Postdoctoral Fellow-MSH-32910-009

Salary: $80,000 - $85,000

Department: Windreich Department of AI & Human Health

Physical work location: The Hamilton and Amabel James Center for Artificial Intelligence and Human Health, 3 E. 101st Street, New York, NY 10029

Name PI or Supervisor: Baihan Lin / baihan.lin@mssm.edu / (206) 915-1164

Web link to Lab: https://www.hpims.org/team/baihan-lin/

Web link to Department: https://icahn.mssm.edu/about/departments-offices/ai-human-health

Administrative Contact: Sparshdeep Kaur / sparshdeep.kaur@mssm.edu / (347) 913-2168

Details of Research Project: 

We are seeking a motivated and talented Postdoctoral Researcher to join a research team focused on advancing our understanding of learning and generalization mechanisms across biological and artificial neural networks. This interdisciplinary project leverages insights from neuroscience, cognitive science, and artificial intelligence to explore universal principles of generalization. The successful candidate will contribute to the development of theoretical frameworks and conduct empirical research aimed at identifying commonalities in generalization processes across diverse neural systems.

Technical Duties: (include any protocols)

• Theoretical and Empirical Research: Design and execute studies using cutting-edge methods from topological data analysis and information theory to investigate neural and artificial network generalization.
• Cross-System Analysis: Engage in comparative research across biological systems (human and animal models) and artificial neural networks to uncover universal patterns of information processing.
• Data Analysis and Model Development: Utilize advanced computational techniques to analyze data from multiple modalities and develop models that generalize across different architectures and species.
• Collaboration and Mentorship: Work collaboratively within an interdisciplinary team, and contribute to mentoring junior researchers, including graduate students and research assistants.

Educational and other Requirements for the position:

Ph.D. in Neuroscience, Cognitive Science, Computer Science, or a related field.

Experience Required:

• Research Experience: Strong background in theoretical neuroscience, machine learning, or AI, with experience in neural data analysis or computational modeling.
• Technical Skills: Proficiency in programming languages such as Python or MATLAB, and experience with deep learning frameworks (e.g., TensorFlow, PyTorch) is highly desirable.
• Analytical Abilities: Excellent analytical skills, with experience in topological data analysis, information theory, or related computational methods.
• Communication Skills: Strong written and verbal communication skills, with a proven track record of research publications.

Goals/Outcomes of the Research Project: 

The ability to learn and generalize from a few examples is one of the most fundamental human cognitive abilities and lies at the core of the theoretical neuroscience field investigating cognitive processes in the brain. Generally regarded as one of the foundational building blocks of intelligence and the most sought-after ability to be replicated in artificial intelligence systems, generalization underlies our essential capacity to learn new information. Thus, the implications of understanding generalization mechanisms in the brain are vast. To name a few, it is central to advancing the treatment of prominent neurological disorders, like ADHD, where information generalization seems to be impaired and currently affects 7 million (11.4%) of children 3-17 years in the US. It is also crucial for the development of early biomarkers in neurodegenerative diseases like Alzheimer’s and Parkinson’s by identifying how the generalization system breaks down.

Despite a wealth of theories and decades of research efforts in cognitive science and neuroscience, the precise nature and underlying mechanisms of generalization remain unknown. And yet, successful generalization can be observed across various neural information processing systems, from humans to animals, and artificial neural networks (ANNs). Importantly, these similarities prevail despite key differences in the systems’ architecture and structure. ANNs for example – although designed as brain-inspired processing systems - employ biologically implausible training and optimization strategies such as backpropagation and regularization. Animal brains, on the other hand, share a great deal of evolutionary roots and neural structures with human brains, but still exhibit less flexibility and tend to rely on more perceptual and contextual clues to learn and generalize. Its ubiquitous presence across neural networks with such different architectures and nevertheless generating similar behavioral responses offer an avenue to detect common patterns that lie at the core of generalization mechanisms of the brain. Therefore, the main goal of this research proposal is to investigate and identify universal principles of generalization that are shared across biological and artificial neural network systems and ultimately advance both the fields of theoretical neuroscience and artificial intelligence. 

Consequently, this research proposes in Objective 1 to develop a theoretical framework that can uncover universal mechanisms of generalization. This will be accomplished by (1) using information theoretical insights with cutting-edge developments in topological data analysis (TDA) and topological deep learning (TDL) to create a theoretical and empirically testable framework with distinct comparative metrics that are invariant to and can abstract away from the system-specific architecture and signal; and (2) by employing a multisystem comparative approach across brains and ANNs, species (humans vs. primates vs. mice), and ANN architectures (image vs. graph networks). Here, the use of ANNs is crucial to the mission of identifying universal principles of generalization in biological systems, since the machine learning (ML) literature offers great insights into the model’s internal dynamics while also having more control over its parameters. Finding the commonalities of biological and artificial NNs could thus help identify universal patterns. 

