Tavoni Lab | Google Scholar Profile
Research
The Tavoni Laboratory investigates the fundamental principles underlying brain function, from sensing the environment to forming memories and making decisions. We study how information is represented and processed within brain networks to optimize behavior in diverse environments. Our theories provide normative predictions about the organization and function of neural circuits, and we integrate these theories with biophysically inspired models and neural data. We strive to develop unified frameworks that connect computational, algorithmic, and implementational perspectives. Ultimately, uncovering the principles of neural coding is essential for understanding disruptions in brain diseases.
Areas of focus in the lab include:
Unveiling neural coding principles via unified multi-level theories
Animals constantly receive sensory stimuli, which are processed by neural networks to generate behavior. While sensory and decision-making circuits are typically studied separately, they work synergistically in the brain: different sensory modalities influence each other, and behavioral goals shape representations at all levels, from the sensory periphery to decision-making centers. We are interested in understanding the principles of neural coding in interacting networks.
Coding principles in multisensory and sensorimotor networks: toward a unified theory of efficient and predictive distributed coding
Recently, we developed a normative coding theory for multimodal canonical circuits—a ubiquitous excitatory-inhibitory motif in the brain. In these networks, cross-modal feedback transmits information between sensory modalities or from motor to sensory systems. The theory provides a unified perspective on efficient and predictive coding, computational and algorithmic principles, and unimodal and multimodal sensory processing [Tavoni, BioRxiv 2025; revised version forthcoming].
First, we showed that in multimodal canonical networks, efficient coding automatically generates predictive codes, in which stimuli from one modality can serve as predictors for stimuli in another modality. Cross-modal predictive computations are an emergent property of context-sensitive efficient codes. Second, we showed how different network components can implement these computations algorithmically, providing an alternative view of predictive coding distinct from classical hierarchical generative models. Third, we showed that this theoretical framework accounts for diverse multimodal phenomena, including audiovisual, visual–olfactory, auditory–somatosensory, and sensorimotor interactions, while recovering classical unimodal coding as a limiting case.
We are now focused on generalizing this framework to other classes of interacting networks and additional contextual effects on neural coding.
From coding principles to their implementation via neuroplasticity
The brain continually adapts its structure to optimally represent and respond to environmental inputs. An active area of research in the lab focuses on uncovering how various mechanisms, including synaptic plasticity, neurogenesis, and neuromodulatory processes contribute to shaping optimal neural codes. These efforts aim to connect computational and implementational perspectives on neural coding.
Network analyses of neural circuits for economic decisions
The goal of this project is to elucidate the principles governing how the values of goods are encoded in the brain and compared within neural circuits to drive economic decisions. We use statistical physics methods grounded in the inference and simulation of maximum-entropy models to reconstruct and analyze functional circuits in the orbitofrontal cortex (OFC), a region that plays a central role in economic decision-making. In parallel, we develop and test theories of how these circuits support the optimal coding of values and choices. Because economic choice behavior is disrupted in a range of clinical disorders, understanding the neural mechanisms underlying these processes is important from both scientific and medical perspectives.
Bayesian and complexity theories of high-level cognition
We continuously gather noisy data through our senses to make inferences about past, present, and future states of the world. Accessible information, time and resources are limited and constrain the accuracy and complexity of viable inference strategies. The lab develops normative theories to understand how efficient inference processes adapt their complexity to environmental uncertainty and task demands [Tavoni, Doi, Pizzica, Balasubramanian, Gold, Nat Hum Behav 2022]. We identified a hierarchical (nested) organization in a wide range of models, from Bayesian probabilistic strategies to simpler, heuristic update processes that are often described as implementing a ‘model-free’ form of learning. By studying this hierarchy, we identified two fundamental principles: (a) a power law of diminishing returns, whereby increasing computational complexity and cognitive effort gives progressively smaller gains in accuracy, and (b) a non-monotonic relationship between cognitive demands and statistical uncertainty in the world, such that complex inference strategies are necessary only in a relatively narrow range of intermediate-noise environments.
Memory optimization in brain-inspired heterogeneous networks
To acquire knowledge, the brain must be able to store and retrieve information efficiently. In this project, we investigate how neural networks can optimize memory capacity. Previous work has primarily focused on estimating the capacity of homogeneous Hopfield networks, but biological neural networks are highly heterogeneous.
