Perceptual learning, roving and the unsupervised bias

Informational feedback accelerates learning in multi-alternative perceptual judgements of orientation

Experience or training can substantially improve perceptual performance through perceptual learning, and the extent and rate of these improvements may be affected by feedback. In this paper, we first developed a neural network model based on the integrated reweighting theory (Dosher et al., 2013) to account for perceptual learning and performance in n-alternative identification tasks and the dependence of learning on different forms of feedback. We then report an experiment comparing the effectiveness of response feedback (RF) versus accuracy feedback (AF) or no feedback (NF) (full versus partial versus no supervision) in learning a challenging eight-alternative visual orientation identification (8AFC) task. Although learning sometimes occurred in the absence of feedback (NF), RF had a clear advantage above AF or NF in this task. Using hybrid supervision learning rules, a new n-alternative identification integrated reweighting theory (I-IRT) explained both the differences in learning curves given different feedback and the dynamic changes in identification confusion data. This study shows that training with more informational feedback (RF) is more effective, though not necessary, in these challenging n-alternative tasks, a result that has implications for developing training paradigms in realistic tasks.

Perceptual learning across saccades: Feature but not location specific

Perceptual learning is the ability to enhance perception through practice. The hallmark of perceptual learning is its specificity for the trained location and stimulus features, such as orientation. For example, training in discriminating a grating's orientation improves performance only at the trained location but not in other untrained locations. Perceptual learning has mostly been studied using stimuli presented briefly while observers maintained gaze at one location. However, in everyday life, stimuli are actively explored through eye movements, which results in successive projections of the same stimulus at different retinal locations. Here, we studied perceptual learning of orientation discrimination across saccades. Observers were trained to saccade to a peripheral grating and to discriminate its orientation change that occurred during the saccade. The results showed that training led to transsaccadic perceptual learning (TPL) and performance improvements which did not generalize to an untrained orientation. Remarkably, however, for the trained orientation, we found a complete transfer of TPL to the untrained location in the opposite hemifield suggesting high flexibility of reference frame encoding in TPL. Three control experiments in which participants were trained without saccades did not show such transfer, confirming that the location transfer was contingent upon eye movements. Moreover, performance at the trained location, but not at the untrained location, was also improved in an untrained fixation task. Our results suggest that TPL has both, a location-specific component that occurs before the eye movement and a saccade-related component that involves location generalization.

Variability in training unlocks generalization in visual perceptual learning through invariant representations

Stimulus and location specificity are long considered hallmarks of visual perceptual learning. This renders visual perceptual learning distinct from other forms of learning, where generalization can be more easily attained, and therefore unsuitable for practical applications, where generalization is key. Based on the hypotheses derived from the structure of the visual system, we test here whether stimulus variability can unlock generalization in perceptual learning. We train subjects in orientation discrimination, while we vary the amount of variability in a task-irrelevant feature, spatial frequency. We find that, independently of task difficulty, this manipulation enables generalization of learning to new stimuli and locations, while not negatively affecting the overall amount of learning on the task. We then use deep neural networks to investigate how variability unlocks generalization. We find that networks develop invariance to the task-irrelevant feature when trained with variable inputs. The degree of learned invariance strongly predicts generalization. A reliance on invariant representations can explain variability-induced generalization in visual perceptual learning. This suggests new targets for understanding the neural basis of perceptual learning in the higher-order visual cortex and presents an easy-to-implement modification of common training paradigms that may benefit practical applications.

Supervised Learning Occurs in Visual Perceptual Learning of Complex Natural Images

There have been long-standing debates regarding whether supervised or unsupervised learning mechanisms are involved in visual perceptual learning (VPL) [1-14]. However, these debates have been based on the effects of simple feedback only about response accuracy in detection or discrimination tasks of low-level visual features such as orientation [15-22]. Here, we examined whether the content of response feedback plays a critical role for the acquisition and long-term retention of VPL of complex natural images. We trained three groups of human subjects (n = 72 in total) to better detect "grouped microcalcifications" or "architectural distortion" lesions (referred to as calcification and distortion in the following) in mammograms either with no trial-by-trial feedback, partial trial-by-trial feedback (response correctness only), or detailed trial-by-trial feedback (response correctness and target location). Distortion lesions consist of more complex visual structures than calcification lesions [23-26]. We found that partial feedback is necessary for VPL of calcifications, whereas detailed feedback is required for VPL of distortions. Furthermore, detailed feedback during training is necessary for VPL of distortion and calcification lesions to be retained for 6 months. These results show that although supervised learning is heavily involved in VPL of complex natural images, the extent of supervision for VPL varies across different types of complex natural images. Such differential requirements for VPL to improve the detectability of lesions in mammograms are potentially informative for the professional training of radiologists.

