Dissecting the Roles of Supervised and Unsupervised Learning in Perceptual Discrimination Judgments

Our ability to compare sensory stimuli is a fundamental cognitive function, which is known to be affected by two biases: choice bias, which reflects a preference for a given response, and contraction bias, which reflects a tendency to perceive stimuli as similar to previous ones. To test whether both reflect supervised processes, we designed feedback protocols aimed to modify them and tested them in human participants. Choice bias was readily modifiable. However, contraction bias was not. To compare these results to those predicted from an optimal supervised process, we studied a noise-matched optimal linear discriminator (Perceptron). In this model, both biases were substantially modified, indicating that the “resilience” of contraction bias to feedback does not maximize performance. These results suggest that perceptual discrimination is a hierarchical, two-stage process. In the first, stimulus statistics are learned and integrated with representations in an unsupervised process that is impenetrable to external feedback. In the second, a binary judgment, learned in a supervised way, is applied to the combined percept.SIGNIFICANCE STATEMENT The seemingly effortless process of inferring physical reality from the sensory input is highly influenced by previous knowledge, leading to perceptual biases. Two common ones are contraction bias (the tendency to perceive stimuli as similar to previous ones) and choice bias (the tendency to prefer a specific response). Combining human psychophysical experiments with computational modeling we show that they reflect two different learning processes. Contraction bias reflects unsupervised learning of stimuli statistics, whereas choice bias results from supervised or reinforcement learning. This dissociation reveals a hierarchical, two-stage process. The first, where stimuli statistics are learned and integrated with representations, is unsupervised. The second, where a binary judgment is applied to the combined percept, is learned in a supervised way.     

Dynamic evolving spiking neural networks for on-line spatio- and spectro-temporal pattern recognition

On-line learning and recognition of spatio- and spectro-temporal data (SSTD) is a very challenging task and an important one for the future development of autonomous machine learning systems with broad applications. Models based on spiking neural networks (SNN) have already proved their potential in capturing spatial and temporal data. One class of them, the evolving SNN (eSNN), uses a one-pass rank-order learning mechanism and a strategy to evolve a new spiking neuron and new connections to learn new patterns from incoming data. So far these networks have been mainly used for fast image and speech frame-based recognition. Alternative spike-time learning methods, such as Spike-Timing Dependent Plasticity (STDP) and its variant Spike Driven Synaptic Plasticity (SDSP), can also be used to learn spatio-temporal representations, but they usually require many iterations in an unsupervised or semi-supervised mode of learning. This paper introduces a new class of eSNN, dynamic eSNN, that utilise both rank-order learning and dynamic synapses to learn SSTD in a fast, on-line mode. The paper also introduces a new model called deSNN, that utilises rank-order learning and SDSP spike-time learning in unsupervised, supervised, or semi-supervised modes. The SDSP learning is used to evolve dynamically the network changing connection weights that capture spatio-temporal spike data clusters both during training and during recall. The new deSNN model is first illustrated on simple examples and then applied on two case study applications: (1) moving object recognition using address-event representation (AER) with data collected using a silicon retina device; (2) EEG SSTD recognition for brain–computer interfaces. The deSNN models resulted in a superior performance in terms of accuracy and speed when compared with other SNN models that use either rank-order or STDP learning. The reason is that the deSNN makes use of both the information contained in the order of the first input spikes (which information is explicitly present in input data streams and would be crucial to consider in some tasks) and of the information contained in the timing of the following spikes that is learned by the dynamic synapses as a whole spatio-temporal pattern.     

Continual task learning in natural and artificial agents

How do humans and other animals learn new tasks? A wave of brain recording studies has investigated how neural representations change during task learning, with a focus on how tasks can be acquired and coded in ways that minimise mutual interference. We review recent work that has explored the geometry and dimensionality of neural task representations in neocortex, and computational models that have exploited these findings to understand how the brain may partition knowledge between tasks. We discuss how ideas from machine learning, including those that combine supervised and unsupervised learning, are helping neuroscientists understand how natural tasks are learned and coded in biological brains.     

