Sarah Marzen
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Sarah Marzen has a total of 42 co-authors. Here are the top co-authors by number of shared publications:
- James P. Crutchfield
Complexity Sciences Center and Physics Department, University of California at Davis, One Shields Avenue, Davis, CA 95616, USA
24 shared publications Aiden Gao
Affiliation Pioneer Academics, Jenkintown, Pennsylvania, United States of America
13 shared publicationsElianna Schimke
Affiliation Department of Natural Sciences, Scripps and Pitzer Colleges, Claremont, California, United States of America
13 shared publicationsJoost le Feber
Affiliation Department of Clinical Neurophysiology, University of Twente, Enschede, The Netherlands
13 shared publicationsMartina Lamberti
Affiliations Université Claude Bernard Lyon1, CNRS UMR5292, INSERM U1028, Sleep Team, Center for Research in Neuroscience of LYON, Bron, France, Department of Clinical Neurophysiology, University of Twente, Enschede, The Netherlands
13 shared publicationsTristan Kako
Affiliations School of Arts and Sciences, Tufts University, Medford, Massachusetts, United States of America, Pioneer Academics, Jenkintown, Pennsylvania, United States of America
13 shared publications
Latest Publications
- Open Access
- 11 August 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (8) , e0342165
Analyzing an organism’s sensors using Maximum Entropy (MaxEnt) models with bias, variance, and confusion matrices
Biological organisms have sensors that communicate information about the environment. Analyzing how well these biological sensors function has usually been done with mutual information between the sensor signal and the environment, but that can be computationally intractable and summarize something quite complex with just a single number. We suggest that alternatively, one may profitably analyze these biosensors using bias and variance or confusion matrices, depending on the kind of environment. Stimulus-dependent Maximum Entropy models are used to develop estimators of the environmental state given the sensor state, and these estimators in turn are then used to calculate either the bias and variance of the estimator or confusion matrices. We focus on several examples to understand the utility of non-information-based analyses: ligand-receptor binding models spanning from genetic regulation to neuronal communication to bacterial chemotaxis and spin-glass Ising models for neural activity in cultured neurons. These new computationally efficient analyses add insight into existing analyses based on mutual information; in particular, mutual information estimates give one number to characterize responses to all environmental inputs, and this analysis method characterizes how sensors respond to each environmental input. Categorical analyses, meanwhile, indicate the presence of memory without much prediction in confusion matrix elements in cultured neural networks, adding to previous understanding from mutual information estimates.
- 11 August 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (8) , e0342165
Open AccessAnalyzing an organism’s sensors using Maximum Entropy (MaxEnt) models with bias, variance, and confusion matrices
Biological organisms have sensors that communicate information about the environment. Analyzing how well these biological sensors function has usually been done with mutual information between the sensor signal and the environment, but that can be computationally intractable and summarize something quite complex with just a single number. We suggest that alternatively, one may profitably analyze these biosensors using bias and variance or confusion matrices, depending on the kind of environment. Stimulus-dependent Maximum Entropy models are used to develop estimators of the environmental state given the sensor state, and these estimators in turn are then used to calculate either the bias and variance of the estimator or confusion matrices. We focus on several examples to understand the utility of non-information-based analyses: ligand-receptor binding models spanning from genetic regulation to neuronal communication to bacterial chemotaxis and spin-glass Ising models for neural activity in cultured neurons. These new computationally efficient analyses add insight into existing analyses based on mutual information; in particular, mutual information estimates give one number to characterize responses to all environmental inputs, and this analysis method characterizes how sensors respond to each environmental input. Categorical analyses, meanwhile, indicate the presence of memory without much prediction in confusion matrix elements in cultured neural networks, adding to previous understanding from mutual information estimates.