Roving: The causes of interference and re-enabled learning in multi-task visual training

People routinely perform multiple visual judgments in the real world, yet, intermixing tasks or task variants during training can damage or even prevent learning. This paper explores why. We challenged theories of visual perceptual learning focused on plastic retuning of low-level retinotopic cortical representations by placing different task variants in different retinal locations, and tested theories of perceptual learning through reweighting (changes in readout) by varying task similarity. Discriminating different (but equivalent) and similar orientations in separate retinal locations interfered with learning, whereas training either with identical orientations or sufficiently different ones in different locations released rapid learning. This location crosstalk during learning renders it unlikely that the primary substrate of learning is retuning in early retinotopic visual areas; instead, learning likely involves reweighting from location-independent representations to a decision. We developed an Integrated Reweighting Theory (IRT), which has both V1-like location-specific representations and higher level (V4/IT or higher) location-invariant representations, and learns via reweighting the readout to decision, to predict the order of learning rates in different conditions. This model with suitable parameters successfully fit the behavioral data, as well as some microstructure of learning performance in a new trial-by-trial analysis.

Spatial Frequency Tuning and Transfer of Perceptual Learning for Motion Coherence Reflects the Tuning Properties of Global Motion Processing

Perceptual learning is typically highly specific to the stimuli and task used during training. However, recently, it has been shown that training on global motion can transfer to untrained tasks, reflecting the generalising properties of mechanisms at this level of processing. We investigated (i) if feedback was required for learning in a motion coherence task, (ii) the transfer across the spatial frequency of training on a global motion coherence task and (iii) the transfer of this training to a measure of contrast sensitivity. For our first experiment, two groups, with and without feedback, trained for ten days on a broadband motion coherence task. Results indicated that feedback was a requirement for robust learning. For the second experiment, training consisted of five days of direction discrimination using one of three motion coherence stimuli (where individual elements were comprised of either broadband Gaussian blobs or low- or high-frequency random-dot Gabor patches), with trial-by-trial auditory feedback. A pre- and post-training assessment was conducted for each of the three types of global motion coherence conditions and high and low spatial frequency contrast sensitivity (both without feedback). Our training paradigm was successful at eliciting improvement in the trained tasks over the five days. Post-training assessments found evidence of transfer for the motion coherence task exclusively for the group trained on low spatial frequency elements. For the contrast sensitivity tasks, improved performance was observed for low- and high-frequency stimuli, following motion coherence training with broadband stimuli, and for low-frequency stimuli, following low-frequency training. Our findings are consistent with perceptual learning, which depends on the global stage of motion processing in higher cortical areas, which is broadly tuned for spatial frequency, with a preference for low frequencies.

Comparing continual task learning in minds and machines

Humans can learn to perform multiple tasks in succession over the lifespan ("continual" learning), whereas current machine learning systems fail. Here, we investigated the cognitive mechanisms that permit successful continual learning in humans and harnessed our behavioral findings for neural network design. Humans categorized naturalistic images of trees according to one of two orthogonal task rules that were learned by trial and error. Training regimes that focused on individual rules for prolonged periods (blocked training) improved human performance on a later test involving randomly interleaved rules, compared with control regimes that trained in an interleaved fashion. Analysis of human error patterns suggested that blocked training encouraged humans to form "factorized" representation that optimally segregated the tasks, especially for those individuals with a strong prior bias to represent the stimulus space in a well-structured way. By contrast, standard supervised deep neural networks trained on the same tasks suffered catastrophic forgetting under blocked training, due to representational interference in the deeper layers. However, augmenting deep networks with an unsupervised generative model that allowed it to first learn a good embedding of the stimulus space (similar to that observed in humans) reduced catastrophic forgetting under blocked training. Building artificial agents that first learn a model of the world may be one promising route to solving continual task performance in artificial intelligence research.