Adaptive Resonance Theory: How a brain learns to consciously attend,learn, and recognize a changing world

Adaptive Resonance Theory, or ART, is a cognitive and neural theory of how the brain autonomously learns to categorize, recognize, and predict objects and events in a changing world. This article reviews classical and recent developments of ART, and provides a synthesis of concepts, principles, mechanisms, architectures, and the interdisciplinary data bases that they have helped to explain and predict. The review illustrates that ART is currently the most highly developed cognitive and neural theory available, with the broadest explanatory and predictive range. Central to ART’s predictive power is its ability to carry out fast, incremental, and stable unsupervised and supervised learning in response to a changing world. ART specifies mechanistic links between processes of consciousness, learning, expectation, attention, resonance, and synchrony during both unsupervised and supervised learning. ART provides functional and mechanistic explanations of such diverse topics as laminar cortical circuitry; invariant object and scenic gist learning and recognition; prototype, surface, and boundary attention; gamma and beta oscillations; learning of entorhinal grid cells and hippocampal place cells; computation of homologous spatial and temporal mechanisms in the entorhinal–hippocampal system; vigilance breakdowns during autism and medial temporal amnesia; cognitive–emotional interactions that focus attention on valued objects in an adaptively timed way; item–order–rank working memories and learned list chunks for the planning and control of sequences of linguistic, spatial, and motor information; conscious speech percepts that are influenced by future context; auditory streaming in noise during source segregation; and speaker normalization. Brain regions that are functionally described include visual and auditory neocortex; specific and nonspecific thalamic nuclei; inferotemporal, parietal, prefrontal, entorhinal, hippocampal, parahippocampal, perirhinal, and motor cortices; frontal eye fields; supplementary eye fields; amygdala; basal ganglia: cerebellum; and superior colliculus. Due to the complementary organization of the brain, ART does not describe many spatial and motor behaviors whose matching and learning laws differ from those of ART. ART algorithms for engineering and technology are listed, as are comparisons with other types of models.     

What are the computations of the cerebellum,the basal ganglia and the cerebral cortex?

The classical notion that the cerebellum and the basal ganglia are dedicated to motor control is under dispute given increasing evidence of their involvement in non-motor functions. Is it then impossible to characterize the functions of the cerebellum, the basal ganglia and the cerebral cortex in a simplistic manner? This paper presents a novel view that their computational roles can be characterized not by asking what are the “goals” of their computation, such as motor or sensory, but by asking what are the “methods” of their computation, specifically, their learning algorithms. There is currently enough anatomical, physiological, and theoretical evidence to support the hypotheses that the cerebellum is a specialized organism for supervised learning, the basal ganglia are for reinforcement learning, and the cerebral cortex is for unsupervised learning.This paper investigates how the learning modules specialized for these three kinds of learning can be assembled into goal-oriented behaving systems. In general, supervised learning modules in the cerebellum can be utilized as “internal models” of the environment. Reinforcement learning modules in the basal ganglia enable action selection by an “evaluation” of environmental states. Unsupervised learning modules in the cerebral cortex can provide statistically efficient representation of the states of the environment and the behaving system. Two basic action selection architectures are shown, namely, reactive action selection and predictive action selection. They can be implemented within the anatomical constraint of the network linking these structures. Furthermore, the use of the cerebellar supervised learning modules for state estimation, behavioral simulation, and encapsulation of learned skill is considered. Finally, the usefulness of such theoretical frameworks in interpreting brain imaging data is demonstrated in the paradigm of procedural learning.     

Human Semi‐Supervised Learning

Most empirical work in human categorization has studied learning in either fully supervised or fully unsupervised scenarios. Most real‐world learning scenarios, however, are semi‐supervised: Learners receive a great deal of unlabeled information from the world, coupled with occasional experiences in which items are directly labeled by a knowledgeable source. A large body of work in machine learning has investigated how learning can exploit both labeled and unlabeled data provided to a learner. Using equivalences between models found in human categorization and machine learning research, we explain how these semi‐supervised techniques can be applied to human learning. A series of experiments are described which show that semi‐supervised learning models prove useful for explaining human behavior when exposed to both labeled and unlabeled data. We then discuss some machine learning models that do not have familiar human categorization counterparts. Finally, we discuss some challenges yet to be addressed in the use of semi‐supervised models for modeling human categorization.     

