Inferring neural activity before plasticity as a foundation for learning beyond backpropagation

Nature Neuroscience, Journal Year: 2024, Volume and Issue: 27(2), P. 348 - 358

Published: Jan. 3, 2024

Abstract For both humans and machines, the essence of learning is to pinpoint which components in its information processing pipeline are responsible for an error output, a challenge that known as ‘credit assignment’. It has long been assumed credit assignment best solved by backpropagation, also foundation modern machine learning. Here, we set out fundamentally different principle on called ‘prospective configuration’. In prospective configuration, network first infers pattern neural activity should result from learning, then synaptic weights modified consolidate change activity. We demonstrate this distinct mechanism, contrast (1) underlies well-established family models cortical circuits, (2) enables more efficient effective many contexts faced biological organisms (3) reproduces surprising patterns behavior observed diverse human rat experiments.

Language: Английский

---

Prediction of future input explains lateral connectivity in primary visual cortex

Current Biology, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 1, 2025

Language: Английский

---

Balancing prior knowledge and sensory data in a predictive coding model of coherent motion detection

PLoS Computational Biology, Journal Year: 2025, Volume and Issue: 21(5), P. e1013116 - e1013116

Published: May 21, 2025

This study introduces a neurobiologically inspired computational model based on the predictive coding algorithm, providing insights into coherent motion detection processes. The is designed to reflect key principles observed in visual system, particularly MT neurons and their surround suppression mechanisms, which play critical role detecting global motion. By integrating these principles, simulates how structures are decomposed individual shared sources, mirroring brain’s strategy for extracting patterns. results obtained from random dot stimuli underscore delicate balance between sensory data prior knowledge detection. Model testing across varying noise levels reveals that, as increases, takes longer stabilize its estimates, consistent with psychophysical experiments showing that response duration (e.g., reaction time or decision-making time) also increases under higher conditions. suggests an excessive emphasis prolongs stabilization detection, whereas optimal integration of expectations enhances accuracy efficiency by preventing disturbances due noise. These findings contribute potential explanations deficiencies schizophrenia.

Language: Английский

---

Sequential Memory with Temporal Predictive Coding

arXiv (Cornell University), Journal Year: 2023, Volume and Issue: unknown

Published: Jan. 1, 2023

Forming accurate memory of sequential stimuli is a fundamental function biological agents. However, the computational mechanism underlying in brain remains unclear. Inspired by neuroscience theories and recent successes applying predictive coding (PC) to \emph{static} tasks, this work we propose novel PC-based model for \emph{sequential} memory, called \emph{temporal coding} (tPC). We show that our tPC models can memorize retrieve inputs accurately with biologically plausible neural implementation. Importantly, analytical study reveals be viewed as classical Asymmetric Hopfield Network (AHN) an implicit statistical whitening process, which leads more stable performance tasks structured inputs. Moreover, find exhibits properties consistent behavioral observations neuroscience, thereby strengthening its relevance. Our establishes possible also theoretically interpreted using existing frameworks.

Language: Английский

---

Temporal prediction captures key differences between spiking excitatory and inhibitory V1 neurons

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

Published: May 14, 2024

Abstract Neurons in primary visual cortex (V1) respond to natural scenes with a sparse and irregular spike code that is carefully balanced by an interplay between excitatory inhibitory neurons. These neuron classes differ their statistics, tuning preferences, connectivity statistics temporal dynamics. To date, no single computational principle has been able account for these properties. We developed recurrently connected spiking network of units trained efficient prediction movie clips. found the model exhibited simple complex cell-like tuning, V1-like and, notably, also captured key differences V1 This suggests properties collectively serve facilitate sensory future.