Deep Neural Networks for Modeling Visual Perceptual Learning

Understanding visual perceptual learning (VPL) has become increasingly more challenging as new phenomena are discovered with novel stimuli and training paradigms. Although existing models aid our knowledge of critical aspects of VPL, the connections shown by these models between behavioral learning and plasticity across different brain areas are typically superficial. Most models explain VPL as readout from simple perceptual representations to decision areas and are not easily adaptable to explain new findings. Here, we show that a well -known instance of deep neural network (DNN), whereas not designed specifically for VPL, provides a computational model of VPL with enough complexity to be studied at many levels of analyses. After learning a Gabor orientation discrimination task, the DNN model reproduced key behavioral results, including increasing specificity with higher task precision, and also suggested that learning precise discriminations could transfer asymmetrically to coarse discriminations when the stimulus conditions varied. Consistent with the behavioral findings, the distribution of plasticity moved toward lower layers when task precision increased and this distribution was also modulated by tasks with different stimulus types. Furthermore, learning in the network units demonstrated close resemblance to extant electrophysiological recordings in monkey visual areas. Altogether, the DNN fulfilled predictions of existing theories regarding specificity and plasticity and reproduced findings of tuning changes in neurons of the primate visual areas. Although the comparisons were mostly qualitative, the DNN provides a new method of studying VPL, can serve as a test bed for theories, and assists in generating predictions for physiological investigations.SIGNIFICANCE STATEMENT Visual perceptual learning (VPL) has been found to cause changes at multiple stages of the visual hierarchy. We found that training a deep neural network (DNN) on an orientation discrimination task produced behavioral and physiological patterns similar to those found in human and monkey experiments. Unlike existing VPL models, the DNN was pre-trained on natural images to reach high performance in object recognition, but was not designed specifically for VPL; however, it fulfilled predictions of existing theories regarding specificity and plasticity and reproduced findings of tuning changes in neurons of the primate visual areas. When used with care, this unbiased and deep-hierarchical model can provide new ways of studying VPL from behavior to physiology.

Visual Perceptual Learning and Models

Visual perceptual learning through practice or training can significantly improve performance on visual tasks. Originally seen as a manifestation of plasticity in the primary visual cortex, perceptual learning is more readily understood as improvements in the function of brain networks that integrate processes, including sensory representations, decision, attention, and reward, and balance plasticity with system stability. This review considers the primary phenomena of perceptual learning, theories of perceptual learning, and perceptual learning's effect on signal and noise in visual processing and decision. Models, especially computational models, play a key role in behavioral and physiological investigations of the mechanisms of perceptual learning and for understanding, predicting, and optimizing human perceptual processes, learning, and performance. Performance improvements resulting from reweighting or readout of sensory inputs to decision provide a strong theoretical framework for interpreting perceptual learning and transfer that may prove useful in optimizing learning in real-world applications.

Neuromodulated Spike-Timing-Dependent Plasticity, and Theory of Three-Factor Learning Rules

Classical Hebbian learning puts the emphasis on joint pre- and postsynaptic activity, but neglects the potential role of neuromodulators. Since neuromodulators convey information about novelty or reward, the influence of neuromodulators on synaptic plasticity is useful not just for action learning in classical conditioning, but also to decide "when" to create new memories in response to a flow of sensory stimuli. In this review, we focus on timing requirements for pre- and postsynaptic activity in conjunction with one or several phasic neuromodulatory signals. While the emphasis of the text is on conceptual models and mathematical theories, we also discuss some experimental evidence for neuromodulation of Spike-Timing-Dependent Plasticity. We highlight the importance of synaptic mechanisms in bridging the temporal gap between sensory stimulation and neuromodulatory signals, and develop a framework for a class of neo-Hebbian three-factor learning rules that depend on presynaptic activity, postsynaptic variables as well as the influence of neuromodulators.

The effects of stimulus variability on the perceptual learning of speech and non-speech stimuli

