1
|
Zhu V, Rosenbaum R. Learning Fixed Points of Recurrent Neural Networks by Reparameterizing the Network Model. Neural Comput 2024; 36:1568-1600. [PMID: 39028956 DOI: 10.1162/neco_a_01681] [Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [MESH Headings] [Track Full Text] [Journal Information] [Subscribe] [Scholar Register] [Received: 08/30/2023] [Accepted: 03/18/2024] [Indexed: 07/21/2024]
Abstract
In computational neuroscience, recurrent neural networks are widely used to model neural activity and learning. In many studies, fixed points of recurrent neural networks are used to model neural responses to static or slowly changing stimuli, such as visual cortical responses to static visual stimuli. These applications raise the question of how to train the weights in a recurrent neural network to minimize a loss function evaluated on fixed points. In parallel, training fixed points is a central topic in the study of deep equilibrium models in machine learning. A natural approach is to use gradient descent on the Euclidean space of weights. We show that this approach can lead to poor learning performance due in part to singularities that arise in the loss surface. We use a reparameterization of the recurrent network model to derive two alternative learning rules that produce more robust learning dynamics. We demonstrate that these learning rules avoid singularities and learn more effectively than standard gradient descent. The new learning rules can be interpreted as steepest descent and gradient descent, respectively, under a non-Euclidean metric on the space of recurrent weights. Our results question the common, implicit assumption that learning in the brain should be expected to follow the negative Euclidean gradient of synaptic weights.
Collapse
Affiliation(s)
- Vicky Zhu
- Babson College, Mathematics, Analytics, Science, and Technology Division, Wellesley, MA 02481, U.S.A.
| | - Robert Rosenbaum
- University of Notre Dame, Department of Applied and Computational Mathematics and Statistics, Notre Dame, IN 46556, U.S.A.
| |
Collapse
|
2
|
Luppi AI, Liu ZQ, Milisav F, Bazinet V, Hansen J, Misic B. From abstract networks to biological realities. Phys Life Rev 2024; 49:12-14. [PMID: 38471192 DOI: 10.1016/j.plrev.2024.02.009] [Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [MESH Headings] [Track Full Text] [Journal Information] [Subscribe] [Scholar Register] [Received: 02/26/2024] [Accepted: 02/28/2024] [Indexed: 03/14/2024]
Affiliation(s)
- Andrea I Luppi
- Montreal Neurological Institute, McGill University, Montreal, QC, Canada.
| | - Zhen-Qi Liu
- Montreal Neurological Institute, McGill University, Montreal, QC, Canada
| | - Filip Milisav
- Montreal Neurological Institute, McGill University, Montreal, QC, Canada
| | - Vincent Bazinet
- Montreal Neurological Institute, McGill University, Montreal, QC, Canada
| | - Justine Hansen
- Montreal Neurological Institute, McGill University, Montreal, QC, Canada
| | - Bratislav Misic
- Montreal Neurological Institute, McGill University, Montreal, QC, Canada
| |
Collapse
|
3
|
Vinograd A, Nair A, Linderman SW, Anderson DJ. Intrinsic Dynamics and Neural Implementation of a Hypothalamic Line Attractor Encoding an Internal Behavioral State. BIORXIV : THE PREPRINT SERVER FOR BIOLOGY 2024:2024.05.21.595051. [PMID: 38826298 PMCID: PMC11142118 DOI: 10.1101/2024.05.21.595051] [Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [Grants] [Track Full Text] [Subscribe] [Scholar Register] [Indexed: 06/04/2024]
Abstract
Line attractors are emergent population dynamics hypothesized to encode continuous variables such as head direction and internal states. In mammals, direct evidence of neural implementation of a line attractor has been hindered by the challenge of targeting perturbations to specific neurons within contributing ensembles. Estrogen receptor type 1 (Esr1)-expressing neurons in the ventrolateral subdivision of the ventromedial hypothalamus (VMHvl) show line attractor dynamics in male mice during fighting. We hypothesized that these dynamics may encode continuous variation in the intensity of an internal aggressive state. Here, we report that these neurons also show line attractor dynamics in head-fixed mice observing aggression. We exploit this finding to identify and perturb line attractor-contributing neurons using 2-photon calcium imaging and holographic optogenetic perturbations. On-manifold perturbations demonstrate that integration and persistent activity are intrinsic properties of these neurons which drive the system along the line attractor, while transient off-manifold perturbations reveal rapid relaxation back into the attractor. Furthermore, stimulation and imaging reveal selective functional connectivity among attractor-contributing neurons. Intriguingly, individual differences among mice in line attractor stability were correlated with the degree of functional connectivity among contributing neurons. Mechanistic modelling indicates that dense subnetwork connectivity and slow neurotransmission are required to explain our empirical findings. Our work bridges circuit and manifold paradigms, shedding light on the intrinsic and operational dynamics of a behaviorally relevant mammalian line attractor.
