Reinforcement Learning on Slow Features of High-Dimensional Input Streams

Humans and animals are able to learn complex behaviors based on a massive stream of sensory information from different modalities. Early animal studies have identified learning mechanisms that are based on reward and punishment such that animals tend to avoid actions that lead to punishment whereas rewarded actions are reinforced. However, most algorithms for reward-based learning are only applicable if the dimensionality of the state-space is sufficiently small or its structure is sufficiently simple. Therefore, the question arises how the problem of learning on high-dimensional data is solved in the brain. In this article, we propose a biologically plausible generic two-stage learning system that can directly be applied to raw high-dimensional input streams. The system is composed of a hierarchical slow feature analysis (SFA) network for preprocessing and a simple neural network on top that is trained based on rewards. We demonstrate by computer simulations that this generic architecture is able to learn quite demanding reinforcement learning tasks on high-dimensional visual input streams in a time that is comparable to the time needed when an explicit highly informative low-dimensional state-space representation is given instead of the high-dimensional visual input. The learning speed of the proposed architecture in a task similar to the Morris water maze task is comparable to that found in experimental studies with rats. This study thus supports the hypothesis that slowness learning is one important unsupervised learning principle utilized in the brain to form efficient state representations for behavioral learning.     

The Wisdom of Networks: A General Adaptation and Learning Mechanism of Complex Systems

I hypothesize that re‐occurring prior experience of complex systems mobilizes a fast response, whose attractor is encoded by their strongly connected network core. In contrast, responses to novel stimuli are often slow and require the weakly connected network periphery. Upon repeated stimulus, peripheral network nodes remodel the network core that encodes the attractor of the new response. This “core‐periphery learning” theory reviews and generalizes the heretofore fragmented knowledge on attractor formation by neural networks, periphery‐driven innovation, and a number of recent reports on the adaptation of protein, neuronal, and social networks. The core‐periphery learning theory may increase our understanding of signaling, memory formation, information encoding and decision‐making processes. Moreover, the power of network periphery‐related “wisdom of crowds” inventing creative, novel responses indicates that deliberative democracy is a slow yet efficient learning strategy developed as the success of a billion‐year evolution. Also see the video abstract here: https://youtu.be/IIjP7zWGjVE .     

Generalization of stochastic visuomotor rotations

Generalization studies examine the influence of perturbations imposed on one movement onto other movements. The strength of generalization is traditionally interpreted as a reflection of the similarity of the underlying neural representations. Uncertainty fundamentally affects both sensory integration and learning and is at the heart of many theories of neural representation. However, little is known about how uncertainty, resulting from variability in the environment, affects generalization curves. Here we extend standard movement generalization experiments to ask how uncertainty affects the generalization of visuomotor rotations. We find that although uncertainty affects how fast subjects learn, the perturbation generalizes independently of uncertainty.     

How Lateral Connections and Spiking Dynamics May Separate Multiple Objects Moving Together

Over successive stages, the ventral visual system of the primate brain develops neurons that respond selectively to particular objects or faces with translation, size and view invariance. The powerful neural representations found in Inferotemporal cortex form a remarkably rapid and robust basis for object recognition which belies the difficulties faced by the system when learning in natural visual environments. A central issue in understanding the process of biological object recognition is how these neurons learn to form separate representations of objects from complex visual scenes composed of multiple objects. We show how a one-layer competitive network comprised of ‘spiking’ neurons is able to learn separate transformation-invariant representations (exemplified by one-dimensional translations) of visual objects that are always seen together moving in lock-step, but separated in space. This is achieved by combining ‘Mexican hat’ functional lateral connectivity with cell firing-rate adaptation to temporally segment input representations of competing stimuli through anti-phase oscillations (perceptual cycles). These spiking dynamics are quickly and reliably generated, enabling selective modification of the feed-forward connections to neurons in the next layer through Spike-Time-Dependent Plasticity (STDP), resulting in separate translation-invariant representations of each stimulus. Variations in key properties of the model are investigated with respect to the network’s ability to develop appropriate input representations and subsequently output representations through STDP. Contrary to earlier rate-coded models of this learning process, this work shows how spiking neural networks may learn about more than one stimulus together without suffering from the ‘superposition catastrophe’. We take these results to suggest that spiking dynamics are key to understanding biological visual object recognition.     

