The computational foundations of dynamic coding in working memory

Trends in Cognitive Sciences, Journal Year: 2024, Volume and Issue: 28(7), P. 614 - 627

Published: April 4, 2024

Working memory (WM) is a fundamental aspect of cognition. WM maintenance classically thought to rely on stable patterns neural activities. However, recent evidence shows that population activities during undergo dynamic variations before settling into pattern. Although this has been difficult explain theoretically, network models optimized for typically also exhibit such dynamics. Here, we examine versus coding in data, classical models, and task-optimized networks. We review principled mathematical reasons why do not, while naturally coding. suggest an update our understanding maintenance, which computational feature rather than epiphenomenon.

Language: Английский

---

Dimensionality reduction beyond neural subspaces with slice tensor component analysis

Nature Neuroscience, Journal Year: 2024, Volume and Issue: 27(6), P. 1199 - 1210

Published: May 6, 2024

Abstract Recent work has argued that large-scale neural recordings are often well described by patterns of coactivation across neurons. Yet the view variability is constrained to a fixed, low-dimensional subspace may overlook higher-dimensional structure, including stereotyped sequences or slowly evolving latent spaces. Here we argue task-relevant in data can also cofluctuate over trials time, defining distinct ‘covariability classes’ co-occur within same dataset. To demix these covariability classes, develop sliceTCA (slice tensor component analysis), new unsupervised dimensionality reduction method for tensors. In three example datasets, motor cortical activity during classic reaching task primates and recent multiregion mice, show capture more structure using fewer components than traditional methods. Overall, our theoretical framework extends population incorporating additional classes variables capturing structure.

Language: Английский

---

Optimal information loading into working memory explains dynamic coding in the prefrontal cortex

Proceedings of the National Academy of Sciences, Journal Year: 2023, Volume and Issue: 120(48)

Published: Nov. 20, 2023

Working memory involves the short-term maintenance of information and is critical in many tasks. The neural circuit dynamics underlying working remain poorly understood, with different aspects prefrontal cortical (PFC) responses explained by putative mechanisms. By mathematical analysis, numerical simulations, using recordings from monkey PFC, we investigate a but hitherto ignored aspect dynamics: loading. We find that, contrary to common assumptions, optimal loading into inputs that are largely orthogonal, rather than similar, late delay activities observed during maintenance, naturally leading widely phenomenon dynamic coding PFC. Using theoretically principled metric, show PFC exhibits hallmarks also emerges as general dynamical strategy task-optimized recurrent networks. Our theory unifies previous, seemingly conflicting theories based on attractor or purely sequential reveals normative principle coding.

Language: Английский

---

Inferring context-dependent computations through linear approximations of prefrontal cortex dynamics

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2023, Volume and Issue: unknown

Published: Feb. 6, 2023

Abstract The complex neural population activity of prefrontal cortex (PFC) is a hallmark cognitive processes. How these rich dynamics emerge and support computations largely unknown. Here, we infer mechanisms underlying the context-dependent selection integration sensory inputs by fitting dynamical models to PFC responses behaving monkeys. A class implementing linear driven external accurately captured within each context, achieving performance comparable without constraints. Two distinct input were equally consistent with data. One implemented recurrent dynamics, as previously proposed, relied on transient amplification. other subtle contextual modulation inputs, providing quantitative constraints attentional effects in areas required explain flexible behavior. Both consistently revealed properties missing more simplified, incomplete descriptions responses. By revealing cortical our modeling approach provides principled general framework link computation.

Language: Английский

---

Signatures of task learning in neural representations

Current Opinion in Neurobiology, Journal Year: 2023, Volume and Issue: 83, P. 102759 - 102759

Published: Sept. 12, 2023

Language: Английский

---

Flexible neural population dynamics govern the speed and stability of sensory encoding in mouse visual cortex

Nature Communications, Journal Year: 2024, Volume and Issue: 15(1)

Published: July 30, 2024

Time courses of neural responses underlie real-time sensory processing and perception. How these temporal dynamics change may be fundamental to how systems adapt different perceptual demands. By simultaneously recording from hundreds neurons in mouse primary visual cortex, we examined population stimuli at sub-second timescales, during behavioural states. We discovered that active states characterised by locomotion, single-neurons shift transient sustained response modes, facilitating rapid emergence stimulus tuning. Differences single-neuron were associated with changes correlations, including faster stabilisation stimulus-evoked the structure correlations locomotion. Using Factor Analysis, latent trajectories activity make more direct transitions between baseline stimulus-encoding This could partly explained dampening oscillatory present stationary Functionally, collectively enabled faster, stable efficient encoding new information These findings reveal a principle demands, where flexible govern speed stability encoding.