Key to this quest is a mutual cross-fertilization of recent scientific breakthroughs in the fields of cognitive science, neuroscience, and AI that offer new explanations and potentially universal characteristics of generalization. The first breakthrough comes from cognitive science, using information theory to extend Roger Shepard’s Universal Law of Generalization, a foundational theory in cognitive science that describes how organisms generalize from one stimulus to another. Based on similarity relations between stimuli, this law formalizes the idea that generalization probability decreases exponentially as the psychological distance between stimuli increases. Recently, Sims proposed an extension and further generalization of Shepard’s law based on information theory. Since its inception, information theory and the framework of viewing the brain as a distributed information processing system has found important applications in neuroscience. Within the overall goal of information theory, which is quantifying, encoding, transmitting, and interpreting information efficiently and accurately, the concepts of efficient coding and information compression are the most relevant to the study of generalization. The representations resulting from the process of generalization are an integral part of achieving communication (e.g. information retrieval) as they facilitate fundamental cognitive functions such as categorization (information compression and abstraction) and object recognition and retrieval (information transmission). To fulfill these functions, the information contained within the representation needs to be reconstructable. A solution to this comes from rate-distortion theory, which provides the foundations for determining the maximum loss of data compression so that the resulting information can still be reconstructed. Indeed, a recent breakthrough in cognitive science by Sims empirically demonstrates that the principles of efficient coding and rate-distortion theory can explain and successfully model human generalization behavior in various datasets. In fact, human perceptual generalization shows a near perfect empirical fit when modeled with rate-distortion theory. However, its explanatory power has yet to be proven in data from other than human behavior. Thus, Objective 2 of this research is to empirically model perceptual generalization data obtained from humans, animals, and ANNs using rate-distortion theory to test its validity as a universal principle. 

A second ground-breaking discovery comes from the fields of neuroscience and AI. Recent advances in ML have combined learning theory and TDA to demonstrate a direct link between a deep neural network’s (DNN) generalization performance and the topology of its representations. This was done by successfully showing that it is possible to predict the DNN’s generalization error by computing the network’s intrinsic dimension over the course of learning. This is because the learning trajectory of DNNs contains fractal structures, whose complexity can be formally linked to the network’s generalization error. In a novel approach, Birdal and colleagues proved that the generalization error can be equivalently bounded in terms of what is called the PHD – an intrinsic dimensionality metric - and developed an algorithm to estimate the PHD from the generalization error. While these findings have important theoretical and practical implications for ML, they are also consistent with novel discoveries in neuroscience demonstrating a positive link between forming a decision variable (indicative of successful learning) and decreased representational dimensionality. Taking this parallel between brains and machines as a starting point, Objective 3 will estimate the PHD of the generalization error in human, animal, and ANN data to evaluate the universal boundedness of the generalization error and representational topology. The application of these findings from learning theory to the brain come at a time where there has been a recent shift in computational neuroscience towards investigating neural and ANN representations using neural geometric patterns and studying them under the manifold hypothesis. This hypothesis posits that there is a low-dimensional subspace that underlies neural activities of a high-dimensional neural state space. Relatedly, with the formalization of the concept of PHD, TDA – previously constraint to theoretical mathematics – started finding applications in biology and ML but did not reach neuroscience until more recently. 

To accomplish this proposal’s ambitious multi-systems comparison goal, I propose a theoretical framework for modeling generalization behavioral across biological and artificial neural networks (brains vs. ANNs), species (humans vs. non-human primates vs. mice) and ANN architectures (image vs. graph networks). For this, I will take advantage of a wealth of data made available via open science practices outlined in Table 1. This data consists of ten behavioral and neural datasets of visual discrimination tasks – a well-established paradigm to investigate generalization performance - performed by primates (humans and macaque monkeys) and rodents (mice), as well as the use of five ANNs of various architectures (image networks and graph networks) performing the same visual discrimination tasks. 

Taken together, this project will leverage a three-way transfer of information across disciplines: Theoretical findings from cognitive sciences will be applied to answer questions in neuroscience and AI, while breakthroughs in AI and TDL will innovatively be expanded to examine behavior and dynamics of biological systems studied in neuroscience and cognitive science. Drawing on such fundamental theories and novel breakthroughs in these fields will make it possible to inspect generalization processes in a novel way by abstracting away from the neural signal and architecture-specific constraints to, which is crucial for a multi-system comparison, gaining insight into shared principles and divergences between the different types of systems (humans, animals, and ANNs), and the discovery of universal principles that govern the building blocks of natural and artificial intelligence.

