Deep neural networks have long been criticized for lacking the ability to perform analogical visual reasoning. Here, we propose a neural network model to solve Raven's Progressive Matrices (RPM) - one of the standard intelligence tests in human psychology. Specifically, we design a reasoning block based on the well-known concept of prediction error (PE) in neuroscience. Our reasoning block uses convolution to extract abstract rules from high-level visual features of the 8 context images and generates the features of a predicted answer. PEs are then calculated between the predicted features and those of the 8 candidate answers, and are then passed to the next stage. We further integrate our novel reasoning blocks into a residual network and build a new Predictive Reasoning Network (PredRNet). Extensive experiments show that our proposed PredRNet achieves state-of-the-art average performance on several important RPM benchmarks. PredRNet also shows good generalization abilities in a variety of out-of-distribution scenarios and other visual reasoning tasks. Most importantly, our PredRNet forms low-dimensional representations of abstract rules and minimizes hierarchical prediction errors during model training, supporting the critical role of PE minimization in visual reasoning. Our work highlights the potential of using neuroscience theories to solve abstract visual reasoning problems in artificial intelligence. The code is available at https://github.com/ZjjConan/AVR-PredRNet.

Raven's Progressive Matrices (RPM), one of the standard intelligence tests in human psychology, has recently emerged as a powerful tool for studying abstract visual reasoning (AVR) abilities in machines. Although existing computational models for RPM problems achieve good performance, they require a large number of labeled training examples for supervised learning. In contrast, humans can efficiently solve unlabeled RPM problems after learning from only a few example questions. Here, we develop a semi-supervised learning (SSL) method, called RuleMatch, to train deep models with a small number of labeled RPM questions along with other unlabeled questions. Moreover, instead of using pixel-level augmentation in object perception tasks, we exploit the nature of RPM problems and augment the data at the level of abstract rules. Specifically, we disrupt the possible rules contained among context images in an RPM question and force the two augmented variants of the same unlabeled sample to obey the same abstract rule and predict a common pseudo label for training. Extensive experiments show that the proposed RuleMatch achieves state-of-the-art performance on two popular RAVEN datasets. Our work makes an important stride in aligning abstract analogical visual reasoning abilities in machines and humans. Our Code is at https://github.com/ZjjConan/AVR-RuleMatch.

Although the neural mechanism underlying risk decision has been extensively investigated, the neural origination of risk attitude and probability distortion need to be further elucidated. In this study, the Rescorla–Wagner model with learning rates a+/a- upon gain/loss evaluates the risky outcome and forms the subjective values of risky options through the learning process, and the softmax function of subjective values produces the choice probability between options. Our model demonstrates that risk attitude is determined by the undervaluation/overvaluation of risky outcome, the standard deviation of the subjective value, and the discrimination ability between subjective value. Our model further displays that overweighting/underweighting of small probabilities results from asymmetric learning rates and the discrimination ability between subjective value. These findings suggest that risk attitude and probability distortion share a common neural mechanism.

Silicon neuron circuits emulate the electrophysiological behavior of real neurons and conductances. A detailed model mapped onto silicon neuron can be beneficial to the improvement of circuit design. Here we present a dynamic systems approach to obtain a detailed mathematical model describing the dynamics of a biophysically realistic silicon neuron. The approximate analytic solution of its firing rate fits simulation data from our neuron chip fabricated using a 130 nm CMOS process very well. Meanwhile, transistors that contribute critically to variation of firing activities are discovered, which helps the improvement of neuron mismatch.

Emulation of decision making in neuromorphic system can be useful for developing remarkable bio-mimetic devices capable of implementing adaptive and interactive behaviors with the varying environment. Both experimental and theoretical studies reveal that the dynamics of NMDA receptors (NMDARs) plays a critical role in attractor dynamics of decision. However, the functionality of NMDARs has not been considered in current neuromorphic decision circuit. Here we present a novel method for the neuromorphic implementation of a two-variable decision circuit with NMDARs. Circuit simulations and theoretical analysis reveal that the decision circuit can present not only the winner-take-all mechanism but also the slow integration of sensory evidences, both of which are embedded in gradually ramp-up activities observed in our simulations and in the electrophysiological recording in monkey's experiments. We demonstrated that the decision circuit we built can generate reliable attractor dynamics capable of reproducing both neurophysiological and behavioral observations during decision tasks.

