /> \Delta V = \alpha \beta (\lambda - \sum V)<br />$
where λ is the maximum associative strength, ΣV is the sum of associative strengths, and α, β are learning rate parameters.
Conceptual linkages: Behavioral data constrain the form and parameters of these equations; neural data (e.g., dopamine signals) are interpreted as neural correlates of prediction error under certain conditions.

Connections to prior lectures and real-world relevance

Marr’s levels framework (computational, algorithmic, implementational) is invoked to illustrate how behavioral data guide the formulation of algorithms and predictions that can be tested neurally.
The discussion connects foundational learning theory to clinical phenomena, showing how basic behavioral insights can illuminate mechanisms underlying psychiatric disorders and inform treatment strategies.
The emphasis on designing tasks that selectively engage cognitive processes aligns with best practices in cognitive neuroscience and psychology, illustrating how theory-driven experiments yield more interpretable neural data.

Ethical, philosophical, and practical implications

Ethically, there is a call to optimize resource use by prioritizing behavioral experiments that can yield rapid, cost-effective insights and inform clinical practice, rather than defaulting to costly neural investigations when not strictly necessary.
Philosophically, the piece challenges the notion that neural data are the ultimate arbiter of mind and cognition; instead, it advocates a computationally informed, behavior-first paradigm in which neural data refine but do not replace behavioral explanations.
Practically, the author argues for a true merger of cognitive science and neuroscience, highlighting the complementary strengths of each and warning against an exclusive emphasis on any single data modality.

Equations, numbers, and key references mentioned

Reward prediction error modeling and neural encoding: see δ_t formula above.
Classic learning equations: ΔV in Rescorla–Wagner, as shown above.
Temporal dynamics and frequencies: attention sampling at approximately 8 Hz; the discussion emphasizes the role of rhythmic processes in perception and attention.
Notable references include: Kamin (1968); Rescorla & Wagner (1972); Barto, Sutton, & Anderson (1983); Sutton (1988); Ljungberg, Apicella, & Schultz (1992); Packard & McGaugh (1996); Daw, Niv, & Dayan (2005); McDougle et al. (2016, 2019); Rouhani, Norman, & Niv (2018, 2020); Luck & Vogel (1997); Katus & Eimer (2018); Fiebelkorn, Saalmann, & Kastner (2013).

Concluding synthesis
The overarching message is a call for balance and integration: behavioral data should ground theory and computational models, while neural data should constrain and test those theories. Pure behavioral research is not opposed to neuroscience; rather, it is the bedrock upon which effective neuroscience and computational explanations are built. Restoring behavior to its historical primacy promises faster, cheaper, and more reliable progress in understanding cognition and in translating this understanding to clinical practice. The author’s stance is not anti-neuroscience but pro-integration, arguing that the best progress arises when psychology, computation, and neuroscience work in tandem, using behavior as the firm base for cognitive explanations and neural models.