Primate Ventral Stream Dynamics

• Ventral stream dynamics are the rapidly evolving neural processes that convert raw visual data into invariant object representations through integrated feedforward, recurrent, and top-down mechanisms.
• High-density electrophysiological recordings and representational similarity analyses reveal that multi-timescale, context-dependent computations yield superior object discrimination compared to static models.
• Recurrent neural network decoders and circuit-level models underscore the importance of temporal order, active sensing, and inter-areal feedback for robust visual inference in primate vision.

Ventral stream dynamics denote the temporally and spatially evolving computational processes by which the primate ventral visual stream (VVS) transforms raw sensory input into high-level, invariant object representations. These dynamics integrate feedforward, recurrent, and modulatory mechanisms across a multi-area cortical cascade (V1→V2→V4→IT), shaping sensory signals over time through both intrinsic and stimulus-driven processes. Recent advances in neurophysiology, system identification, and biologically inspired models reveal that ventral stream dynamics are not well approximated by static, stage-like architectures; instead, they comprise a rich tapestry of multi-timescale, context-dependent computations that support robust, flexible visual inference in dynamic environments (Dunnhofer et al., 18 Jan 2026).

1. Anatomical and Functional Basis of Ventral Stream Dynamics

The ventral stream consists of a hierarchically organized occipitotemporal pathway extending from V1 through intermediate areas (V2, V4) to inferior temporal (IT) cortex. Each region exhibits distinct cytoarchitectures, latency profiles, and functional roles. V1 and V2 encode retinotopic edge and junction features; V4 introduces greater tolerance to position, size, and simple transformations; IT supports category-invariant and view-invariant object identity representations (Xie et al., 2024, Shao et al., 2024).

Dynamically, latency of visual information increases from V1 (∼40 ms) through V4 (∼80–100 ms) to IT (∼110–160 ms), with late IT responses extending beyond 250 ms post-stimulus and encoding finer categorical distinctions (Kietzmann et al., 2019, Anthes et al., 14 Apr 2026, Masclef et al., 20 Jun 2026). Empirical observations reveal that IT responses are not static: the geometry of population trajectories evolves within and across images, exhibiting both early categorical encoding and later identity refinement (Dunnhofer et al., 18 Jan 2026, Anthes et al., 14 Apr 2026).

2. Empirical Measurement and Computational Characterization

High-density simultaneous electrophysiological recordings have facilitated time-resolved analysis of ventral stream population dynamics. Local field potential (LFP) arrays across V1, V4, and IT in macaque reveal that even within the "feedforward" window (first 100 ms), information transfer from V4 to IT is biphasic, with early and late windows exhibiting distinct semantic content—first low-level, then high-level geometry (Anthes et al., 14 Apr 2026). Representational Similarity Analysis (RSA) of population activity matrices (RDMs, $\mathrm{RDM}_{ij}=1-\mathrm{corr}(r_i, r_j)$) exposes a dynamic "handover" of geometry within IT, with early subspaces corresponding to coarse discrimination and late subspaces to finer identity (Anthes et al., 14 Apr 2026).

Multivariate decoding approaches, including recurrent neural networks (RNNs) trained on full temporal trajectories, demonstrate that time-evolving neural state vectors carry categorical information substantially beyond what can be extracted from static spatial patterns. For example, GRU-based decoders on IT trajectories achieve >17 pp higher accuracy than any snapshot-based method, and this advantage dissipates when temporal order is randomized, confirming the necessity of ordered neural dynamics for invariant object recognition (Anthes et al., 14 Apr 2026).

3. Mechanisms Driving Dynamic Computation: Recurrent, Intrinsic, and Active Components

Ventral stream dynamics arise from a combination of feedforward propagation, within-area horizontal recurrence, inter-areal feedback, and active top-down modulation. Empirical evidence from MEG, fMRI, and neural perturbation studies—coupled with BLT (bottom-up, lateral, top-down) network simulations—shows that recurrence is essential for capturing staggered, reversible emergence of categorical structure: GIST features rise and fall in V1→IT, animacy surfaces in IT before re-emerging in V4, and late IT responses depend critically on top-down and lateral interconnectivity (Kietzmann et al., 2019).

Across three domains, distinct dynamic modes emerge (Dunnhofer et al., 18 Jan 2026):

• Intrinsic dynamics (static stimuli): Late-phase responses in IT reflect within-area and inter-areal recurrence, E/I balance, and top-down projections (e.g., vlPFC, hippocampus); these later signals are behaviorally necessary for challenging discrimination tasks.
• Stimulus-driven dynamics (videos): IT robustly tracks not only object identity but also motion direction and speed. However, current video-based ANNs (e.g., 3D-CNNs, SlowFast networks) capture only appearance-bound temporal structure and fail to generalize to appearance-free manipulations, indicating missing dynamic invariances in artificial systems (Dunnhofer et al., 6 Jan 2026).
• Active sensing (eye movements): During natural scanpaths, fixation-locked ventral stream dynamics are governed by a closed oculo-motor loop; foveated glimpses are encoded, integrated, and drive policy learning where ventral-stream outputs guide subsequent fixation targets. Reinforcement learning objectives drive coupling between ventral/dorsal stream states and saccade planning (Ibrayev et al., 2024, Choi et al., 2023).

