Multi-timescale reinforcement learning in the brain.

To thrive in complex environments, animals and artificial agents must learn to act adaptively to maximize fitness and rewards. Such adaptive behavior can be learned through reinforcement learning1, a class of algorithms that has been successful at training artificial agents2-6 and at characterizing the firing of dopamine neurons in the midbrain7-9. In classical reinforcement learning, agents discount future rewards exponentially according to a single time scale, controlled by the discount factor. Here, we explore the presence of multiple timescales in biological reinforcement learning. We first show that reinforcement agents learning at a multitude of timescales possess distinct computational benefits. Next, we report that dopamine neurons in mice performing two behavioral tasks encode reward prediction error with a diversity of discount time constants. Our model explains the heterogeneity of temporal discounting in both cue-evoked transient responses and slower timescale fluctuations known as dopamine ramps. Crucially, the measured discount factor of individual neurons is correlated across the two tasks suggesting that it is a cell-specific property. Together, our results provide a new paradigm to understand functional heterogeneity in dopamine neurons, a mechanistic basis for the empirical observation that humans and animals use non-exponential discounts in many situations10-14, and open new avenues for the design of more efficient reinforcement learning algorithms.

Competitive integration of time and reward explains value-sensitive foraging decisions and frontal cortex ramping dynamics.

The ability to make advantageous decisions is critical for animals to ensure their survival. Patch foraging is a natural decision-making process in which animals decide when to leave a patch of depleting resources to search for a new one. To study the algorithmic and neural basis of patch foraging behavior in a controlled laboratory setting, we developed a virtual foraging task for head-fixed mice. Mouse behavior could be explained by ramp-to-threshold models integrating time and rewards antagonistically. Accurate behavioral modeling required inclusion of a slowly varying "patience" variable, which modulated sensitivity to time. To investigate the neural basis of this decision-making process, we performed dense electrophysiological recordings with Neuropixels probes broadly throughout frontal cortex and underlying subcortical areas. We found that decision variables from the reward integrator model were represented in neural activity, most robustly in frontal cortical areas. Regression modeling followed by unsupervised clustering identified a subset of neurons with ramping activity. These neurons' firing rates ramped up gradually in single trials over long time scales (up to tens of seconds), were inhibited by rewards, and were better described as being generated by a continuous ramp rather than a discrete stepping process. Together, these results identify reward integration via a continuous ramping process in frontal cortex as a likely candidate for the mechanism by which the mammalian brain solves patch foraging problems.

Distinct roles of the orbitofrontal cortex, ventral striatum, and dopamine neurons in counterfactual thinking of decision outcomes.

Individuals often assess past decisions by comparing what was gained with what would have been gained had they acted differently. Thoughts of past alternatives that counter what actually happened are called "counterfactuals." Recent theories emphasize the role of the prefrontal cortex in processing counterfactual outcomes in decision-making, although how subcortical regions contribute to this process remains to be elucidated. Here we report a clear distinction among the roles of the orbitofrontal cortex, ventral striatum and midbrain dopamine neurons in processing counterfactual outcomes in monkeys. Our findings suggest that actually gained and counterfactual outcome signals are both processed in the cortico-subcortical network constituted by these regions but in distinct manners and integrated only in the orbitofrontal cortex in a way to compare these outcomes. This study extends the prefrontal theory of counterfactual thinking and provides key insights regarding how the prefrontal cortex cooperates with subcortical regions to make decisions using counterfactual information.

Prior expectation enhances sensorimotor behavior by modulating population tuning and subspace activity in sensory cortex.

Prior knowledge facilitates our perception and goal-directed behaviors, particularly when sensory input is lacking or noisy. However, the neural mechanisms underlying the improvement in sensorimotor behavior by prior expectations remain unknown. In this study, we examine the neural activity in the middle temporal (MT) area of visual cortex while monkeys perform a smooth pursuit eye movement task with prior expectation of the visual target's motion direction. Prior expectations discriminately reduce the MT neural responses depending on their preferred directions, when the sensory evidence is weak. This response reduction effectively sharpens neural population direction tuning. Simulations with a realistic MT population demonstrate that sharpening the tuning can explain the biases and variabilities in smooth pursuit, suggesting that neural computations in the sensory area alone can underpin the integration of prior knowledge and sensory evidence. State-space analysis further supports this by revealing neural signals of prior expectations in the MT population activity that correlate with behavioral changes.

A neural mechanism for detecting object motion during self-motion.

Detection of objects that move in a scene is a fundamental computation performed by the visual system. This computation is greatly complicated by observer motion, which causes most objects to move across the retinal image. How the visual system detects scene-relative object motion during self-motion is poorly understood. Human behavioral studies suggest that the visual system may identify local conflicts between motion parallax and binocular disparity cues to depth and may use these signals to detect moving objects. We describe a novel mechanism for performing this computation based on neurons in macaque middle temporal (MT) area with incongruent depth tuning for binocular disparity and motion parallax cues. Neurons with incongruent tuning respond selectively to scene-relative object motion, and their responses are predictive of perceptual decisions when animals are trained to detect a moving object during self-motion. This finding establishes a novel functional role for neurons with incongruent tuning for multiple depth cues.

The role of state uncertainty in the dynamics of dopamine.

