Informing deep neural networks by multiscale principles of neuromodulatory systems

Mei, Jie; Müller, Eilif; Ramaswamy, Srikanth

doi:10.1016/j.tins.2021.12.008

Cited by 27 publications

(21 citation statements)

References 158 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…synaptic plasticity to shape decision making, learning, attention, memory, and motivation [33,36,31,11,35,41]. The multiscale organization of neuromodulatory systems provides high-level contextual control of cognition and behavior, and new approaches to inform and evaluate ANNs can be inspired by integrating them with ANNs [24].…”

Section: Biological Neuromodulation and Artificial Neural Networkmentioning

confidence: 99%

“…Creating such an array of complex neural networks is generally impractical, and so methods like model-agnostic continual learning [12] instead train two networks -one that encodes a task identity as a non-trivial latent representation, and another that maps that metarepresentation to weights that perform the actual task-specific inference. Here, we propose a novel bio-inspired neuromodulation [2,21,24,27] paradigm that reduces task-specific representational footprints so much that it is practical to store each task representation separately and eliminates the need for non-trivial task encoders for forgetting-free continual learning.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Learning to modulate random weights can induce task-specific contexts for economical meta and continual learning

Hong¹,

Pavlic²

2022

Preprint

View full text Add to dashboard Cite

Neural networks are vulnerable to catastrophic forgetting when data distributions are non-stationary during continual online learning; learning of a later task often leads to forgetting of an earlier task. One solution approach is model-agnostic continual meta-learning, whereby both task-specific and meta parameters are trained. Here, we depart from this view and introduce a novel neural-network architecture inspired by neuromodulation in biological nervous systems. Neuromodulation is the biological mechanism that dynamically controls and fine-tunes synaptic dynamics to complement the behavioral context in real-time, which has received limited attention in machine learning. We introduce a singlehidden-layer network that learns only a relatively small context vector per task (task-specific parameters) that neuromodulates unchanging, randomized weights (meta parameters) that transform the input. We show that when task boundaries are available, this approach can eliminate catastrophic forgetting entirely while also drastically reducing the number of learnable parameters relative to other context-vector-based approaches. Furthermore, by combining this model with a simple meta-learning approach for inferring task identity, we demonstrate that the model can be generalized into a framework to perform continual learning without knowledge of task boundaries. Finally, we showcase the framework in a supervised continual online learning scenario and discuss the implications of the proposed formalism.

show abstract

Section: Biological Neuromodulation and Artificial Neural Networkmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Learning to modulate random weights can induce task-specific contexts for economical meta and continual learning

Hong¹,

Pavlic²

2022

Preprint

View full text Add to dashboard Cite

show abstract

“…However, this also means that the NMC model represents only what it seeks to model closely: an isolated cortical microcircuit. As such, without the vast majority of its synaptic and neuromodulatory inputs, its activity would not resemble a plausible cortical state (Mei et al, 2022; Lee and Dan, 2012). While we were able to simulate in vivo -like (asynchronous sparse) activity through artificial depolarization and mimicking the effects of extracellular calcium (Markram et al, 2015), we cannot expect more complex emergent phenomena that would require, for example, inter-region connectivity and feedback connections (Felleman and Van Essen, 1991; Shepherd and Yamawaki, 2021).…”

Section: Introductionmentioning

confidence: 99%

Modeling and Simulation of Neocortical Micro- and Mesocircuitry. Part I: Anatomy

Reimann

Bolaños-Puchet

Courcol

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

The function of the neocortex is fundamentally determined by its repeating microcircuit motif, but also by its rich, hierarchical, interregional structure with a highly specific laminar architecture. The last decade has seen the emergence of extensive new data sets on anatomy and connectivity at the whole brain scale, providing promising new directions for studies of cortical function that take into account the inseparability of whole-brain and microcircuit architectures. Here, we present a data-driven computational model of the anatomy of non-barrel primary somatosensory cortex of juvenile rat, which integrates whole-brain scale data while providing cellular and subcellular specificity. This multiscale integration was achieved by building the morphologically detailed model of cortical circuitry embedded within a volumetric, digital brain atlas. The model consists of 4.2 million morphologically detailed neurons belonging to 60 different morphological types, placed in the nonbarrel subregions of the Paxinos and Watson atlas. They are connected by 13.2 billion synapses determined by axo-dendritic overlap, comprising local connectivity and long-range connectivity defined by topographic mappings between subregions and laminar axonal projection profiles, both parameterized by whole brain data sets. Additionally, we incorporated core- and matrix-type thalamocortical projection systems, associated with sensory and higher-order extrinsic inputs, respectively. An analysis of the modeled synaptic connectivity revealed a highly nonrandom topology with substantial structural differences but also synergy between local and long-range connectivity. Long-range connections featured a more divergent structure with a comparatively small group of neurons serving as hubs to distribute excitation to far away locations. Taken together with analyses at different spatial granularities, these results support the notion that local and interregional connectivity exist on a spectrum of scales, rather than as separate and distinct networks, as is commonly assumed. Finally, we predicted how the emergence of primary sensory cortical maps is constrained by the anatomy of thalamo-cortical projections. A subvolume of the model comprising 211,712 neurons in the front limb, jaw, and dysgranular zone has been made freely and openly available to the community.

show abstract

“…To further understand how biologically plausible mechanisms may shed light on DNNs and optimization methods, implementations at the dendritic, single-neuron, or microcircuitry levels have been increasingly realized. 1 , 2 , 3 , 4 , 5 Thus far, deep learning and neurorobotics studies 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 have examined whether biological neuromodulation may lead to behavioral benefits. In these studies, neuromodulation was commonly defined as a mechanism that self-reconfigures network hyperparameters and connectivity based on environmental or/and behavioral states of the neural network ( Table 1 ).…”

Section: Introductionmentioning

confidence: 99%

“… 6 , 7 , 8 , 9 The neuromodulatory system consists of the serotonergic (5-HT), dopaminergic (DA), noradrenergic (NA), and cholinergic systems (ACh), which modulate a spectrum of physiological and cognitive processes through highly region- and target-specific projections originating from midbrain, hindbrain, or forebrain areas. 3 , 10 Through neuromodulation-inspired learning, DNNs that employ adaptive learning rules including synaptic plasticity and feedback-based hyperparameter tuning were validated in tasks including spatial learning, cue-reward association, and image classification and recognition ( Table 1 ). As an example, 9 Vecoven et al.…”

Section: Introductionmentioning

confidence: 99%

Effects of neuromodulation-inspired mechanisms on the performance of deep neural networks in a spatial learning task

Mei¹,

Meshkinnejad²,

Mohsenzadeh³

2023

iScience

Self Cite

View full text Add to dashboard Cite

Informing deep neural networks by multiscale principles of neuromodulatory systems

Cited by 27 publications

References 158 publications

Learning to modulate random weights can induce task-specific contexts for economical meta and continual learning

Learning to modulate random weights can induce task-specific contexts for economical meta and continual learning

Modeling and Simulation of Neocortical Micro- and Mesocircuitry. Part I: Anatomy

Effects of neuromodulation-inspired mechanisms on the performance of deep neural networks in a spatial learning task

Contact Info

Product

Resources

About