Previous studies suggest fundamental differences between the perceptual learning of speech and non-speech stimuli. One major difference is in the way variability in the training set affects learning and its generalization to untrained stimuli: training-set variability appears to facilitate speech learning, while slowing or altogether extinguishing non-speech auditory learning. We asked whether the reason for this apparent difference is a consequence of the very different methodologies used in speech and non-speech studies. We hypothesized that speech and non-speech training would result in a similar pattern of learning if they were trained using the same training regimen. We used a 2 (random vs. blocked pre- and post-testing) × 2 (random vs. blocked training) × 2 (speech vs. non-speech discrimination task) study design, yielding 8 training groups. A further 2 groups acted as untrained controls, tested with either random or blocked stimuli. The speech task required syllable discrimination along 4 minimal-pair continua (e.g., bee-dee), and the non-speech stimuli required duration discrimination around 4 base durations (e.g., 50 ms). Training and testing required listeners to pick the odd-one-out of three stimuli, two of which were the base duration or phoneme continuum endpoint and the third varied adaptively. Training was administered in 9 sessions of 640 trials each, spread over 4–8 weeks. Significant learning was only observed following speech training, with similar learning rates and full generalization regardless of whether training used random or blocked schedules. No learning was observed for duration discrimination with either training regimen. We therefore conclude that the two stimulus classes respond differently to the same training regimen. A reasonable interpretation of the findings is that speech is perceived categorically, enabling learning in either paradigm, while the different base durations are not well-enough differentiated to allow for categorization, resulting in disruption to learning.

Fast learning of simple perceptual discriminations reduces brain activation in working memory and in high-level auditory regions

Introducing simple stimulus regularities facilitates learning of both simple and complex tasks. This facilitation may reflect an implicit change in the strategies used to solve the task when successful predictions regarding incoming stimuli can be formed. We studied the modifications in brain activity associated with fast perceptual learning based on regularity detection. We administered a two-tone frequency discrimination task and measured brain activation (fMRI) under two conditions: with and without a repeated reference tone. Although participants could not explicitly tell the difference between these two conditions, the introduced regularity affected both performance and the pattern of brain activation. The "No-Reference" condition induced a larger activation in frontoparietal areas known to be part of the working memory network. However, only the condition with a reference showed fast learning, which was accompanied by a reduction of activity in two regions: the left intraparietal area, involved in stimulus retention, and the posterior superior-temporal area, involved in representing auditory regularities. We propose that this joint reduction reflects a reduction in the need for online storage of the compared tones. We further suggest that this change reflects an implicit strategic shift "backwards" from reliance mainly on working memory networks in the "No-Reference" condition to increased reliance on detected regularities stored in high-level auditory networks.

Improvement of uncorrected visual acuity and contrast sensitivity with perceptual learning and transcranial random noise stimulation in individuals with mild myopia

Perceptual learning has been shown to produce an improvement of visual acuity (VA) and contrast sensitivity (CS) both in subjects with amblyopia and refractive defects such as myopia or presbyopia. Transcranial random noise stimulation (tRNS) has proven to be efficacious in accelerating neural plasticity and boosting perceptual learning in healthy participants. In this study, we investigated whether a short behavioral training regime using a contrast detection task combined with online tRNS was as effective in improving visual functions in participants with mild myopia compared to a 2-month behavioral training regime without tRNS (Camilleri et al., 2014). After 2 weeks of perceptual training in combination with tRNS, participants showed an improvement of 0.15 LogMAR in uncorrected VA (UCVA) that was comparable with that obtained after 8 weeks of training with no tRNS, and an improvement in uncorrected CS (UCCS) at various spatial frequencies (whereas no UCCS improvement was seen after 8 weeks of training with no tRNS). On the other hand, a control group that trained for 2 weeks without stimulation did not show any significant UCVA or UCCS improvement. These results suggest that the combination of behavioral and neuromodulatory techniques can be fast and efficacious in improving sight in individuals with mild myopia.

The role of alpha-rhythm states in perceptual learning: insights from experiments and computational models

During the past two decades growing evidence indicates that brain oscillations in the alpha band (~10 Hz) not only reflect an "idle" state of cortical activity, but also take a more active role in the generation of complex cognitive functions. A recent study shows that more than 60% of the observed inter-subject variability in perceptual learning can be ascribed to ongoing alpha activity. This evidence indicates a significant role of alpha oscillations for perceptual learning and hence motivates to explore the potential underlying mechanisms. Hence, it is the purpose of this review to highlight existent evidence that ascribes intrinsic alpha oscillations a role in shaping our ability to learn. In the review, we disentangle the alpha rhythm into different neural signatures that control information processing within individual functional building blocks of perceptual learning. We further highlight computational studies that shed light on potential mechanisms regarding how alpha oscillations may modulate information transfer and connectivity changes relevant for learning. To enable testing of those model based hypotheses, we emphasize the need for multidisciplinary approaches combining assessment of behavior and multi-scale neuronal activity, active modulation of ongoing brain states and computational modeling to reveal the mathematical principles of the complex neuronal interactions. In particular we highlight the relevance of multi-scale modeling frameworks such as the one currently being developed by "The Virtual Brain" project.