Accuracy of latent-variable estimation in Bayesian semi-supervised learning

Hierarchical probabilistic models, such as Gaussian mixture models, are widely used for unsupervised learning tasks. These models consist of observable and latent variables, which represent the observable data and the underlying data-generation process, respectively. Unsupervised learning tasks, such as cluster analysis, are regarded as estimations of latent variables based on the observable ones. The estimation of latent variables in semi-supervised learning, where some labels are observed, will be more precise than that in unsupervised, and one of the concerns is to clarify the effect of the labeled data. However, there has not been sufficient theoretical analysis of the accuracy of the estimation of latent variables. In a previous study, a distribution-based error function was formulated, and its asymptotic form was calculated for unsupervised learning with generative models. It has been shown that, for the estimation of latent variables, the Bayes method is more accurate than the maximum-likelihood method. The present paper reveals the asymptotic forms of the error function in Bayesian semi-supervised learning for both discriminative and generative models. The results show that the generative model, which uses all of the given data, performs better when the model is well specified.     

Learning invariant face recognition from examples

Autonomous learning is demonstrated by living beings that learn visual invariances during their visual experience. Standard neural network models do not show this sort of learning. On the example of face recognition in different situations we propose a learning process that separates learning of the invariance proper from learning new instances of individuals. The invariance is learned by a set of examples called model, which contains instances of all situations. New instances are compared with these on the basis of rank lists, which allow generalization across situations. The result is also implemented as a spike-time-based neural network, which is shown to be robust against disturbances. The learning capability is demonstrated by recognition experiments on a set of standard face databases.     

A kernel-based framework to tensorial data analysis

Tensor-based techniques for learning allow one to exploit the structure of carefully chosen representations of data. This is a desirable feature in particular when the number of training patterns is small which is often the case in areas such as biosignal processing and chemometrics. However, the class of tensor-based models is somewhat restricted and might suffer from limited discriminative power. On a different track, kernel methods lead to flexible nonlinear models that have been proven successful in many different contexts. Nonetheless, a naïve application of kernel methods does not exploit structural properties possessed by the given tensorial representations. The goal of this work is to go beyond this limitation by introducing non-parametric tensor-based models. The proposed framework aims at improving the discriminative power of supervised tensor-based models while still exploiting the structural information embodied in the data. We begin by introducing a feature space formed by multilinear functionals. The latter can be considered as the infinite dimensional analogue of tensors. Successively we show how to implicitly map input patterns in such a feature space by means of kernels that exploit the algebraic structure of data tensors. The proposed tensorial kernel links to the MLSVD and features an interesting invariance property; the approach leads to convex optimization and fits into the same primal-dual framework underlying SVM-like algorithms.     

Incremental Adaptive Learning Vector Quantization for Character Recognition with Continuous Style Adaptation

Incremental learning enables continuous model adaptation based on a constantly arriving data stream. It is a way relevant to human cognitive system, which learns to predict objects in a changing world. Incremental learning for character recognition is a typical scenario that characters appear sequentially and the font/writing style changes irregularly. In the paper, we investigate how to classify characters incrementally (i.e., input patterns appear once at a time). A reasonable assumption is that adjacent characters from the same font or the same writer share the same style in a short period while style variation occurs in characters printed by different fonts or written by different persons during a long period. The challenging issue here is how to take advantage of the local style consistency and adapt to the continuous style variation as well incrementally. For this purpose, we propose a continuous incremental adaptive learning vector quantization (CIALVQ) method, which incrementally learns a self-adaptive style transfer matrix for mapping input patterns from style-conscious space onto style-free space. After style transformation, this problem is casted into a common character recognition task and an incremental learning vector quantization (ILVQ) classifier is used. In this framework, we consider two learning modes: supervised incremental learning and active incremental learning. In the latter mode, samples receiving low confidence from the classifier are requested class labels. We evaluated the classification performance of CIALVQ in two scenarios, interleaved test-then-train and style-specific classification on NIST hand-printed data sets. The results show that local style consistency improves the accuracies of both two test scenarios, and for both supervised and active incremental learning modes.     

Fuzzy classification by fuzzy labeled neural gas.