In a recent study [Zhang, Tavoni, PRX Life 2025], we used methods from the statistical mechanics of spin glasses to calculate the memory capacity of a broader class of brain-inspired networks with heterogeneous connectivity and arbitrary coding levels (i.e., varying neuron activation rates in stored patterns). This framework allowed us to derive normative predictions for the conditions that optimize memory function in these networks.
In particular, we predicted that the number of inward connections and neuronal coding levels should be correlated to maximize capacity. This prediction held across multiple biologically relevant scenarios, including the storage of independent patterns in both classical and dendritic networks, as well as the storage of example patterns clustered around prototypes representing concepts. In the latter case, heterogeneity similarly influenced the capacity for both examples and concepts.
Finally, we analyzed bipartite models of the CA3–DG hippocampal circuit, which is known to play a central role in memory encoding. In these networks, capacity was maximized by a quasi-indexing encoding scheme, in which each dentate gyrus (DG) neuron binds subsets of features from a few memory patterns stored in CA3. Compared to a complete-indexing scheme, where each DG neuron binds all features of a single pattern, quasi-indexing substantially improved both memory capacity and robustness to neuronal ablation. These findings extend the hippocampal indexing theory of memory encoding and retrieval while suggesting specific neural substrates for these functions.
Ultimately, understanding the fundamental principles governing memory function is important for identifying therapeutic targets for Alzheimer’s disease and other conditions associated with memory decline.
If you are interested in joining the lab as a postdoc or graduate student, please email Gaia Tavoni at gaia.tavoni@wustl.edu.
Selected publications and preprints
- Tavoni G. Convergence of efficient and predictive coding in multimodal sensory processing. BioRxiv. 2025. doi: 10.1101/2025.02.24.639817.
- Zhang K, Tavoni G. Maximizing memory capacity in heterogeneous networks. PRX Life. June 20, 2025; 3(2), 023016.
- Tavoni G, Doi T, Pizzica C, Balasubramanian V, Gold JI. Human inference reflects a normative balance of complexity and accuracy. Nature Human Behaviour. May 30, 2022; doi: 10.1038/s41562-022-01357-z.
- Tavoni G, Kersen DEC, Balasubramanian V. Cortical feedback and gating in odor discrimination and generalization. PLoS Computational Biology. Oct 11, 2021; 17(10): e1009479. doi: 10.1371/journal.pcbi.1009479.
- Tavoni G, Balasubramanian V, Gold JI. What is optimal in optimal inference? Current Opinion in Behavioral Sciences. 2019; 29:117-126.
- Tavoni G, Ferrari U, Battaglia FP, Cocco S, Monasson R. Functional coupling networks inferred from prefrontal cortex activity show experience-related effective plasticity. Network Neuroscience. 2017; 1(3):275-301.
- Cocco S, Monasson R, Posani L, Tavoni G. Functional networks from inverse modeling of neural population activity. Current Opinion in Systems Biology. 2017; 3:103-110.
- Tavoni G, Cocco S, Monasson R. Neural assemblies revealed by inferred connectivity-based models of prefrontal cortex recordings. Journal of Computational Neuroscience. 2016; 41(3): 269-293.
Education
2010, BS in Physical Engineering, Polytechnic of Turin
2012, International MS in Physics of Complex Systems, International university consortium (Polytechnic of Turin, International School for Advanced Studies and International Centre for Theoretical Physics of Trieste, Universities Pierre & Marie Curie, Paris Diderot, Paris-Sud and École Normale Supérieure at Cachan)
2015, PhD in Physics, École Normale Supérieure (Paris), Laboratories of Statistical and Theoretical Physics
2020, Post-doc in Theoretical Neuroscience, University of Pennsylvania, Computational Neuroscience Initiative
Selected honors
2024, Sloan Fellowship Award
2024, PRX Life Reviewer Excellence Award
2017, Swartz Foundation Fellowship Award for Theory in Neuroscience
2015, Computational Neuroscience Initiative Postdoctoral Fellowship
2010, admitted to the Alta Scuola Politecnica (declined)