- Open Access
Analyzing an organism’s sensors using Maximum Entropy models with bias, variance, and confusion matrices
Biological organisms have sensors that communicate information about the environment. Analyzing how well these biological sensors function has usually been done with mutual information between the sensor signal and the environment, but that can be computationally intractable and summarize something quite complex with just a single number. We suggest that alternatively, one may profitably analyze these biosensors using bias and variance or confusion matrices, depending on the kind of environment. Stimulus-dependent Maximum Entropy models are used to develop estimators of the environmental state given the sensor state, and these estimators in turn are then used to calculate either the bias and variance of the estimator or confusion matrices. We focus on several examples to understand the utility of non-information-based analyses: ligand-receptor binding models spanning genetic regulation to neuronal communication to bacterial chemotaxis, and spin-glass Ising models for neural activity in cultured neurons. These new computationally-efficient analyses add insight to existing analyses based on mutual information; in particular, mutual information estimates give one number to characterize responses to all environmental inputs, and this analysis method characterizes how sensors respond to each environmental input. Categorical analyses, meanwhile, indicate the presence of memory without much prediction in confusion matrix elements in cultured neural networks, adding to previous understanding from mutual information estimates. Author summary: All living organisms use external stimuli to navigate their environment via their sensors. Because encoding information costs energy, organisms retain only a fraction of the information received from their sensors, ideally information that maximizes their ability to remember past environmental states or predict future ones, key functions that support survival. To better understand how well sensor systems absorb stimulus information, we used stimulus-dependent Maximum Entropy (MaxEnt) models with maximum likelihood estimation and typical statistical metrics, such as confusion matrices or bias and variance. This approach provides two primary benefits over previous approaches: it is more computationally efficient, and it provides a more information-rich picture on how sensors interact with stimulus.
Open AccessAnalyzing an organism’s sensors using Maximum Entropy models with bias, variance, and confusion matrices
Biological organisms have sensors that communicate information about the environment. Analyzing how well these biological sensors function has usually been done with mutual information between the sensor signal and the environment, but that can be computationally intractable and summarize something quite complex with just a single number. We suggest that alternatively, one may profitably analyze these biosensors using bias and variance or confusion matrices, depending on the kind of environment. Stimulus-dependent Maximum Entropy models are used to develop estimators of the environmental state given the sensor state, and these estimators in turn are then used to calculate either the bias and variance of the estimator or confusion matrices. We focus on several examples to understand the utility of non-information-based analyses: ligand-receptor binding models spanning genetic regulation to neuronal communication to bacterial chemotaxis, and spin-glass Ising models for neural activity in cultured neurons. These new computationally-efficient analyses add insight to existing analyses based on mutual information; in particular, mutual information estimates give one number to characterize responses to all environmental inputs, and this analysis method characterizes how sensors respond to each environmental input. Categorical analyses, meanwhile, indicate the presence of memory without much prediction in confusion matrix elements in cultured neural networks, adding to previous understanding from mutual information estimates. Author summary: All living organisms use external stimuli to navigate their environment via their sensors. Because encoding information costs energy, organisms retain only a fraction of the information received from their sensors, ideally information that maximizes their ability to remember past environmental states or predict future ones, key functions that support survival. To better understand how well sensor systems absorb stimulus information, we used stimulus-dependent Maximum Entropy (MaxEnt) models with maximum likelihood estimation and typical statistical metrics, such as confusion matrices or bias and variance. This approach provides two primary benefits over previous approaches: it is more computationally efficient, and it provides a more information-rich picture on how sensors interact with stimulus.
- Open Access
- 1 January 2026
- Published by CourseSource in CourseSource
- Vol. 13
Can Large Language Models Replace a Professor?
QUBES Hub - The Power of Biology, Math, and Community
- 1 January 2026
- Published by CourseSource in CourseSource
- Vol. 13
Open AccessCan Large Language Models Replace a Professor?
QUBES Hub - The Power of Biology, Math, and Community
- 12 December 2025
- Published by The Royal Society in Interface Focus
- Vol. 15 (5)
Resource-rational reinforcement learning and sensorimotor causal states, and resource-rational maximiners
We propose a new computational-level objective function for theoretical biology and theoretical neuroscience that combines: reinforcement learning, the study of learning with feedback via rewards; rate-distortion theory, a branch of information theory that deals with compressing signals to retain relevant information; and computational mechanics, the study of minimal sufficient statistics of prediction, also known as causal states. We highlight why this proposal is likely only an approximation, but is likely to be an interesting one, and propose a new algorithm for evaluating it to obtain the newly coined ‘reward-rate manifold’. The performance of real and artificial agents in partially observable environments can be newly benchmarked using these reward-rate manifolds. To that end, we calculate an example reward-rate manifold utilizing new equations reminiscent of the Blahut–Arimoto algorithm and propose a general algorithm for computing reward-rate manifolds. Finally, we describe experiments that can probe whether or not biological organisms are resource-rational reinforcement learners, using as an example maximin strategies, as bacteria have been shown to be approximate maximiners—doing their best in the worst-case environment, regardless of what is actually happening. This proposal explains why ‘good enough’ for organisms might actually be near-optimal, if viewed correctly.