Synchronization of neuronal assemblies in reciprocally connected cortical areas

Summary To investigate scene segmentation in the visual system we present a model of two reciprocally connected visual areas comprising spiking neurons. The peripheral area P is modeled similar to the primary visual cortex, while the central area C is modeled as an associative memory representing stimulus objects according to Hebbian learning. Without feedback from area C, spikes corresponding to stimulus representations in P are synchronized only locally (slow state). Feedback from C can induce fast oscillations and an increase of synchronization ranges (fast state). Presenting a superposition of several stimulus objects, scene segmentation happens on a time scale of hundreds of milliseconds by alternating epochs of the slow and fast state, where neurons representing the same object are simultaneously in the fast state. We relate our simulation results to various phenomena observed in neurophysiological experiments, such as stimulus-dependent synchronization of fast oscillations, synchronization on different time scales, ongoing activity, and attention-dependent neural activity.     

Sequence learning recodes cortical representations instead of strengthening initial ones

We contrast two computational models of sequence learning. The associative learner posits that learning proceeds by strengthening existing association weights. Alternatively, recoding posits that learning creates new and more efficient representations of the learned sequences. Importantly, both models propose that humans act as optimal learners but capture different statistics of the stimuli in their internal model. Furthermore, these models make dissociable predictions as to how learning changes the neural representation of sequences. We tested these predictions by using fMRI to extract neural activity patterns from the dorsal visual processing stream during a sequence recall task. We observed that only the recoding account can explain the similarity of neural activity patterns, suggesting that participants recode the learned sequences using chunks. We show that associative learning can theoretically store only very limited number of overlapping sequences, such as common in ecological working memory tasks, and hence an efficient learner should recode initial sequence representations.     

Local adaptations of two naturally occurring neuronal conductances,gK+(A) and gK+(Ca), allow for associative conditioning and contiguity judgements in artificial neural networks

Features of two potassium conductances implicated in the acquisition of conditioned reflexes, the slow calcium dependent conductance (gK+(Ca)) and the fast transient conductance (gK+(A)), were incorporated into a 6 × 6 element artificial neural network. Adaptive algorithms derived from observations of cortical neurons during associative learning changed gK+(A) in proportion to the product of this current and an EPSP-induced second messenger concentration, and changed gK+(Ca) as a function of a spike-induced second messenger concentration. This network concurrently acquired two distinct representations in response to presentation of stimuli: one resembled associative conditioning (defined in terms of its senstivity to forward pairing vs. simultaneous or backward pairing); the other reflected contiguous pairings of stimuli. The acquisition of one representation did not markedly interfere with acquisition of the other. This network may accordingly serve as an example of a self-organizing system which minimizes the postulated inherent cross talk between functionally dissiminar representations (Minsky and Papert 1988).     

Information processing in the olfactory systems of insects and vertebrates

Insects and vertebrates separately evolved remarkably similar mechanisms to process olfactory information. Odors are sampled by huge numbers of receptor neurons, which converge type-wise upon a much smaller number of principal neurons within glomeruli. There, odor information is transformed by inhibitory interneuron-mediated, cross-glomerular circuit interactions that impose slow temporal structures and fast oscillations onto the firing patterns of principal neurons. The transformations appear to improve signal-to-noise characteristics, define odor categories, achieve precise odor identification, extract invariant features, and begin the process of sparsening the neural representations of odors for efficient discrimination, memorization, and recognition.     

Motor learning with unstable neural representations

It is often assumed that learning takes place by changing an otherwise stable neural representation. To test this assumption, we studied changes in the directional tuning of primate motor cortical neurons during reaching movements performed in familiar and novel environments. During the familiar task, tuning curves exhibited slow random drift. During learning of the novel task, random drift was accompanied by systematic shifts of tuning curves. Our analysis suggests that motor learning is based on a surprisingly unstable neural representation. To explain these results, we propose that motor cortex is a redundant neural network, i.e., any single behavior can be realized by multiple configurations of synaptic strengths. We further hypothesize that synaptic modifications underlying learning contain a random component, which causes wandering among synaptic configurations with equivalent behaviors but different neural representations. We use a simple model to explore the implications of these assumptions.     