Compensation Statement

The Mount Sinai Health System (MSHS) provides a salary range to comply with the New York City Law on Salary Transparency in Job Advertisements. The salary range for this role is $73,588.00 - $80,000.00 Annually. Actual salaries depend on a variety of factors, including experience, education, and hospital need. The salary range or contractual rate listed does not include bonuses/incentive, differential pay or other forms of compensation or benefits.

SPOC-UAW Local 4100 at Icahn School of Medicine (Post Docs), J18 - AI and Human Health - ISM, Icahn School of Medicine

EOE Minorities/Women/Disabled/Veterans

The Mount Sinai Health System is an equal opportunity employer. We comply with applicable Federal civil rights laws and does not discriminate, exclude, or treat people differently on the basis of race, color, national origin, age, religion, disability, sex, sexual orientation, gender identity, or gender expression. We are passionately committed to addressing racism and its effects on our faculty, staff, students, trainees, patients, visitors, and the communities we serve. Our goal is for Mount Sinai to become an anti-racist health care and learning institution that intentionally addresses structural racism.”

Strength Through Diversity

The Mount Sinai Health System believes that diversity, equity, and inclusion are key drivers for excellence. We share a common devotion to delivering exceptional patient care. When you join us, you become a part of Mount Sinai’s unrivaled record of achievement, education, and advancement as we revolutionize medicine together. We invite you to participate actively as a part of the Mount Sinai Health System team by:  

• Using a lens of equity in all aspects of patient care delivery, education, and research to promote policies and practices to allow opportunities for all to thrive and reach their potential.
• Serving as a role model confronting racist, sexist, or other inappropriate actions by speaking up, challenging exclusionary organizational practices, and standing side-by-side in support of colleagues who experience discrimination.
• Inspiring and fostering an environment of anti-racist behaviors among and between departments and co-workers.

At Mount Sinai, our leaders strive to learn, empower others, and embrace change to further advance equity and improve the well-being of staff, patients, and the organization.  We expect our leaders to embrace anti-racism, create a collaborative and respectful environment, and constructively disrupt the status quo to improve the system and enhance care for our patients. We work hard to create an inclusive, welcoming and nurturing work environment where all feel they are valued, belong and are able to advance professionally.

 Explore more about this opportunity and how you can help us write a new chapter in our history!

“About the Mount Sinai Health System:

Mount Sinai Health System is one of the largest academic medical systems in the New York metro area, with more than 43,000 employees working across eight hospitals, more than 400 outpatient practices, more than 300 labs, a school of nursing, and a leading school of medicine and graduate education. Mount Sinai advances health for all people, everywhere, by taking on the most complex health care challenges of our time — discovering and applying new scientific learning and knowledge; developing safer, more effective treatments; educating the next generation of medical leaders and innovators; and supporting local communities by delivering high-quality care to all who need it. Through the integration of its hospitals, labs, and schools, Mount Sinai offers comprehensive health care solutions from birth through geriatrics, leveraging innovative approaches such as artificial intelligence and informatics while keeping patients’ medical and emotional needs at the center of all treatment. The Health System includes approximately 7,400 primary and specialty care physicians; 13 joint-venture outpatient surgery centers throughout the five boroughs of New York City, Westchester, Long Island, and Florida; and more than 30 affiliated community health centers. We are consistently ranked by U.S. News & World Report's Best Hospitals, receiving high "Honor Roll" status, and are highly ranked: No. 1 in Geriatrics and top 20 in Cardiology/Heart Surgery, Diabetes/Endocrinology, Gastroenterology/GI Surgery, Neurology/Neurosurgery, Orthopedics, Pulmonology/Lung Surgery, Rehabilitation, and Urology. New York Eye and Ear Infirmary of Mount Sinai is ranked No. 12 in Ophthalmology. U.S. News & World Report’s “Best Children’s Hospitals” ranks Mount Sinai Kravis Children's Hospital among the country’s best in several pediatric specialties. The Icahn School of Medicine at Mount Sinai is ranked No. 14 nationwide in National Institutes of Health funding and in the 99th percentile in research dollars per investigator according to the Association of American Medical Colleges. Newsweek’s “The World’s Best Smart Hospitals” ranks The Mount Sinai Hospital as No. 1 in New York and in the top five globally, and Mount Sinai Morningside in the top 20 globally.

The Mount Sinai Health System is an equal opportunity employer. We comply with applicable Federal civil rights laws and does not discriminate, exclude, or treat people differently on the basis of race, color, national origin, age, religion, disability, sex, sexual orientation, gender identity, or gender expression. We are passionately committed to addressing racism and its effects on our faculty, staff, students, trainees, patients, visitors, and the communities we serve. Our goal is for Mount Sinai to become an anti-racist health care and learning institution that intentionally addresses structural racism.”

EOE Minorities/Women/Disabled/Veterans