- 12 December 2025
- Published by The Royal Society in Interface Focus
- Vol. 15 (5)
Resource-rational reinforcement learning and sensorimotor causal states, and resource-rational maximiners
We propose a new computational-level objective function for theoretical biology and theoretical neuroscience that combines: reinforcement learning, the study of learning with feedback via rewards; rate-distortion theory, a branch of information theory that deals with compressing signals to retain relevant information; and computational mechanics, the study of minimal sufficient statistics of prediction, also known as causal states. We highlight why this proposal is likely only an approximation, but is likely to be an interesting one, and propose a new algorithm for evaluating it to obtain the newly coined ‘reward-rate manifold’. The performance of real and artificial agents in partially observable environments can be newly benchmarked using these reward-rate manifolds. To that end, we calculate an example reward-rate manifold utilizing new equations reminiscent of the Blahut–Arimoto algorithm and propose a general algorithm for computing reward-rate manifolds. Finally, we describe experiments that can probe whether or not biological organisms are resource-rational reinforcement learners, using as an example maximin strategies, as bacteria have been shown to be approximate maximiners—doing their best in the worst-case environment, regardless of what is actually happening. This proposal explains why ‘good enough’ for organisms might actually be near-optimal, if viewed correctly.
- 1 October 2025
- Published by AIP Publishing in Chaos: An Interdisciplinary Journal of Nonlinear Science
- Vol. 35 (10)
New analytic formulas for memory and prediction functions in reservoir computers with time delays
Time delays increase the effective dimensionality of reservoirs, thus suggesting that time delays in reservoirs can enhance their performance, particularly their memory and prediction abilities. We find new closed-form expressions for memory and prediction functions of linear time-delayed reservoirs in terms of the power spectrum of the input and the reservoir transfer function. We confirm this relationship numerically for some time-delayed reservoirs using simulations, including when the reservoir can be linearized but is actually nonlinear. Finally, we use these closed-form formulas to address the utility of multiple time delays in linear reservoirs in order to perform memory and prediction, finding similar results to previous work on nonlinear reservoirs. We hope that these closed-form formulas can be used to understand memory and predictive capabilities in time-delayed reservoirs.
- 1 October 2025
- Published by AIP Publishing in Chaos: An Interdisciplinary Journal of Nonlinear Science
- Vol. 35 (10)
New analytic formulas for memory and prediction functions in reservoir computers with time delays
Time delays increase the effective dimensionality of reservoirs, thus suggesting that time delays in reservoirs can enhance their performance, particularly their memory and prediction abilities. We find new closed-form expressions for memory and prediction functions of linear time-delayed reservoirs in terms of the power spectrum of the input and the reservoir transfer function. We confirm this relationship numerically for some time-delayed reservoirs using simulations, including when the reservoir can be linearized but is actually nonlinear. Finally, we use these closed-form formulas to address the utility of multiple time delays in linear reservoirs in order to perform memory and prediction, finding similar results to previous work on nonlinear reservoirs. We hope that these closed-form formulas can be used to understand memory and predictive capabilities in time-delayed reservoirs.
- Open Access
- 2 July 2025
- Published by MDPI AG in Electronics
- Vol. 14 (13) , 2676
A Simple Open-Loop Control Method for Optimizing Manufacturing Control Knobs Using Artificial Intelligence
Manufacturing processes are collecting a wealth of data on how operational knobs affect process efficiency and product quality. Yet, optimizing the adjustment of these knobs using artificial intelligence remains a challenge. We propose a simple open-loop control method for optimizing a manufacturing process, with pharmaceutical applications in mind, using artificial intelligence. The first step involves fitting a simple supervised learning model to manufacturing data—typically an artificial neural network with universal approximation guarantees—so that operational knobs (such as concentrations and temperatures) can be used to predict process efficiency (e.g., time-to-product) and/or product quality (e.g., yield or quality score). Assuming the supervised learning model works well, we can perform typical optimization procedures, like gradient ascent, to increase efficiency and product quality. We test this on a publicly available dataset for wine and suggest new values for wine parameters that should produce a higher-quality wine with a greater probability. The result is a setting for the manufacturing knobs that optimizes the product using basic artificial intelligence. This method can be further enhanced by incorporating more advanced and recent AI applications for anomaly and defect detection in manufacturing processes.
