Time without clocks: Human time perception based on perceptual classification

Warrick Roseboom, Zafeirios Fountas, Kyriacos Nikiforou, David Bhowmik, Murray Shanahan, Anil K. SethbioRxiv 2017 · DOI

TL;DR

This paper proposes a novel explanation for human time perception: it's not about an internal clock, but rather about how the brain tracks changes in perceptual input. They built an artificial neural network that estimates duration by accumulating "salient" changes in its visual processing, and found its estimates closely mimic human reports, including common biases.

What Problem Does This Paper Address?

How the human brain generates our subjective experience of time is a long-standing mystery. Traditional models often rely on the idea of an internal "pacemaker" or clock mechanism. However, direct neural evidence for such a mechanism operating at psychologically relevant timescales is scarce. Alternative "state-dependent" models suggest that time perception arises from the natural dynamics of neural processing itself. This paper tackles the question of how subjective time perception, particularly for longer intervals (over 1 second), can be explained without recourse to these internal timing mechanisms, and instead focuses on the role of perceptual content. They aim to bridge the gap between models that focus on internal timing and those that emphasize the influence of what we perceive.

Methods

The core of the proposed system is a deep convolutional neural network (CNN) designed for image classification, specifically AlexNet, pre-trained on ImageNet. This network mimics the hierarchical structure of human visual processing.

Here's how it estimates time:

  1. Perceptual Change Detection: For each layer in the CNN, the system calculates the Euclidean distance between the network's activation patterns for consecutive video frames.
  2. Salient Event Thresholding: This distance is compared against a dynamic threshold. When the distance exceeds the threshold, it's considered a "salient perceptual change," and a unit of "subjective time" is accumulated. This threshold can be modulated, representing "attentional modulation."
  3. Accumulation: Each layer has an accumulator that counts these salient events.
  4. Duration Estimation: The accumulated salient changes across network layers are then fed into a Support Vector Machine (SVM) regressor, which is trained to map these abstract temporal units to standard units of time (seconds) based on the actual physical durations of the videos.

Crucially, the system's performance was also tested under a "Gaze" condition, where the input to the CNN was constrained to a patch around the human participant's actual gaze fixation point, mimicking human visual attention.

Datasets

The study used a dataset of natural scene videos recorded in various locations:

  • City streets
  • University campus
  • Countryside
  • Cafes and offices

These videos were split into durations ranging from 1 to 64 seconds. The dataset was used for both human participant experiments and for training and evaluating the artificial neural network model. No new benchmarks were introduced; existing video stimuli were adapted.

Key Insights and Ideas

  • Time as Perceptual Change: The most significant insight is the proposal that subjective time perception can arise from simply tracking changes in perceptual input, rather than relying on an internal clock.
  • CNNs as Timekeepers: Deep convolutional neural networks, designed for object recognition, can be repurposed to estimate duration by monitoring their internal activation dynamics.
  • Gaze Matters: Constraining the visual input to mimic human gaze significantly improves the model's time estimation accuracy and its similarity to human reports, highlighting the role of attention in time perception.
  • Content-Driven Biases: The model replicates human biases in time estimation based on scene content (e.g., busy city scenes are perceived as longer than quiet cafe scenes), suggesting these biases are rooted in the nature of perceptual processing itself.
  • Attention Modulation: The model's "temporal attention" mechanism, implemented by modulating the salient event threshold, successfully reproduces how attention influences subjective time (e.g., higher attention leading to overestimation).

Key Results

  • Human-like Estimation: The "Gaze" model's duration estimates closely matched human reports, replicating qualitative biases like overestimating short durations and underestimating long durations (Vierordt's Law) and scalar variability.
  • Scene-Based Biases: The model showed content-based biases similar to humans: city scenes were estimated as longer (~24% longer for the system vs. ~6% for humans), while office/cafe scenes were estimated as shorter (~7% for the system vs. ~4% for humans).
  • Gaze Improvement: The "Gaze" model significantly outperformed the "Full-frame" model, demonstrating the importance of attending to specific parts of a scene.
  • Robustness: The model's performance was robust across variations in threshold parameters and did not appear to be due to regression overfitting.
  • Pixel vs. Activation: Estimating time based on raw pixel differences between frames (without the CNN's classification) resulted in significantly poorer performance and did not replicate human biases as effectively, underscoring the role of perceptual classification.

Limitations and Future Work

  • Modality Specificity: The current model is primarily visual. The authors acknowledge that human time perception involves multiple senses and internal bodily states (interoception). Future work could extend the model to include auditory or other sensory inputs.
  • Simplicity of Metric: While Euclidean distance worked, more sophisticated metrics for perceptual change, potentially informed by predictive coding theories, could be explored.
  • Generalization: While the model replicates biases, the exact degree of over/underestimation might depend on the training data composition.
  • Predictive Coding: The authors suggest that incorporating predictive coding principles could further enhance the model's explanatory power by accounting for "top-down" influences and prior expectations.
  • Internal States: Future work could explore how internal bodily states, like those tracked by the insula cortex, contribute to time perception and how they could be integrated into such models.

Summary updated

Abstract

Despite being a fundamental dimension of experience, how the human brain generates the perception of time remains unknown. Here, we provide a novel explanation for how human time perception might be accomplished, based on non-temporal perceptual clas-sification processes. To demonstrate this proposal, we built an artificial neural system centred on a feed-forward image classification network, functionally similar to human visual processing. In this system, input videos of natural scenes drive changes in network activation, and accumulation of salient changes in activation are used to estimate duration. Estimates produced by this system match human reports made about the same videos, replicating key qualitative biases, including differentiating between scenes of walking around a busy city or sitting in a cafe or office. Our approach provides a working model of duration perception from stimulus to estimation and presents a new direction for examining the foundations of this central aspect of human experience.
BibTeX
@misc{roseboom2017time,
  author = {Warrick Roseboom and Zafeirios Fountas and Kyriacos Nikiforou and David Bhowmik and Murray Shanahan and Anil K. Seth},
  title = {{Time without clocks: Human time perception based on perceptual classification}},
  year = {2017},
  doi = {10.1101/172387},
  howpublished = {bioRxiv},
}

Loading related content...

Figure