Dendrites endow artificial neural networks with accurate, robust and parameter-efficient learning

Nature Communications, Journal Year: 2025, Volume and Issue: 16(1)

Published: Jan. 22, 2025

Artificial neural networks (ANNs) are at the core of most Deep Learning (DL) algorithms that successfully tackle complex problems like image recognition, autonomous driving, and natural language processing. However, unlike biological brains who similar in a very efficient manner, DL require large number trainable parameters, making them energy-intensive prone to overfitting. Here, we show new ANN architecture incorporates structured connectivity restricted sampling properties dendrites counteracts these limitations. We find dendritic ANNs more robust overfitting match or outperform traditional on several classification tasks while using significantly fewer parameters. These advantages likely result different learning strategy, whereby nodes respond multiple classes, classical strive for class-specificity. Our findings suggest incorporation can make precise, resilient, parameter-efficient shed light how features impact strategies ANNs.

Language: Английский

---

Cellular psychology: relating cognition to context-sensitive pyramidal cells

Trends in Cognitive Sciences, Journal Year: 2024, Volume and Issue: unknown

Published: Oct. 1, 2024

Language: Английский

---

2D MoS2-based reconfigurable analog hardware

Nature Communications, Journal Year: 2025, Volume and Issue: 16(1)

Published: Jan. 2, 2025

Biological neural circuits demonstrate exceptional adaptability to diverse tasks by dynamically adjusting connections efficiently process information. However, current two-dimension materials-based neuromorphic hardware mainly focuses on specific devices individually mimic artificial synapse or heterosynapse soma and encoding the inner states realize corresponding mock object function. Recent advancements suggest that integrating multiple material brain-like functions including inter-mutual connecting assembly engineering has become a new research trend. In this work, we MoS2-based reconfigurable analog emulate synaptic, heterosynaptic, somatic functionalities. The inner-states inter-connections of all modules co-encode versatile such as analog-to-digital/digital-to-analog conversion, linear/nonlinear computations integration, vector-matrix multiplication, convolution, name few. By assembling fit with different environment-interactive demanding tasks, experimentally achieves reconstruction image sharpening medical images for diagnosis well circuit-level imitation attention-switching visual residual mechanisms smart perception. This innovative promotes development future general-purpose computing machines high flexibility tasks. study introduces integrates It adapts like enhancement perception, advancing flexible, solutions.

Language: Английский

---

Learning of state representation in recurrent network: the power of random feedback and biological constraints

Published: Jan. 14, 2025

How external/internal ‘state’ is represented in the brain crucial, since appropriate representation enables goal-directed behavior. Recent studies suggest that state and value can be simultaneously learnt through reinforcement learning (RL) using reward-prediction-error recurrent-neural-network (RNN) its downstream weights. However, how such neurally implemented remains unclear because training of RNN ‘backpropagation’ method requires weights, which are biologically unavailable at upstream RNN. Here we show random feedback instead weights still works ‘feedback alignment’, was originally demonstrated for supervised learning. We further if constrained to non-negative, occurs without alignment non-negative constraint ensures loose alignment. These results neural mechanisms RL representation/value power biological constraints.

Language: Английский

---

Learning of state representation in recurrent network: the power of random feedback and biological constraints

Published: Jan. 14, 2025

How external/internal ‘state’ is represented in the brain crucial, since appropriate representation enables goal-directed behavior. Recent studies suggest that state and value can be simultaneously learnt through reinforcement learning (RL) using reward-prediction-error recurrent-neural-network (RNN) its downstream weights. However, how such neurally implemented remains unclear because training of RNN ‘backpropagation’ method requires weights, which are biologically unavailable at upstream RNN. Here we show random feedback instead weights still works ‘feedback alignment’, was originally demonstrated for supervised learning. We further if constrained to non-negative, occurs without alignment non-negative constraint ensures loose alignment. These results neural mechanisms RL representation/value power biological constraints.

Language: Английский

---

Editorial overview: Computational neuroscience as a bridge between artificial intelligence, modeling and data

Current Opinion in Neurobiology, Journal Year: 2024, Volume and Issue: 84, P. 102835 - 102835

Published: Jan. 6, 2024

Language: Английский

---

A neural model for V1 that incorporates dendritic nonlinearities and back-propagating action potentials

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: Sept. 17, 2024

Abstract The groundbreaking work of Hubel and Wiesel has been instrumental in shaping our understanding V1, leading to modeling neural responses as cascades linear nonlinear processes what come be known the “standard model” vision. Under this formulation, however, some dendritic properties cannot represented a practical manner, while extensive evidence indicates that are an indispensable element key behaviours. As result, current V1 models fail explain number scenarios. In work, we propose implicit model for considers integration backpropagation action potentials from soma dendrites. This is parsimonious scheme minimizes energy, allows better conceptual processes, explains several neurophysiological phenomena have challenged classical approaches.