- 2 July 2025
- Published by MDPI AG in Electronics
- Vol. 14 (13) , 2676
Open AccessA Simple Open-Loop Control Method for Optimizing Manufacturing Control Knobs Using Artificial Intelligence
Manufacturing processes are collecting a wealth of data on how operational knobs affect process efficiency and product quality. Yet, optimizing the adjustment of these knobs using artificial intelligence remains a challenge. We propose a simple open-loop control method for optimizing a manufacturing process, with pharmaceutical applications in mind, using artificial intelligence. The first step involves fitting a simple supervised learning model to manufacturing data—typically an artificial neural network with universal approximation guarantees—so that operational knobs (such as concentrations and temperatures) can be used to predict process efficiency (e.g., time-to-product) and/or product quality (e.g., yield or quality score). Assuming the supervised learning model works well, we can perform typical optimization procedures, like gradient ascent, to increase efficiency and product quality. We test this on a publicly available dataset for wine and suggest new values for wine parameters that should produce a higher-quality wine with a greater probability. The result is a setting for the manufacturing knobs that optimizes the product using basic artificial intelligence. This method can be further enhanced by incorporating more advanced and recent AI applications for anomaly and defect detection in manufacturing processes.
- Open Access
Humans are resource-rational predictors in a sequence learning task
Organisms can solve complex tasks despite having limited cognitive resources when those resources are used optimally. Doing so optimally makes an organism “resource-rational”. In this paper, we show for the first time that humans are resource-rational at prediction. In a novel sequence learning experiment, participants predict data generated from hidden Markov models (HMMs) and receive online feedback via clicker training. We compute the predictive rateaccuracy (PRA) curve for each HMM to solve for the highest accuracy achievable for a given cognitive capacity, or “rate”, quantified as the mutual information between participants’ predictions and the underlying causal state of the sequence being predicted. We found that the majority of participants achieve near-optimal prediction with various cognitive capacities, despite performing well below the maximum predictive accuracy on the task overall. We also show that this information-theoretic approach to measuring cognitive capacity can be grounded in the established psychological science concept of working memory: participants who extracted higher quantities of mutual information in the sequence learning task showed significantly higher working memory in a complex digit span test. This research provides new avenues for assessing smart behavior in difficult prediction tasks, provides a new methodology for assessing resource-rationality, and provides evidence that humans are resource-rational predictors.
Open AccessHumans are resource-rational predictors in a sequence learning task
Organisms can solve complex tasks despite having limited cognitive resources when those resources are used optimally. Doing so optimally makes an organism “resource-rational”. In this paper, we show for the first time that humans are resource-rational at prediction. In a novel sequence learning experiment, participants predict data generated from hidden Markov models (HMMs) and receive online feedback via clicker training. We compute the predictive rateaccuracy (PRA) curve for each HMM to solve for the highest accuracy achievable for a given cognitive capacity, or “rate”, quantified as the mutual information between participants’ predictions and the underlying causal state of the sequence being predicted. We found that the majority of participants achieve near-optimal prediction with various cognitive capacities, despite performing well below the maximum predictive accuracy on the task overall. We also show that this information-theoretic approach to measuring cognitive capacity can be grounded in the established psychological science concept of working memory: participants who extracted higher quantities of mutual information in the sequence learning task showed significantly higher working memory in a complex digit span test. This research provides new avenues for assessing smart behavior in difficult prediction tasks, provides a new methodology for assessing resource-rationality, and provides evidence that humans are resource-rational predictors.
- 1 October 2024
- Published by American Association of Physics Teachers (AAPT) in American Journal of Physics
- Vol. 92 (10) , 801-808
Cognitive biases can move opinion dynamics from consensus to signatures of transient chaos
Interest in how democracies form consensus has increased recently, with statistical physics and economics approaches both suggesting that there is convergence to a fixed point in belief networks, but with fluctuations in opinions when there are “stubborn” voters. We modify a model of opinion dynamics in which agents are fully Bayesian to account for two cognitive biases: confirmation bias and in-group bias. Confirmation bias occurs when the received information is considered to be more likely to align with the receiver's beliefs. In-group bias occurs when the receiver further considers the information to be more likely when the receiver's beliefs and the sender's beliefs are aligned. We find that when there are no cognitive biases, a network of agents always converges to complete consensus. With confirmation bias alone, polarization can occur. With both biases present, consensus and polarization are possible, but when agents attempt to counteract confirmation bias, there can be signatures of transient chaos and ongoing opinion fluctuations. Based on this simple model, we conjecture that complex opinion fluctuations might be a generic feature of opinion dynamics when agents are Bayesian with biases.