Estimating the hidden learning representations

Successful adaptation relies on the ability to learn the consequence of our actions in different environments. However, understanding the neural bases of this ability still represents one of the great challenges of system neuroscience. In fact, the neuronal plasticity changes occurring during learning cannot be fully controlled experimentally and their evolution is hidden. Our approach is to provide hypotheses about the structure and dynamics of the hidden plasticity changes using behavioral learning theory. In fact, behavioral models of animal learning provide testable predictions about the hidden learning representations by formalizing their relation with the observables of the experiment (stimuli, actions and outcomes). Thus, we can understand whether and how the predicted learning processes are represented at the neural level by estimating their evolution and correlating them with neural data. Here, we present a bayesian model approach to estimate the evolution of the internal learning representations from the observations of the experiment (state estimation), and to identify the set of models' parameters (parameter estimation) and the class of behavioral model (model selection) that are most likely to have generated a given sequence of actions and outcomes. More precisely, we use Sequential Monte Carlo methods for state estimation and the maximum likelihood principle (MLP) for model selection and parameter estimation. We show that the method recovers simulated trajectories of learning sessions on a single-trial basis and provides predictions about the activity of different categories of neurons that should participate in the learning process. By correlating the estimated evolutions of the learning variables, we will be able to test the validity of different models of instrumental learning and possibly identify the neural bases of learning.     

A Self-Organizing Model of the Visual Development of Hand-Centred Representations

We show how hand-centred visual representations could develop in the primate posterior parietal and premotor cortices during visually guided learning in a self-organizing neural network model. The model incorporates trace learning in the feed-forward synaptic connections between successive neuronal layers. Trace learning encourages neurons to learn to respond to input images that tend to occur close together in time. We assume that sequences of eye movements are performed around individual scenes containing a fixed hand-object configuration. Trace learning will then encourage individual cells to learn to respond to particular hand-object configurations across different retinal locations. The plausibility of this hypothesis is demonstrated in computer simulations.     

Neural Modularity Helps Organisms Evolve to Learn New Skills without Forgetting Old Skills

A long-standing goal in artificial intelligence is creating agents that can learn a variety of different skills for different problems. In the artificial intelligence subfield of neural networks, a barrier to that goal is that when agents learn a new skill they typically do so by losing previously acquired skills, a problem called catastrophic forgetting. That occurs because, to learn the new task, neural learning algorithms change connections that encode previously acquired skills. How networks are organized critically affects their learning dynamics. In this paper, we test whether catastrophic forgetting can be reduced by evolving modular neural networks. Modularity intuitively should reduce learning interference between tasks by separating functionality into physically distinct modules in which learning can be selectively turned on or off. Modularity can further improve learning by having a reinforcement learning module separate from sensory processing modules, allowing learning to happen only in response to a positive or negative reward. In this paper, learning takes place via neuromodulation, which allows agents to selectively change the rate of learning for each neural connection based on environmental stimuli (e.g. to alter learning in specific locations based on the task at hand). To produce modularity, we evolve neural networks with a cost for neural connections. We show that this connection cost technique causes modularity, confirming a previous result, and that such sparsely connected, modular networks have higher overall performance because they learn new skills faster while retaining old skills more and because they have a separate reinforcement learning module. Our results suggest (1) that encouraging modularity in neural networks may help us overcome the long-standing barrier of networks that cannot learn new skills without forgetting old ones, and (2) that one benefit of the modularity ubiquitous in the brains of natural animals might be to alleviate the problem of catastrophic forgetting.     

Learning to smell the roses: experience-dependent neural plasticity in human piriform and orbitofrontal cortices

It is widely presumed that odor quality is a direct outcome of odorant structure, but human studies indicate that molecular knowledge of an odorant is not always sufficient to predict odor quality. Indeed, the same olfactory input may generate different odor percepts depending on prior learning and experience. Combining functional magnetic resonance imaging with an olfactory paradigm of perceptual learning, we examined how sensory experience modifies odor perception and odor quality coding in the human brain. Prolonged exposure to a target odorant enhanced perceptual differentiation for odorants related in odor quality or functional group, an effect that was paralleled by learning-induced response increases in piriform cortex and orbitofrontal cortex (OFC). Critically, the magnitude of OFC activation predicted subsequent improvement in behavioral differentiation. Our findings suggest that neural representations of odor quality can be rapidly updated through mere perceptual experience, a mechanism that may underlie the development of odor perception.     

