guest ::
| Home > Publications database > BIO: Leveraging Conscious Dynamics Enhancing ML Performance |
| Home > Publications database > BIO: Leveraging Conscious Dynamics Enhancing ML Performance |
| Poster (After Call) | FZJ-2024-06518 |
; ; ; ;
2024
This record in other databases:
Please use a persistent id in citations: doi:10.34734/FZJ-2024-06518
Abstract: Exploring the Influence of Spontaneous Dynamics on Reservoir Computing PerformanceInsights emerging from the study of brain state-dependent neural dynamics have been slow to influence the optimization and design of Machine Learning (ML) and Artificial Intelligence (AI) algorithms. Analysis of time signals recorded in neuroimaging studies has demonstrated intricate dynamical patterns inherent to conscious brain states, with complexity and proximity to criticality associated with optimal cognitive performance. While substantial investments are made to enhance the performance of ML and AI by designing new hardware resources like GPUs, supercomputers, quantum, and neuromorphic machines, valuable lessons derived from studying the state-dependence of the most efficient computer that we know, the brain, can potentially be better incorporated into ML workflows to ameliorate performance. ML research has begun to achieve improved performance and enhanced learning capabilities using brain-inspired classification algorithms, e.g. programmed ”wake” or ”sleep” states. However, to our knowledge, the implications of neuronal dynamics observed specifically during conscious brain states has not yet been thoroughly examined nor integrated into the design or optimization of ML/AI algorithms. Reservoir Computing (RC) is particularly well-suited for processing time series data, due to its inherent capacity to capture temporal dependencies and dynamics, as its echo state properties amplify and propagate input information over time. To explore the capacity of various dynamical regimes to influence the performance of computations, we utilize RC to solve various cognitive learning tasks implemented in Neurogym (github.com/neurogym). In order to analyze time series data from RC, we implement the Peturbational Complexity Index (PCI) a clinical metric for discriminating various levels of consciousness, and assess the fractal-like dimensions making use of Detrended Fluctuation Analysis (DFA). Furthermore, we analyze the dynamics of the reservoir on exhibiting scale-invariance, sensitivity to perturbations, and operating at a critical state, by computing the Lyapunov exponents. These metrics contribute to our understanding of how reservoir systems adapt and organize information during the acquisition of new tasks. In order to train the RC we make use of high performance computing (HPC) resources and the L2L framework (github.com/Meta-optimization/L2L), a Python implementation of the well known ”Learning to Learn” ML paradigm. Using RC enables the assessment of the complexity and criticality of the signals, and the results can be applied in a reversible manner to guide the network to maintain a specific complexity and operate near criticality. Our preliminary results indicate that, in analogy to enhanced computational properties afforded to the brain during conscious states, ML algorithms parameterized to support near-critical and highly complex operational regimes outperform the same algorithms parameterized with less ”conscious-like” dynamics. This work supports the hypothesis regarding the utility of reverse engineering neural principles derived from the study of brain state-specific dynamics to optimize the computational prowess of current ML/AI algorithms.* 16. Why would attending ASSC benefit your career and the ASSC community? Attending the Association for the Scientific Study of Consciousness (ASSC) conference presents a valuable opportunity to enrich both one's career and the ASSC community. Professionally, participation in ASSC facilitates networking with leading researchers and practitioners in the field of consciousness studies, fostering collaborations and opening avenues for future research endeavors. Engaging with diverse perspectives and cutting-edge research presented at the conference broadens one's knowledge and understanding of consciousness science, contributing to professional growth and development. Furthermore, active involvement in the ASSC community through conference participation, presentations, and discussions allows individuals to contribute their own insights, research findings, and perspectives, thereby enriching the collective knowledge base of the community. By actively participating in ASSC, attendees not only benefit personally and professionally but also contribute to the advancement of consciousness research and the broader scientific community.* 17. How would receipt of the travel award facilitate your attending ASSC? How are you hoping to use the travel award? Receiving the travel award would significantly support my attendance at ASSC by alleviating financial constraints associated with travel expenses and childcare arrangements. The award would enable me to cover the costs of transportation, accommodation, and childcare, making it feasible for me to participate fully in the conference activities. A very short justification on how this work benefits or is linked to the ebrain-Health project The tools used and advocated on this poster presentation are the Peturbational Complexity Index (PCI) a measure used in neuroscience to quantify the complexity (richness and diversity) of brain responses to perturbations or stimuli, and assess the long-range connections via fractal-like dimensions making use of Detrended Fluctuation Analysis (DFA). Furthermore, the dynamics are analyzed on the exhibition of scale-invariance, sensitivity to perturbations, and operating at a critical state, by computing the Lyapunov exponents. The development and utilization of these tools within the project are crucial for enhancing the project's efficacy and achieving its objectives. By enabling the analysis of simulated outputs from The Virtual Brain, these tools provide valuable insights into brain dynamics and their relation to health conditions in the context of the Digital Twin. They facilitate a more thorough examination of the relationships within the simulated functional traces derived from the output of various instances of the model. Additionally, their integration within the workflow of Work Package 8 demonstrates the project's commitment to leveraging advanced computational techniques to advance our understanding of brain health. This strategic use of tools underscores the project's broader goal of improving diagnostics and treatment strategies for neurological disorders. Therefore, their inclusion serves as a vital component in realizing the project's overarching objectives. This work benefits significantly from the discourse at a conference focused on dynamics by offering valuable insights into the effectiveness of various analytical methods as tools for understanding complex dynamic systems. Moreover, it fosters networking opportunities among researchers with shared interests, laying the groundwork for future collaborations and enhancing the project's impact within the scientific community.
|
The record appears in these collections: |
PDF