- 1 October 2024
- Published by American Association of Physics Teachers (AAPT) in American Journal of Physics
- Vol. 92 (10) , 801-808
Cognitive biases can move opinion dynamics from consensus to signatures of transient chaos
Interest in how democracies form consensus has increased recently, with statistical physics and economics approaches both suggesting that there is convergence to a fixed point in belief networks, but with fluctuations in opinions when there are “stubborn” voters. We modify a model of opinion dynamics in which agents are fully Bayesian to account for two cognitive biases: confirmation bias and in-group bias. Confirmation bias occurs when the received information is considered to be more likely to align with the receiver's beliefs. In-group bias occurs when the receiver further considers the information to be more likely when the receiver's beliefs and the sender's beliefs are aligned. We find that when there are no cognitive biases, a network of agents always converges to complete consensus. With confirmation bias alone, polarization can occur. With both biases present, consensus and polarization are possible, but when agents attempt to counteract confirmation bias, there can be signatures of transient chaos and ongoing opinion fluctuations. Based on this simple model, we conjecture that complex opinion fluctuations might be a generic feature of opinion dynamics when agents are Bayesian with biases.
- Open Access
- 16 April 2024
- Published by Springer Nature in Scientific Reports
- Vol. 14 (1) , 1-11
Complexity-calibrated benchmarks for machine learning reveal when prediction algorithms succeed and mislead
Recurrent neural networks are used to forecast time series in finance, climate, language, and from many other domains. Reservoir computers are a particularly easily trainable form of recurrent neural network. Recently, a “next-generation” reservoir computer was introduced in which the memory trace involves only a finite number of previous symbols. We explore the inherent limitations of finite-past memory traces in this intriguing proposal. A lower bound from Fano’s inequality shows that, on highly non-Markovian processes generated by large probabilistic state machines, next-generation reservoir computers with reasonably long memory traces have an error probability that is at least $$\sim 60\%$$ higher than the minimal attainable error probability in predicting the next observation. More generally, it appears that popular recurrent neural networks fall far short of optimally predicting such complex processes. These results highlight the need for a new generation of optimized recurrent neural network architectures. Alongside this finding, we present concentration-of-measure results for randomly-generated but complex processes. One conclusion is that large probabilistic state machines—specifically, large $$\epsilon$$ -machines—are key to generating challenging and structurally-unbiased stimuli for ground-truthing recurrent neural network architectures.- 16 April 2024
- Published by Springer Nature in Scientific Reports
- Vol. 14 (1) , 1-11
Open AccessComplexity-calibrated benchmarks for machine learning reveal when prediction algorithms succeed and mislead
Recurrent neural networks are used to forecast time series in finance, climate, language, and from many other domains. Reservoir computers are a particularly easily trainable form of recurrent neural network. Recently, a “next-generation” reservoir computer was introduced in which the memory trace involves only a finite number of previous symbols. We explore the inherent limitations of finite-past memory traces in this intriguing proposal. A lower bound from Fano’s inequality shows that, on highly non-Markovian processes generated by large probabilistic state machines, next-generation reservoir computers with reasonably long memory traces have an error probability that is at least $$\sim 60\%$$ higher than the minimal attainable error probability in predicting the next observation. More generally, it appears that popular recurrent neural networks fall far short of optimally predicting such complex processes. These results highlight the need for a new generation of optimized recurrent neural network architectures. Alongside this finding, we present concentration-of-measure results for randomly-generated but complex processes. One conclusion is that large probabilistic state machines—specifically, large $$\epsilon$$ -machines—are key to generating challenging and structurally-unbiased stimuli for ground-truthing recurrent neural network architectures. - 19 December 2023
- Published by Elsevier in BioSystems
- Vol. 235, 105107
From reinforcement learning to agency: Frameworks for understanding basal cognition
- 19 December 2023
- Published by Elsevier in BioSystems
- Vol. 235, 105107
From reinforcement learning to agency: Frameworks for understanding basal cognition