4. Representational Evolution, Manifold Geometry, and Robustness

As one ascends the ventral hierarchy, the geometry of object manifolds is progressively "untangled": representations become more compact (decreased manifold radius $R_m$), lower-dimensional (reduced $D_m$), and more linearly separable (increased classification margin $\gamma$), leading to improved adversarial robustness and shape bias (Shao et al., 2024). DNNs trained to align to higher-order VVS regions inherit these geometrical properties, evidenced by increased top-1 accuracy under adversarial PGD attacks ($\epsilon=0.009$: up to 0.40 for TO-guided vs. 0.20 for unaligned), output surface smoothness scores ($S$ rising from 0.82 to 0.95), and increasing percent of shape-based decisions (20 %→35 %) (Shao et al., 2024).

This evolving representational space is not exclusive to category-trained models. Models optimized on spatial latents (object location, pose) show similar layer-by-layer neural alignment to ventral stream data, with high correlation between task performance and overall neural fit (Pearson $r$ for distance = 0.98, translation = 0.94, category = 0.96), indicating that the ventral stream implements a unified "what+where" inference engine (Xie et al., 2024).

5. Modeling Perspectives: Representational, Circuit-level, and Behavioral

Three complementary modeling frameworks underpin current theoretical synthesis:

• Representational models: Losses include static cross-entropy, temporal prediction (slowness, next-frame), and dynamical similarity to neural trajectories (dynamic RSA) (Dunnhofer et al., 18 Jan 2026). Static image-trained ANNs capture early dynamics only; recurrence-equipped models are required for late-phase IT, dynamic videos, and active vision.
• Circuit-level models: Multi-area recurrent systems, incorporating delays, Dale's law (explicit E/I separation), and laminar specificity, emulate the distributed temporal processing of VVS with equations such as

$$\tau^a \dot h^a(t) = -h^a(t) + \phi\left(W^{aa} h^a(t) + \sum_{b\neq a} W^{ab} h^b(t-\Delta^{ab}) + U^a x(t)\right) + \eta^a(t).$$

Inclusion of E/I-balanced energy constraints ensures circuit stability and gain control (Dunnhofer et al., 18 Jan 2026).

• Behavioral models: Policies for saccade selection are learned via RL losses, with gaze sequences determined by current ventral stream state; such systems predict fixation densities and scanpaths, and replicate fixation-locked neural activity with millisecond-scale precision (Ibrayev et al., 2024, Choi et al., 2023, Dunnhofer et al., 18 Jan 2026).

6. Temporal Progression and Interareal Information Flow

Granger-causality analyses of population state evolution reveal bidirectional, temporally structured information flow. Feedforward peaks (V1→V4→IT) occur at ≈70–100 ms, while feedback (IT→V4, V4→V1) peaks later (110–260 ms), coincident with categorical disambiguation (Kietzmann et al., 2019). During object rotation, EEG-based decoding indicates that dorsal (spatial transformation) channels dominate early (0–200 ms), ventral (object identity) engagement rises sharply in the mid-phase (200–400 ms), and late phases (400–800 ms) reflect distributed ventral-dorsal-motor synergy (Masclef et al., 20 Jun 2026). These timings are consistent with both hierarchical feedforward theories and interactive models of perception.

Time (ms)	Dorsal TopBias	Ventral TopBias	Motor TopBias
0–200	0.60	0.20	0.55
200–400	0.30	0.60	0.45
400–800	0.25	0.45	0.30

7. Open Problems and Future Directions

Despite advances, current video-based and two-stream ANN models do not fully capture the appearance-invariant, temporally structured computations of biological ventral stream neurons. For example, when presented with appearance-free motion variants, macaque IT population activity preserves motion information (generalizes), but all current ANN classes fail this test, demonstrating only appearance-bound dynamic matching (Dunnhofer et al., 6 Jan 2026). Closing these gaps requires:

• Novel objectives encoding natural temporal statistics and appearance-invariant motion dynamics.
• Multi-area, E/I-balanced, and top-down-equipped architectures matching empirical constraints at the representational, circuit, and behavioral levels (Dunnhofer et al., 18 Jan 2026).
• More sensitive alignment metrics (unbiased RSA, nonlinear fits) to dissect subtle divergences across models trained on distinct objectives (Xie et al., 2024).
• Integration of active sensing mechanisms and dynamic reinforcement signals as core modeling ingredients (Ibrayev et al., 2024, Choi et al., 2023).

Converging evidence supports the interpretation of the ventral stream as a dynamic, multi-purpose inference architecture, whose temporal evolution and recurrent connectivity underpin the flexibility, robustness, and context sensitivity of primate object perception.