Simon J. Thorpe

Fast saccades toward faces: Face detection in just 100 ms

Previous work has demonstrated that the human visual system can detect animals in complex natural scenes very efficiently and rapidly. In particular, using a saccadic choice task, H. Kirchner and S. J. Thorpe (2006) found that when two images are simultaneously flashed in the left and right visual fields, saccades toward the side with an animal can be initiated in as little as 120-130 ms. Here we show that saccades toward human faces are even faster, with the earliest reliable saccades occurring in just 100-110 ms, and mean reaction times of roughly 140 ms. Intriguingly, it appears that these very fast saccades are not completely under instructional control, because when faces were paired with photographs of vehicles, fast saccades were still biased toward faces even when the subject was targeting vehicles. Finally, we tested whether these very fast saccades might only occur in the simple case where the images are presented left and right of fixation by showing they also occur when the images are presented above and below fixation. Such results impose very serious constraints on the sorts of processing model that can be invoked and demonstrate that face-selective behavioral responses can be generated extremely rapidly.

Rapid Formation of Robust Auditory Memories: Insights from Noise

Before a natural sound can be recognized, an auditory signature of its source must be learned through experience. Here we used random waveforms to probe the formation of new memories for arbitrary complex sounds. A behavioral measure was designed, based on the detection of repetitions embedded in noises up to 4 s long. Unbeknownst to listeners, some noise samples reoccurred randomly throughout an experimental block. Results showed that repeated exposure induced learning for otherwise totally unpredictable and meaningless sounds. The learning was unsupervised and resilient to interference from other task-relevant noises. When memories were formed, they emerged rapidly, performance became abruptly near-perfect, and multiple noises were remembered for several weeks. The acoustic transformations to which recall was tolerant suggest that the learned features were local in time. We propose that rapid sensory plasticity could explain how the auditory brain creates useful memories from the ever-changing, but sometimes repeating, acoustical world.

2012 Special Issue: Extraction of temporally correlated features from dynamic vision sensors with spike-timing-dependent plasticity

A biologically inspired approach to learning temporally correlated patterns from a spiking silicon retina is presented. Spikes are generated from the retina in response to relative changes in illumination at the pixel level and transmitted to a feed-forward spiking neural network. Neurons become sensitive to patterns of pixels with correlated activation times, in a fully unsupervised scheme. This is achieved using a special form of Spike-Timing-Dependent Plasticity which depresses synapses that did not recently contribute to the post-synaptic spike activation, regardless of their activation time. Competitive learning is implemented with lateral inhibition. When tested with real-life data, the system is able to extract complex and overlapping temporally correlated features such as car trajectories on a freeway, after only 10 min of traffic learning. Complete trajectories can be learned with a 98% detection rate using a second layer, still with unsupervised learning, and the system may be used as a car counter. The proposed neural network is extremely robust to noise and it can tolerate a high degree of synaptic and neuronal variability with little impact on performance. Such results show that a simple biologically inspired unsupervised learning scheme is capable of generating selectivity to complex meaningful events on the basis of relatively little sensory experience.

Exploration of rank order coding with spiking neural networks for speech recognition

Speech recognition is very difficult in the context of noisy and corrupted speech. Most conventional techniques need huge databases to estimate speech (or noise) density probabilities to perform recognition. We discuss the potential of perceptive speech analysis and processing in combination with biologically plausible neural network processors. We illustrate the potential of such non-linear processing of speech by means of a preliminary test with recognition of French spoken digits from a small speech database

Low-level cues and ultra-fast face detection.

Recent experimental work has demonstrated the existence of extremely rapid saccades toward faces in natural scenes that can be initiated only 100u2009ms after image onset (Crouzet et al., 2010). These ultra-rapid saccades constitute a major challenge to current models of processing in the visual system because they do not seem to leave enough time for even a single feed-forward pass through the ventral stream. Here we explore the possibility that the information required to trigger these very fast saccades could be extracted very early on in visual processing using relatively low-level amplitude spectrum (AS) information in the Fourier domain. Experiment 1 showed that AS normalization can significantly alter face-detection performance. However, a decrease of performance following AS normalization does not alone prove that AS-based information is used (Gaspar and Rousselet, 2009). In Experiment 2, following the Gaspar and Rousselet paper, we used a swapping procedure to clarify the role of AS information in fast object detection. Our experiment is composed of three conditions: (i) original images, (ii) category swapped, in which the face image has the AS of a vehicle, and the vehicle has the AS of a face, and (iii) identity swapped, where the face has the AS of another face image, and the vehicle has the AS of another vehicle image. The results showed very similar levels of performance in the original and identity swapped conditions, and a clear drop in the category swapped condition. This result demonstrates that, in the early temporal window offered by the saccadic choice task, the visual saccadic system does indeed rely on low-level AS information in order to rapidly detect faces. This sort of crude diagnostic information could potentially be derived very early on in the visual system, possibly as early as V1 and V2.