We extend the neural gas for supervised fuzzy classification. In this way we are able to learn crisp as well as fuzzy clustering, given labeled data. Based on the neural gas cost function, we propose three different ways to incorporate the additional class information into the learning algorithm. We demonstrate the effect on the location of the prototypes and the classification accuracy. Further, we show that relevance learning can be easily included.     

Multiple functional networks modeling for autism spectrum disorder diagnosis

Despite countless studies on autism spectrum disorder (ASD), diagnosis relies on specific behavioral criteria and neuroimaging biomarkers for the disorder are still relatively scarce and irrelevant for diagnostic workup. Many researchers have focused on functional networks of brain activities using resting‐state functional magnetic resonance imaging (rsfMRI) to diagnose brain diseases, including ASD. Although some existing methods are able to reveal the abnormalities in functional networks, they are either highly dependent on prior assumptions for modeling these networks or do not focus on latent functional connectivities (FCs) by considering discriminative relations among FCs in a nonlinear way. In this article, we propose a novel framework to model multiple networks of rsfMRI with data‐driven approaches. Specifically, we construct large‐scale functional networks with hierarchical clustering and find discriminative connectivity patterns between ASD and normal controls (NC). We then learn features and classifiers for each cluster through discriminative restricted Boltzmann machines (DRBMs). In the testing phase, each DRBM determines whether a test sample is ASD or NC, based on which we make a final decision with a majority voting strategy. We assess the diagnostic performance of the proposed method using public datasets and describe the effectiveness of our method by comparing it to competing methods. We also rigorously analyze FCs learned by DRBMs on each cluster and discover dominant FCs that play a major role in discriminating between ASD and NC. Hum Brain Mapp 38:5804–5821, 2017. © 2017 Wiley Periodicals, Inc.     

Self-organizing ARTMAP rule discovery

The self-organizing ARTMAP rule discovery (SOARD) system derives relationships among recognition classes during online learning. SOARD training on input/output pairs produces the basic competence of direct recognition of individual class labels for new test inputs. As a typical supervised system, it learns many-to-one maps, which recognize different inputs (Spot, Rex) as belonging to one class (dog). As an ARTMAP system, it also learns one-to-many maps, allowing a given input (Spot) to learn a new class (animal) without forgetting its previously learned output (dog), even as it corrects erroneous predictions (cat). As it learns individual input/output class predictions, SOARD employs distributed code representations that support online rule discovery. When the input Spot activates the classes dogand animal, confidence in the rule dog→animal begins to grow. When other inputs simultaneously activate classes cat and animal, confidence in the converse rule, animal→dog, decreases. Confidence in a self-organized rule is encoded as the weight in a path from one class node to the other. An experience-based mechanism modulates the rate of rule learning, to keep inaccurate predictions from creating false rules during early learning. Rules may be excitatory or inhibitory so that rule-based activation can add missing classes and remove incorrect ones. SOARD rule activation also enables inputs to learn to make direct predictions of output classes that they have never experienced during supervised training. When input Rex activates its learned class dog, the rule dog→animal indirectly activates the output class animal. The newly activated class serves as a teaching signal which allows input Rex to learn direct activation of the output class animal. Simulations using small-scale and large-scale datasets demonstrate functional properties of the SOARD system in both spatial and time-series domains.     

Clustering: A neural network approach

Clustering is a fundamental data analysis method. It is widely used for pattern recognition, feature extraction, vector quantization (VQ), image segmentation, function approximation, and data mining. As an unsupervised classification technique, clustering identifies some inherent structures present in a set of objects based on a similarity measure. Clustering methods can be based on statistical model identification (McLachlan & Basford, 1988) or competitive learning. In this paper, we give a comprehensive overview of competitive learning based clustering methods. Importance is attached to a number of competitive learning based clustering neural networks such as the self-organizing map (SOM), the learning vector quantization (LVQ), the neural gas, and the ART model, and clustering algorithms such as the $C$-means, mountain/subtractive clustering, and fuzzy $C$-means (FCM) algorithms. Associated topics such as the under-utilization problem, fuzzy clustering, robust clustering, clustering based on non-Euclidean distance measures, supervised clustering, hierarchical clustering as well as cluster validity are also described. Two examples are given to demonstrate the use of the clustering methods.     