Perceptual expertise and top-down expectation of musical notation engages the primary visual cortex

Most theories of visual processing propose that object recognition is achieved in higher visual cortex. However, we show that category selectivity for musical notation can be observed in the first ERP component called the C1 (measured 40-60 msec after stimulus onset) with music-reading expertise. Moreover, the C1 note selectivity was observed only when the stimulus category was blocked but not when the stimulus category was randomized. Under blocking, the C1 activity for notes predicted individual music-reading ability, and behavioral judgments of musical stimuli reflected music-reading skill. Our results challenge current theories of object recognition, indicating that the primary visual cortex can be selective for musical notation within the initial feedforward sweep of activity with perceptual expertise and with a testing context that is consistent with the expertise training, such as blocking the stimulus category for music reading.

Coexistence of reward and unsupervised learning during the operant conditioning of neural firing rates

A fundamental goal of neuroscience is to understand how cognitive processes, such as operant conditioning, are performed by the brain. Typical and well studied examples of operant conditioning, in which the firing rates of individual cortical neurons in monkeys are increased using rewards, provide an opportunity for insight into this. Studies of reward-modulated spike-timing-dependent plasticity (RSTDP), and of other models such as R-max, have reproduced this learning behavior, but they have assumed that no unsupervised learning is present (i.e., no learning occurs without, or independent of, rewards). We show that these models cannot elicit firing rate reinforcement while exhibiting both reward learning and ongoing, stable unsupervised learning. To fix this issue, we propose a new RSTDP model of synaptic plasticity based upon the observed effects that dopamine has on long-term potentiation and depression (LTP and LTD). We show, both analytically and through simulations, that our new model can exhibit unsupervised learning and lead to firing rate reinforcement. This requires that the strengthening of LTP by the reward signal is greater than the strengthening of LTD and that the reinforced neuron exhibits irregular firing. We show the robustness of our findings to spike-timing correlations, to the synaptic weight dependence that is assumed, and to changes in the mean reward. We also consider our model in the differential reinforcement of two nearby neurons. Our model aligns more strongly with experimental studies than previous models and makes testable predictions for future experiments.

Deleterious effects of roving on learned tasks

In typical perceptual learning experiments, one stimulus type (e.g., a bisection stimulus offset either to the left or right) is presented per trial. In roving, two different stimulus types (e.g., a 30' and a 20' wide bisection stimulus) are randomly interleaved from trial to trial. Roving can impair both perceptual learning and task sensitivity. Here, we investigate the relationship between the two. Using a bisection task, we found no effect of roving before training. We next trained subjects and they improved. A roving condition applied after training impaired sensitivity.

Tactile perceptual learning: learning curves and transfer to the contralateral finger

Tactile perceptual learning has been shown to improve performance on tactile tasks, but there is no agreement about the extent of transfer to untrained skin locations. The lack of such transfer is often seen as a behavioral index of the contribution of early somatosensory brain regions. Moreover, the time course of improvements has never been described explicitly. Sixteen subjects were trained on the Ludvigh task (a tactile vernier task) on four subsequent days. On the fifth day, transfer of learning to the non-trained contralateral hand was tested. In five subjects, we explored to what extent training effects were retained approximately 1.5 years after the final training session, expecting to find long-term retention of learning effects after training. Results showed that tactile perceptual learning mainly occurred offline, between sessions. Training effects did not transfer initially, but became fully available to the untrained contralateral hand after a few additional training runs. After 1.5 years, training effects were not fully washed out and could be recuperated within a single training session. Interpreted in the light of theories of visual perceptual learning, these results suggest that tactile perceptual learning is not fundamentally different from visual perceptual learning, but might proceed at a slower pace due to procedural and task differences, thus explaining the apparent divergence in the amount of transfer and long-term retention.

Theory and Simulation in Neuroscience

Modeling work in neuroscience can be classified using two different criteria. The first one is the complexity of the model, ranging from simplified conceptual models that are amenable to mathematical analysis to detailed models that require simulations in order to understand their properties. The second criterion is that of direction of workflow, which can be from microscopic to macroscopic scales (bottom-up) or from behavioral target functions to properties of components (top-down). We review the interaction of theory and simulation using examples of top-down and bottom-up studies and point to some current developments in the fields of computational and theoretical neuroscience.