Neuro-robotics study on integrative learning of proactive visual attention and motor behaviors

The current paper proposes a novel model for integrative learning of proactive visual attention and sensory-motor control as inspired by the premotor theory of visual attention. The model is characterized by coupling a slow dynamics network with a fast dynamics network and by inheriting our prior proposed multiple timescales recurrent neural networks model (MTRNN) that may correspond to the fronto-parietal networks in the cortical brains. The neuro-robotics experiments in a task of manipulating multiple objects utilizing the proposed model demonstrated that some degrees of generalization in terms of position and object size variation can be achieved by organizing seamless integration of the proactive object-related visual attention and the related sensory-motor control into a set of action primitives in the distributed neural activities appearing in the fast dynamics network. It was also shown that such action primitives can be combined in compositional ways in acquiring novel actions in the slow dynamics network. The experimental results presented substantiate the premotor theory of visual attention.     

Distributed representations accelerate evolution of adaptive behaviours

Animals with rudimentary innate abilities require substantial learning to transform those abilities into useful skills, where a skill can be considered as a set of sensory–motor associations. Using linear neural network models, it is proved that if skills are stored as distributed representations, then within-lifetime learning of part of a skill can induce automatic learning of the remaining parts of that skill. More importantly, it is shown that this “free-lunch” learning (FLL) is responsible for accelerated evolution of skills, when compared with networks which either 1) cannot benefit from FLL or 2) cannot learn. Specifically, it is shown that FLL accelerates the appearance of adaptive behaviour, both in its innate form and as FLL-induced behaviour, and that FLL can accelerate the rate at which learned behaviours become innate.     

A Model of Self-Organizing Head-Centered Visual Responses in Primate Parietal Areas

We present a hypothesis for how head-centered visual representations in primate parietal areas could self-organize through visually-guided learning, and test this hypothesis using a neural network model. The model consists of a competitive output layer of neurons that receives afferent synaptic connections from a population of input neurons with eye position gain modulated retinal receptive fields. The synaptic connections in the model are trained with an associative trace learning rule which has the effect of encouraging output neurons to learn to respond to subsets of input patterns that tend to occur close together in time. This network architecture and synaptic learning rule is hypothesized to promote the development of head-centered output neurons during periods of time when the head remains fixed while the eyes move. This hypothesis is demonstrated to be feasible, and each of the core model components described is tested and found to be individually necessary for successful self-organization.     

Beetles, boxes and brain cells: neural mechanisms underlying valuation and learning

Sensory cues in the environment can predict the availability of reward. Through experience, humans and animals learn these predictions and use them to guide their actions. For example, we can learn to discriminate chanterelles from ordinary champignons through experience. Assuming the development of a taste for the complex and lingering flavors of chanterelles, we therefore learn to value the same action--picking mushrooms--differentially depending upon the appearance of a mushroom. One major goal of cognitive neuroscience is to understand the neural mechanisms that underlie this sort of learning. Because the acquisition of rewards motivates much behavior, recent efforts have focused on describing the neural signals related to learning the value of stimuli and actions. Neurons in the basal ganglia, in midbrain dopamine areas, in frontal and parietal cortices and in other brain areas, all modulate their activity in relation to aspects of learning. By training monkeys on various behavioral tasks, recent studies have begun to characterize how neural signals represent distinct processes, such as the timing of events, motivation, absolute (objective) and relative (subjective) valuation, and the formation of associative links between stimuli and potential actions. In addition, a number of studies have either further characterized dopamine signals or sought to determine how such signaling might interact with target structures, such as the striatum and rhinal cortex, to underlie learning.     

How Attention Can Create Synaptic Tags for the Learning of Working Memories in Sequential Tasks

Intelligence is our ability to learn appropriate responses to new stimuli and situations. Neurons in association cortex are thought to be essential for this ability. During learning these neurons become tuned to relevant features and start to represent them with persistent activity during memory delays. This learning process is not well understood. Here we develop a biologically plausible learning scheme that explains how trial-and-error learning induces neuronal selectivity and working memory representations for task-relevant information. We propose that the response selection stage sends attentional feedback signals to earlier processing levels, forming synaptic tags at those connections responsible for the stimulus-response mapping. Globally released neuromodulators then interact with tagged synapses to determine their plasticity. The resulting learning rule endows neural networks with the capacity to create new working memory representations of task relevant information as persistent activity. It is remarkably generic: it explains how association neurons learn to store task-relevant information for linear as well as non-linear stimulus-response mappings, how they become tuned to category boundaries or analog variables, depending on the task demands, and how they learn to integrate probabilistic evidence for perceptual decisions.