Collapse
Affiliation(s)
- Amit Vinograd
- Division of Biology and Biological Engineering, California Institute of Technology; Pasadena, USA
- Tianqiao and Chrissy Chen Institute for Neuroscience Caltech; Pasadena, USA
- Howard Hughes Medical Institute; Chevy Chase, USA
| | - Aditya Nair
- Division of Biology and Biological Engineering, California Institute of Technology; Pasadena, USA
- Tianqiao and Chrissy Chen Institute for Neuroscience Caltech; Pasadena, USA
- Howard Hughes Medical Institute; Chevy Chase, USA
| | - Scott W. Linderman
- Department of Statistics, Stanford University, Stanford, USA
- Wu Tsai Neurosciences Institute, Stanford University, Stanford, USA
| | - David J. Anderson
- Division of Biology and Biological Engineering, California Institute of Technology; Pasadena, USA
- Tianqiao and Chrissy Chen Institute for Neuroscience Caltech; Pasadena, USA
- Howard Hughes Medical Institute; Chevy Chase, USA
| |
Collapse
|
4
|
Jiang D, Wang K, Li H, Zhang Y. Efficient Near-Infrared Spectrum Detection in Nondestructive Wood Testing via Transfer Network Redesign. SENSORS (BASEL, SWITZERLAND) 2024; 24:1245. [PMID: 38400402 PMCID: PMC10893441 DOI: 10.3390/s24041245] [Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [Key Words] [Track Full Text] [Subscribe] [Scholar Register] [Received: 01/08/2024] [Revised: 02/11/2024] [Accepted: 02/11/2024] [Indexed: 02/25/2024]
Abstract
This study systematically developed a deep transfer network for near-infrared spectrum detection using convolutional neural network modules as key components. Through meticulous evaluation, specific modules and structures suitable for constructing the near-infrared spectrum detection model were identified, ensuring its effectiveness. This study extensively analyzed the basic network components and explored three unsupervised domain adaptation structures, highlighting their applications in the nondestructive testing of wood. Additionally, five transfer networks were strategically redesigned to substantially enhance their performance. The experimental results showed that the Conditional Domain Adversarial Network and Globalized Loss Optimization Transfer network outperformed the Direct Standardization, Piecewise Direct Standardization, and Spectral Space Transformation models. The coefficients of determination for the Conditional Domain Adversarial Network and Globalized Loss Optimization Transfer network are 82.11% and 83.59%, respectively, with root mean square error prediction values of 12.237 and 11.582, respectively. These achievements represent considerable advancements toward the practical implementation of an efficient and reliable near-infrared spectrum detection system using a deep transfer network.
Collapse
Affiliation(s)
- Dapeng Jiang
- College of Computer and Control Engineering, Northeast Forestry University, 26 Hexing Rd., Harbin 150040, China; (D.J.); (K.W.)
| | - Keqi Wang
- College of Computer and Control Engineering, Northeast Forestry University, 26 Hexing Rd., Harbin 150040, China; (D.J.); (K.W.)
| | - Hongbo Li
- College of Electrical and Information, Northeast Agricultural University, Harbin 150030, China;
| | - Yizhuo Zhang
- College of Computer Science and Artificial Intelligence, Changzhou University, 1 Gehu Middle Rd., Changzhou 213164, China
| |
Collapse
|
5
|
Pals M, Macke JH, Barak O. Trained recurrent neural networks develop phase-locked limit cycles in a working memory task. PLoS Comput Biol 2024; 20:e1011852. [PMID: 38315736 PMCID: PMC10868787 DOI: 10.1371/journal.pcbi.1011852] [Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [MESH Headings] [Grants] [Track Full Text] [Journal Information] [Subscribe] [Scholar Register] [Received: 04/13/2023] [Revised: 02/15/2024] [Accepted: 01/22/2024] [Indexed: 02/07/2024] Open
Abstract
Neural oscillations are ubiquitously observed in many brain areas. One proposed functional role of these oscillations is that they serve as an internal clock, or 'frame of reference'. Information can be encoded by the timing of neural activity relative to the phase of such oscillations. In line with this hypothesis, there have been multiple empirical observations of such phase codes in the brain. Here we ask: What kind of neural dynamics support phase coding of information with neural oscillations? We tackled this question by analyzing recurrent neural networks (RNNs) that were trained on a working memory task. The networks were given access to an external reference oscillation and tasked to produce an oscillation, such that the phase difference between the reference and output oscillation maintains the identity of transient stimuli. We found that networks converged to stable oscillatory dynamics. Reverse engineering these networks revealed that each phase-coded memory corresponds to a separate limit cycle attractor. We characterized how the stability of the attractor dynamics depends on both reference oscillation amplitude and frequency, properties that can be experimentally observed. To understand the connectivity structures that underlie these dynamics, we showed that trained networks can be described as two phase-coupled oscillators. Using this insight, we condensed our trained networks to a reduced model consisting of two functional modules: One that generates an oscillation and one that implements a coupling function between the internal oscillation and external reference. In summary, by reverse engineering the dynamics and connectivity of trained RNNs, we propose a mechanism by which neural networks can harness reference oscillations for working memory. Specifically, we propose that a phase-coding network generates autonomous oscillations which it couples to an external reference oscillation in a multi-stable fashion.
Collapse
Affiliation(s)
- Matthijs Pals
- Machine Learning in Science, Excellence Cluster Machine Learning, University of Tübingen, Tübingen, Germany
- Tübingen AI Center, University of Tübingen, Tübingen, Germany
| | - Jakob H. Macke
- Machine Learning in Science, Excellence Cluster Machine Learning, University of Tübingen, Tübingen, Germany
- Tübingen AI Center, University of Tübingen, Tübingen, Germany
- Department Empirical Inference, Max Planck Institute for Intelligent Systems, Tübingen, Germany
| | - Omri Barak
- Rappaport Faculty of Medicine Technion, Israel Institute of Technology, Haifa, Israel
- Network Biology Research Laboratory, Israel Institute of Technology, Haifa, Israel
| |
Collapse
|