An incremental network for on-line unsupervised classification and topology learning.

This paper presents an on-line unsupervised learning mechanism for unlabeled data that are polluted by noise. Using a similarity threshold-based and a local error-based insertion criterion, the system is able to grow incrementally and to accommodate input patterns of on-line non-stationary data distribution. A definition of a utility parameter, the error-radius, allows this system to learn the number of nodes needed to solve a task. The use of a new technique for removing nodes in low probability density regions can separate clusters with low-density overlaps and dynamically eliminate noise in the input data. The design of two-layer neural network enables this system to represent the topological structure of unsupervised on-line data, report the reasonable number of clusters, and give typical prototype patterns of every cluster without prior conditions such as a suitable number of nodes or a good initial codebook.     

On-line learning in changing environments with applications in supervised and unsupervised learning.

An adaptive on-line algorithm extending the learning of learning idea is proposed and theoretically motivated. Relying only on gradient flow information it can be applied to learning continuous functions or distributions, even when no explicit loss function is given and the Hessian is not available. The framework is applied for unsupervised and supervised learning. Its efficiency is demonstrated for drifting and switching non-stationary blind separation tasks of acoustic signals. Furthermore applications to classification (US postal service data set) and time-series prediction in changing environments are presented.     

Biologically Inspired Tensor Features

According to the research results reported in the past decades, it is well acknowledged that face recognition is not a trivial task. With the development of electronic devices, we are gradually revealing the secret of object recognition in the primate’s visual cortex. Therefore, it is time to reconsider face recognition by using biologically inspired features. In this paper, we represent face images by utilizing the C1 units, which correspond to complex cells in the visual cortex, and pool over S1 units by using a maximum operation to reserve only the maximum response of each local area of S1 units. The new representation is termed C1 Face. Because C1 Face is naturally a third-order tensor (or a three dimensional array), we propose three-way discriminative locality alignment (TWDLA), an extension of the discriminative locality alignment, which is a top-level discriminate manifold learning-based subspace learning algorithm. TWDLA has the following advantages: (1) it takes third-order tensors as input directly so the structure information can be well preserved; (2) it models the local geometry over every modality of the input tensors so the spatial relations of input tensors within a class can be preserved; (3) it maximizes the margin between a tensor and tensors from other classes over each modality so it performs well for recognition tasks and (4) it has no under sampling problem. Extensive experiments on YALE and FERET datasets show (1) the proposed C1Face representation can better represent face images than raw pixels and (2) TWDLA can duly preserve both the local geometry and the discriminative information over every modality for recognition.     

Training with differential outcomes enhances discriminative learning and visuospatial recognition memory in children born prematurely

Previous studies have demonstrated that discriminative learning is facilitated when a particular outcome is associated with each relation to be learned. When this training procedure is applied (the differential outcome procedure; DOP), learning is faster and more accurate than when the more common non-differential outcome procedure is used. This enhancement of accuracy and acquisition has been called the differential outcome effect (DOE). Our primary purpose in the present study was to explore the DOE in children born with great prematurity performing a discriminative learning task (Experiment 1) or a delayed visuospatial recognition task (Experiment 2). In Experiment 1, participants showed a faster learning and a better performance when differential outcomes were used. In Experiment 2, a significant DOE was also observed. That is, premature children performed the visuospatial recognition task better when they received differential outcomes following their correct responses. By contrast, the overall performance of full-term children was similar in both differential and non-differential conditions. These results are first to show that the DOP can enhance learning of conditional discriminations and recognition memory in children born prematurely with very low birth-weight.     

Improved learning of Riemannian metrics for exploratory analysis.

We have earlier introduced a principle for learning metrics, which shows how metric-based methods can be made to focus on discriminative properties of data. The main applications are in supervising unsupervised learning to model interesting variation in data, instead of modeling all variation as plain unsupervised learning does. The metrics are derived by approximations to an information-geometric formulation. In this paper, we review the theory, introduce better approximations to the distances, and show how to apply them in two different kinds of unsupervised methods: prototype-based and pairwise distance-based. The two examples are self-organizing maps and multidimensional scaling (Sammon's mapping).