Language: Английский

---

Simulations predict differing phase responses to excitation vs. inhibition in theta-resonant pyramidal neurons

Journal of Neurophysiology, Journal Year: 2023, Volume and Issue: 130(4), P. 910 - 924

Published: Aug. 23, 2023

Rhythmic activity is ubiquitous in neural systems, with theta-resonant pyramidal neurons integrating rhythmic inputs many cortical structures. Impedance analysis has been widely used to examine frequency-dependent responses of neuronal membranes inputs, but it assumes that the membrane a linear system, requiring use small signals stay near-linear regime. However, postsynaptic potentials are often large and trigger nonlinear mechanisms (voltage-gated ion channels). The goals this work were

Language: Английский

---

Carbon-Aware Machine Learning: A Case Study on Cellular Traffic Forecasting with Spiking Neural Networks

IFIP advances in information and communication technology, Journal Year: 2024, Volume and Issue: unknown, P. 178 - 191

Published: Jan. 1, 2024

Language: Английский

---

Synaptic microarchitecture: the role of spatial interplay between excitatory and inhibitory inputs in shaping dendritic plasticity and neuronal output

Frontiers in Cellular Neuroscience, Journal Year: 2024, Volume and Issue: 18

Published: Dec. 20, 2024

Pyramidal neurons (PNs) receive and integrate thousands of synaptic inputs impinging onto their dendritic arbor to shape the neuronal output. The richness complexity such input-output transformation primarily relies on ability generate different forms local spikes, regenerative events originating in dendrites profoundly influencing probability temporal structure somatic spiking. Extensive work during last two decades has identified impact clustering cooperative plasticity among glutamatergic synapse promoting spikes. However, role inhibitory synapses processes remains elusive. In this opinion paper, following a general introduction input spatial distribution activity, we highlight coordinated excitatory as an emerging key factor organization architecture. particular will emphasize that relative positioning diverse at microscale level is major challenge for understanding how dynamics circuit function brain.In brain areas, distinct converging PNs show macro-scale across large compartments. For instance, hippocampal formation, fibers from entorhinal cortex (EC) project distal portions apical CA1 through perforant path (PP), while Schaffer collaterals (SCs) CA3 area mainly contact proximal (Megias et al., 2001) (Fig. 1). Similarly, neocortex, intra-cortical layer 2/3 (L2/3) axons (feedforward information) 5 (L5) PNs, with cortico-cortical high-order cortical areas (feedback targeting (Larkum, 2012). This illustrates large-scale connectivity scheme wherein either distant or regions preferentially dendrites, respectively (Felleman Van Essen, 1991). Such broad-scale reflects important functional properties where activation typically produces single action potentials co-activation can calcium plateau -specific spikes initiated region -leading neuron burst firing (Jarsky 2005;Takahashi Magee, 2009;Larkum 1999). supra-linear integration provides biophysical basis fundamental associative process combine compare types information cell (Bittner 2015;Larkum, Along same line, differential effect feedforward triggering combined feedback inducing firing, opportunity independent transmission these signals pathway (multiplexing) (Naud Sprekeler, 2017).Intriguingly, GABAergic are also non-randomly distributed along axo-dendritic axis PNs. Diverse subclasses interneurons (INs) target specific sub-regions including axon initial segment, soma, critically contributing e.g. oscillations (Klausberger Somogyi, 2007;Tzilivaki 2023). both hippocampus proximo-distal compartmentalization creates pattern broadly align subsets inputs. example, hippocampus, oriens-lacunosum-moleculare (O-LM), neurogliaform, (PP)-associated INs pyramidal aligning PP EC. Comparably, bistratified, SC-associated Ivy match (Klausberger, 2009, Lovett-Barron (Fig 1)The existence structured patterns localization persists smaller scales. At side, computational experimental works showed favors initiation (Mel, 1993;Poirazi Mel, 2001, Poirazi 2003a, 2003b, Larkum 2009). L5 instance within ~ 40 m range undergo summation due N-methyl-D-aspartate (NMDA) receptor-dependent mechanism, whereas more than 80 apart linearly, indicating determinants (Polsky 2004). been observed directly L2/3 spontaneous activity likely co-activate neighboring spines rather spines, thus forming "assemblets" 10 (Takahashi clustered underpins level. visual cortex, similarly tuned aids edge detection contour (Iacaruso 2017), motor task-related cluster μm