Language: Английский

---

Effects of noise and metabolic cost on cortical task representations

eLife, Journal Year: 2025, Volume and Issue: 13

Published: Jan. 21, 2025

Cognitive flexibility requires both the encoding of task-relevant and ignoring task-irrelevant stimuli. While neural coding stimuli is increasingly well understood, mechanisms for remain poorly understood. Here, we study how task performance biological constraints jointly determine relevant irrelevant in circuits. Using mathematical analyses task-optimized recurrent networks, show that circuits can exhibit a range representational geometries depending on strength noise metabolic cost. By comparing these results with recordings from primate prefrontal cortex (PFC) over course learning, activity PFC changes line minimal strategy. Specifically, our reveal suppression dynamically achieved by activity-silent, sub-threshold dynamics. Our provide normative explanation as to why implements an adaptive,

Language: Английский

---

Auditory Cortex Learns to Discriminate Audiovisual Cues through Selective Multisensory Enhancement

Published: April 2, 2025

Multisensory object discrimination is essential in everyday life, yet the neural mechanisms underlying this process remain unclear. In study, we trained rats to perform a two-alternative forced-choice task using both auditory and visual cues. Our findings reveal that multisensory perceptual learning actively engages cortex (AC) neurons audiovisual processing. Importantly, many AC exhibited experience-dependent associations between their preferences, displaying unique integration model. This model employed selective enhancement for auditory-visual pairing guiding contralateral choice, which correlated with improved discrimination. Furthermore, effectively distinguished whether preferred stimulus was paired its associated distinct integrative mechanism. results highlight capability of sensory cortices develop sophisticated strategies, adapting demands enhance abilities.

Language: Английский

---

Interdigitating Modules for Visual Processing During Locomotion and Rest in Mouse V1

Published: April 16, 2025

Layer 1 of V1 has been shown to receive locomotion-related signals from the dorsal lateral geniculate (dLGN) and posterior (LP) thalamic nuclei (Roth et al., 2016). Inputs dLGN terminate in M2+ patches while inputs LP target M2− interpatches (D’Souza 2019) suggesting that motion related are processed distinct networks. Here, we investigated by calcium imaging head-fixed awake mice whether L2/3 neurons underneath L1 modules differentially activated locomotion, networks feedback connections higher cortical areas may contribute these differences. We found strongly locomotion-modulated cell clusters during visual stimulation were aligned with interpatches, weakly modulated cells clustered under patches. Unlike patch cells, pairs interpatch showed increased correlated variability transients when sites visuotopic map far apart, activity is integrated across large parts field. Pathway tracing further suggests strong locomotion modulation relies on looped, like-to-like between apical dendrites MOs-, PM- RSP-projecting input L1. SST neurons, interneurons influence firing specific subnetworks controlling excitability interpatches.

Language: Английский

---

Interdigitating Modules for Visual Processing During Locomotion and Rest in Mouse V1

Published: April 16, 2025

Layer 1 of V1 has been shown to receive locomotion-related signals from the dorsal lateral geniculate (dLGN) and posterior (LP) thalamic nuclei (Roth et al., 2016). Inputs dLGN terminate in M2+ patches while inputs LP target M2− interpatches (D’Souza 2019) suggesting that motion related are processed distinct networks. Here, we investigated by calcium imaging head-fixed awake mice whether L2/3 neurons underneath L1 modules differentially activated locomotion, networks feedback connections higher cortical areas may contribute these differences. We found strongly locomotion-modulated cell clusters during visual stimulation were aligned with interpatches, weakly modulated cells clustered under patches. Unlike patch cells, pairs interpatch showed increased correlated variability transients when sites visuotopic map far apart, activity is integrated across large parts field. Pathway tracing further suggests strong locomotion modulation relies on looped, like-to-like between apical dendrites MOs-, PM- RSP-projecting input L1. SST neurons, interneurons influence firing specific subnetworks controlling excitability interpatches.

Language: Английский

---