Language: Английский

---

Learning probability distributions of sensory inputs with Monte Carlo predictive coding

PLoS Computational Biology, Journal Year: 2024, Volume and Issue: 20(10), P. e1012532 - e1012532

Published: Oct. 30, 2024

It has been suggested that the brain employs probabilistic generative models to optimally interpret sensory information. This hypothesis formalised in distinct frameworks, focusing on explaining separate phenomena. On one hand, classic predictive coding theory proposed how can be learned by networks of neurons employing local synaptic plasticity. other neural sampling theories have demonstrated stochastic dynamics enable circuits represent posterior distributions latent states environment. These frameworks were brought together variational filtering introduced coding. Here, we consider a variant for static inputs, which refer as Monte Carlo (MCPC). We demonstrate integration with results network learns precise using computation and The MCPC infer presence generate likely inputs their absence. Furthermore, captures experimental observations variability activity during perceptual tasks. By combining sampling, account both sets data previously had explained these individual frameworks.

Language: Английский

---

Predictive and error coding for vocal communication signals in the songbird auditory forebrain

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

Published: Feb. 26, 2024

Abstract Predictive coding posits that sensory signals are compared to internal models, with resulting prediction-error carried in the spiking responses of single neurons. Despite its proposal as a general cortical mechanism, including for speech processing, whether or how predictive functions single-neuron vocal communication is unknown. As proxy model, we developed neural network uses current context predict future spectrotemporal features signal, birdsong. We then represent birdsong either weighted sets latent evolving time, time-varying prediction-errors reflect difference between ongoing network-predicted and actual song. Using these spectrotemporal, predictive, song representations, fit linear/non-linear receptive fields neuron recorded from caudomedial nidopallium (NCM), caudal mesopallium (CMM) Field L, analogs mammalian auditory cortices, anesthetized European starlings, Sturnus vulgaris , listening conspecific songs. In all three regions, yield best model song-evoked responses, but unique information about representations (signal, prediction, error) The relative weighting this varies across contrast many computational models neither nor error segregated separate continuous interplay prediction consistent relevance processing temporally patterned signals, new integrated neurons required.

Language: Английский

---

Prediction of future input explains lateral connectivity in primary visual cortex

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

Published: June 1, 2024

Neurons in primary visual cortex (V1) show a remarkable functional specificity their pre- and postsynaptic partners. Recent work has revealed variety of wiring biases describing how the short- long-range connections V1 neurons relate to tuning properties. However, it is less clear whether these connectivity rules are based on some underlying principle cortical organization. Here, we that emerges naturally recurrent neural network optimized predict upcoming sensory inputs for natural stimuli. This temporal prediction model reproduces complex relationships between orientation direction preferences, tendency highly connected respond more similarly movies, differences excitatory inhibitory populations. Together, findings provide principled explanation anatomical properties early cortex.

Language: Английский

---

Balancing Prior Knowledge and Sensory Data in a Predictive Coding Model: Insights into Coherent Motion Detection in Schizophrenia

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

Published: June 1, 2024

Abstract This study introduces a biologically plausible computational model based on the predictive coding algorithm, providing insights into motion detection processes and potential deficiencies in schizophrenia. The decomposes structures individual shared sources, highlighting critical role of surround suppression detecting global motion. sheds light how brain extracts structure comprehends or coherent within visual field. results obtained from random dot stimuli underscore delicate balance between sensory data prior knowledge detection. Model testing across varying noise levels reveals longer convergence times with higher noise, consistent psychophysical experiments showing that response duration (e.g., reaction time decision-making time) also increases levels. suggests an excessive emphasis extends Conversely, for faster convergence, requires certain level to prevent disturbance due noise. These findings contribute explanations observed

Language: Английский

---

Policy optimization emerges from noisy representation learning

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

Published: Nov. 3, 2024

A bstract Nervous systems learn representations of the world and policies to act within it. We present a framework that uses reward-dependent noise facilitate policy opti- mization in representation learning networks. These networks balance extracting normative features task-relevant information solve tasks. Moreover, their changes reproduce several experimentally observed shifts neural code during task learning. Our presents biologically plausible mechanism for emergent optimization amid evidence plays vital role governing dynamics. Code is available at: NeuralThermalOptimization.

Language: Английский

---