subdomains support decisionmaking (Kerlin 2019). Besides relevance tight proximity between active (synaptic clustering), strongly depends morphology. thin short branches, high resistance determines low attenuation depolarization produced by individual signal branch (Kastellakis Poirazi, timely 20 radial oblique 100 initiate sodium spike regardless relationship branch, determining in-branch (Losonczy 2006). Anatomical studies SCs have revealed nonuniform structure. particular, number inter-spine distances well per was greater chance level, supporting modes, (Druckman 2014). Similar findings were thalamocortical PN (Rah 2013). Collectively, evidence indicates that, scales, arrangement crucially shapes transfer depolarization/spiking (Ujfalussy Makara, 2020;Kastellakis 2019).As inputs, several lines inhibition fine respect (Boivin Nedivi, 2018). Modeling suggest positioned distally (off-path) efficiently raise threshold initiating compared proximally-placed ones (on-path), on-path location effective shunting already-triggered (Gidon Segev, Both predictions corroborated experimentally ex vivo confirming branches determinant shaping excitability (Jadi concern, report highly sub-branch levels. O-LM (somatostatin+, SOM+) neurogliaform (neural nitric oxide synthase+, nNOS+) ending intermediated terminal domain bistratified (neuropeptide Y+, NPY+) origin basal (Bloss 2016). addition, study whole density significantly vary sub-regions, its ratio remarkably balanced (Iascone 2020). Finally, be located effectively controlling spine 2018;Chiu 2013).Extensive reports expression long-term potentiation (LTP) lower induction spreading signaling molecules small GTPases potentiated stretches ̴ m: establishes subset contiguous ultimately leading formation (Harvey Svoboda 2007;Harvey 2008;Hedrick Yasuda, 2017). On other hand, stimulation depress nearby diffusion phosphatase calcineurin, mechanism expected increase structural identity clusters (Oh Likewise, depression (LTD) potentiate (Chater Goda, 2021). Overall, observations short-range interplay define synapses. contribute largely obscure. Traditionally, considered poorly plastic take part phenomena adjusting (Steele modeling placement spatially constraint hence degree (Bar-Ilan GABAA receptors GABA uncaging leads shrinkage 15 m, reinforcing competitive selection (Hayama 2013).Nevertheless, indicate express (Chiu prompts questions interact -topics far investigated indirect approaches (Chapman 2022). After spike-timing-dependent population auditory PN, unstimulated found co-tuned achieve precise excitation-to-inhibition set point (Field, Interestingly, plasticity-induced remodeling suggesting (Chen thalamic afferents induces LTP formed SOM+ portion, hinting interaction al, Extending framework, identifies presence organizers (Kirchner Gjorgjieva, 2021).A recent (Ravasenga By single-spine pairing glutamate potential train, they 3-4 around depressed. Although factors could limit generalization finding physiological lack data, dependence excitation plays First, considering heterosynaptic disinhibit winner-takes-all process, concurrent maintaining global homeostatic balance. Second, activity-dependent complementing mentioned above. light interplay, convergence stretch network allowing, differentially control interneuron subtypes. together PP-associated interneurons, which mediate feed-back feed-forward inhibition, respectively. If ,differently EC consistently "interplay range" weaken bias balance feed-forward, thereby altering incoming signals. Thus, analogy aforementioned matching compartments, it co-alignment may serve 'fingerprint' subtypes, consistent act 'synaptic motifs'. broader excitatory-inhibitory assessed topology rules available models predicting spiking output receiving realistic cellular allow understand modulate tuning contribution enable learning gating areas. Importantly, inform aberrant would lead disruption coordination subtypes causing pathology. long run, refined about include computation large-networks functions designing neuromorphic efficient deep networks (DNNs) (Pagkalos 2024(Pagkalos , 2023)). (A) Representative received dendrites. Specific aligned fibers. Proximal stratum radiatum targeted SC (orange), amygdala projections (green), (pink) (purple). contrast, lacunosum moleculare thalamus (yellow), (red), (dark blue), (light blue). Dashed box delineates portion represented B. (B) Two possible arrangements segment. Left panel: (striped light-dark blue) (d1) beyond (d2). points combinations, certain paired Right randomly arrangement, there no (s.o., oriens; s.p., pyramidale; s.r., radiatum; s.lm., moleculare; PP, path; SC, Scaffer Collaterals; Thal, Thalamic; Amyg, Amygdala)

Language: Английский

---