Understanding how to reproduce robust and reliable decision making behavior in neuromorphic systems can be useful for developing information processing architectures in subthreshold analog circuits as well as future emerging nano-technologies, that comprise inhomogeneous and unreliable components. To this end, we explore the computational properties of a recurrent neural network, implemented in a custom mixed signal analog/digital neuromorphic chip, for realizing perceptual decision-making, bi-stable perception, and working memory. The chip comprises conductance-based integrate-and-fire neurons and configurable synapses with realistic dynamics. These circuits are configured to implement a recurrent neural network, composed of excitatory and inhibitory pools of silicon neurons coupled with local excitation and global inhibition. We show how the interplay between excitation and inhibition produces competitive winner-take-all dynamics, which is a feature of decision-making and persistent activity models, and demonstrate that the system generates reliable dynamics capable of reproducing both neuro-physiological data and psycho-physical performances in coding and collective distributed computation.

Neuroscientists have carried out comprehensive experiments to reveal the neural mechanisms underlying the perceptual decision making that pervades daily life. These experiments have illuminated salient features of decision making, including probabilistic choice behavior, the ramping activity of decision-related neurons, and the dependence of decision time and accuracy on the difficulty of the task. Spiking network models have reproduced these features, and a two-dimensional mean field model has demonstrated that the saddle node structure underlies two-alternative decision making. Here, we reduced a spiking network model to an analytically tractable, partial integro-differential system and characterized not only multiple-choice decision behaviors but also the time course of neural activities underlying decisions, providing a mechanistic explanation for the observations noted in the experiments. First, we observed that a two-bump unstable steady state of the system is responsible for two-choice decision making, similar to the saddle node structure in the two-dimensional mean field model. However, for four-choice decision making, three types of unstable steady states collectively predominate the time course of the evolution from the initial state to the stable states. Second, the time constant of the unstable steady state can explain the fact that four-choice decision making requires a longer time than two-choice decision making. However, the quicker decision, given a stronger motion strength, cannot be explained by the time constant of the unstable steady state. Rather, the decision time can be attributed to the projection coefficient of the difference between the initial state and the unstable steady state on the eigenvector corresponding to the largest positive eigenvalue.

Our actions take place in space and time, but despite the role of time in decision theory and the growing acknowledgement that the encoding of time is crucial to behaviour, few studies have considered the interactions between neural codes for objects in space and for elapsed time during perceptual decisions. The speed-accuracy trade-off (SAT) provides a window into spatiotemporal interactions. Our hypothesis is that temporal coding determines the rate at which spatial evidence is integrated, controlling the SAT by gain modulation. Here, we propose that local cortical circuits are inherently suited to the relevant spatial and temporal coding. In simulations of an interval estimation task, we use a generic local-circuit model to encode time by 'climbing' activity, seen in cortex during tasks with a timing requirement. The model is a network of simulated pyramidal cells and inhibitory interneurons, connected by conductance synapses. A simple learning rule enables the network to quickly produce new interval estimates, which show signature characteristics of estimates by experimental subjects. Analysis of network dynamics formally characterizes this generic, local-circuit timing mechanism. In simulations of a perceptual decision task, we couple two such networks. Network function is determined only by spatial selectivity and NMDA receptor conductance strength; all other parameters are identical. To trade speed and accuracy, the timing network simply learns longer or shorter intervals, driving the rate of downstream decision processing by spatially non-selective input, an established form of gain modulation. Like the timing network's interval estimates, decision times show signature characteristics of those by experimental subjects. Overall, we propose, demonstrate and analyse a generic mechanism for timing, a generic mechanism for modulation of decision processing by temporal codes, and we make predictions for experimental verification.