Effects of Task Requirements on Rapid Natural Scene Processing: From Common Sensory Encoding to Distinct Decisional Mechanisms

Using manual responses, human participants are remarkably fast and accurate at deciding if a natural scene contains an animal, but recent data show that they are even faster to indicate with saccadic eye movements which of 2 scenes contains an animal. How could it be that 2 images can apparently be processed faster than a single image? To better understand the origin of this speed advantage in forced-choice categorization, the present study used a masking procedure to compare 4 tasks in which sensory, decisional, and motor aspects were systematically varied. With stimulus onset asynchronies (SOAs) above 40 ms, there were substantial differences in sensitivity between tasks, as determined by d measurements, with an advantage for tasks using a single image. However, with SOAs below 30-40 ms, sensitivity was similar for all experiments, despite very large differences in reaction time. This suggests that the initial part of the sensory encoding relies on common and parallel processing across a large range of tasks, whether participants have to categorize the image or locate a target in 1 of 2 scenes.

Limits of Event-related Potential Differences in Tracking Object Processing Speed

We report results from two experiments in which subjects had to categorize briefly presented upright or inverted natural scenes. In the first experiment, subjects decided whether images contained animals or human faces presented at different scales. Behavioral results showed virtually identical processing speed between the two categories and very limited effects of inversion. One type of event-related potential (ERP) comparison, potentially capturing low-level physical differences, showed large effects with onsets at about 150 msec in the animal task. However, in the human face task, those differences started as early as 100 msec. In the second experiment, subjects responded to close-up views of animal faces or human faces in an attempt to limit physical differences between image sets. This manipulation almost completely eliminated small differences before 100 msec in both tasks. But again, despite very similar behavioral performances and short reaction times in both tasks, human faces were associated with earlier ERP differences compared with animal faces. Finally, in both experiments, as an alternative way to determine processing speed, we compared the ERP with the same images when seen as targets and nontargets in different tasks. Surprisingly, all task-dependent ERP differences had relatively long latencies. We conclude that task-dependent ERP differences fail to capture object processing speed, at least for some categories like faces. We discuss models of object processing that might explain our results, as well as alternative approaches.

Fast recognition of musical sounds based on timbre.

Human listeners seem to have an impressive ability to recognize a wide variety of natural sounds. However, there is surprisingly little quantitative evidence to characterize this fundamental ability. Here the speed and accuracy of musical-sound recognition were measured psychophysically with a rich but acoustically balanced stimulus set. The set comprised recordings of notes from musical instruments and sung vowels. In a first experiment, reaction times were collected for three target categories: voice, percussion, and strings. In a go/no-go task, listeners reacted as quickly as possible to members of a target category while withholding responses to distractors (a diverse set of musical instruments). Results showed near-perfect accuracy and fast reaction times, particularly for voices. In a second experiment, voices were recognized among strings and vice-versa. Again, reaction times to voices were faster. In a third experiment, auditory chimeras were created to retain only spectral or temporal features of the voice. Chimeras were recognized accurately, but not as quickly as natural voices. Altogether, the data suggest rapid and accurate neural mechanisms for musical-sound recognition based on selectivity to complex spectro-temporal signatures of sound sources.

At 120 msec you can spot the animal but you don't yet know it's a dog

Earlier studies suggested that the visual system processes information at the basic level (e.g., dog) faster than at the subordinate (e.g., Dalmatian) or superordinate (e.g., animals) levels. However, the advantage of the basic category over the superordinate category in object recognition has been challenged recently, and the hierarchical nature of visual categorization is now a matter of debate. To address this issue, we used a forced-choice saccadic task in which a target and a distractor image were displayed simultaneously on each trial and participants had to saccade as fast as possible toward the image containing animal targets based on different categorization levels. This protocol enables us to investigate the first 100–120 msec, a previously unexplored temporal window, of visual object categorization. The first result is a surprising stability of the saccade latency (median RT ∼155 msec) regardless of the animal target category and the dissimilarity of target and distractor image sets. Accuracy was high (around 80% correct) for categorization tasks that can be solved at the superordinate level but dropped to almost chance levels for basic level categorization. At the basic level, the highest accuracy (62%) was obtained when distractors were restricted to another dissimilar basic category. Computational simulations based on the saliency map model showed that the results could not be predicted by pure bottom–up saliency differences between images. Our results support a model of visual recognition in which the visual system can rapidly access relatively coarse visual representations that provide information at the superordinate level of an object, but where additional visual analysis is required to allow more detailed categorization at the basic level.

Animal Detection Precedes Access to Scene Category

The processes underlying object recognition are fundamental for the understanding of visual perception. Humans can recognize many objects rapidly even in complex scenes, a task that still presents major challenges for computer vision systems. A common experimental demonstration of this ability is the rapid animal detection protocol, where human participants earliest responses to report the presence/absence of animals in natural scenes are observed at 250–270 ms latencies. One of the hypotheses to account for such speed is that people would not actually recognize an animal per se, but rather base their decision on global scene statistics. These global statistics (also referred to as spatial envelope or gist) have been shown to be computationally easy to process and could thus be used as a proxy for coarse object recognition. Here, using a saccadic choice task, which allows us to investigate a previously inaccessible temporal window of visual processing, we showed that animal – but not vehicle – detection clearly precedes scene categorization. This asynchrony is in addition validated by a late contextual modulation of animal detection, starting simultaneously with the availability of scene category. Interestingly, the advantage for animal over scene categorization is in opposition to the results of simulations using standard computational models. Taken together, these results challenge the idea that rapid animal detection might be based on early access of global scene statistics, and rather suggests a process based on the extraction of specific local complex features that might be hardwired in the visual system.