Reinforcement learning models of the basal ganglia map the phasic dopamine signal to reward prediction errors (RPEs). Conventional models assert that, when a stimulus predicts a reward with fixed delay, dopamine activity during the delay should converge to baseline through learning. However, recent studies have found that dopamine ramps up before reward in certain conditions even after learning, thus challenging the conventional models. In this work, we show that sensory feedback causes an unbiased learner to produce RPE ramps. Our model predicts that when feedback gradually decreases during a trial, dopamine activity should resemble a "bump," whose ramp-up phase should, furthermore, be greater than that of conditions where the feedback stays high. We trained mice on a virtual navigation task with varying brightness, and both predictions were empirically observed. In sum, our theoretical and experimental results reconcile the seemingly conflicting data on dopamine behaviors under the RPE hypothesis.

A Unified Framework for Dopamine Signals across Timescales.

Rapid phasic activity of midbrain dopamine neurons is thought to signal reward prediction errors (RPEs), resembling temporal difference errors used in machine learning. However, recent studies describing slowly increasing dopamine signals have instead proposed that they represent state values and arise independent from somatic spiking activity. Here we developed experimental paradigms using virtual reality that disambiguate RPEs from values. We examined dopamine circuit activity at various stages, including somatic spiking, calcium signals at somata and axons, and striatal dopamine concentrations. Our results demonstrate that ramping dopamine signals are consistent with RPEs rather than value, and this ramping is observed at all stages examined. Ramping dopamine signals can be driven by a dynamic stimulus that indicates a gradual approach to a reward. We provide a unified computational understanding of rapid phasic and slowly ramping dopamine signals: dopamine neurons perform a derivative-like computation over values on a moment-by-moment basis.

Gain Modulation as a Mechanism for Coding Depth from Motion Parallax in Macaque Area MT.

Observer translation produces differential image motion between objects that are located at different distances from the observer's point of fixation [motion parallax (MP)]. However, MP can be ambiguous with respect to depth sign (near vs far), and this ambiguity can be resolved by combining retinal image motion with signals regarding eye movement relative to the scene. We have previously demonstrated that both extra-retinal and visual signals related to smooth eye movements can modulate the responses of neurons in area MT of macaque monkeys, and that these modulations generate neural selectivity for depth sign. However, the neural mechanisms that govern this selectivity have remained unclear. In this study, we analyze responses of MT neurons as a function of both retinal velocity and direction of eye movement, and we show that smooth eye movements modulate MT responses in a systematic, temporally precise, and directionally specific manner to generate depth-sign selectivity. We demonstrate that depth-sign selectivity is primarily generated by multiplicative modulations of the response gain of MT neurons. Through simulations, we further demonstrate that depth can be estimated reasonably well by a linear decoding of a population of MT neurons with response gains that depend on eye velocity. Together, our findings provide the first mechanistic description of how visual cortical neurons signal depth from MP.SIGNIFICANCE STATEMENT Motion parallax is a monocular cue to depth that commonly arises during observer translation. To compute from motion parallax whether an object appears nearer or farther than the point of fixation requires combining retinal image motion with signals related to eye rotation, but the neurobiological mechanisms have remained unclear. This study provides the first mechanistic account of how this interaction takes place in the responses of cortical neurons. Specifically, we show that smooth eye movements modulate the gain of responses of neurons in area MT in a directionally specific manner to generate selectivity for depth sign from motion parallax. We also show, through simulations, that depth could be estimated from a population of such gain-modulated neurons.

The neural basis of depth perception from motion parallax.

In addition to depth cues afforded by binocular vision, the brain processes relative motion signals to perceive depth. When an observer translates relative to their visual environment, the relative motion of objects at different distances (motion parallax) provides a powerful cue to three-dimensional scene structure. Although perception of depth based on motion parallax has been studied extensively in humans, relatively little is known regarding the neural basis of this visual capability. We review recent advances in elucidating the neural mechanisms for representing depth-sign (near versus far) from motion parallax. We examine a potential neural substrate in the middle temporal visual area for depth perception based on motion parallax, and we explore the nature of the signals that provide critical inputs for disambiguating depth-sign.This article is part of the themed issue 'Vision in our three-dimensional world'.

A simple approach to ignoring irrelevant variables by population decoding based on multisensory neurons.

Sensory input reflects events that occur in the environment, but multiple events may be confounded in sensory signals. For example, under many natural viewing conditions, retinal image motion reflects some combination of self-motion and movement of objects in the world. To estimate one stimulus event and ignore others, the brain can perform marginalization operations, but the neural bases of these operations are poorly understood. Using computational modeling, we examine how multisensory signals may be processed to estimate the direction of self-motion (i.e., heading) and to marginalize out effects of object motion. Multisensory neurons represent heading based on both visual and vestibular inputs and come in two basic types: "congruent" and "opposite" cells. Congruent cells have matched heading tuning for visual and vestibular cues and have been linked to perceptual benefits of cue integration during heading discrimination. Opposite cells have mismatched visual and vestibular heading preferences and are ill-suited for cue integration. We show that decoding a mixed population of congruent and opposite cells substantially reduces errors in heading estimation caused by object motion. In addition, we present a general formulation of an optimal linear decoding scheme that approximates marginalization and can be implemented biologically by simple reinforcement learning mechanisms. We also show that neural response correlations induced by task-irrelevant variables may greatly exceed intrinsic noise correlations. Overall, our findings suggest a general computational strategy by which neurons with mismatched tuning for two different sensory cues may be decoded to perform marginalization operations that dissociate possible causes of sensory inputs.

A functional link between MT neurons and depth perception based on motion parallax.

As an observer translates, objects lying at different distances from the observer have differential image motion on the retina (motion parallax). It is well established psychophysically that humans perceive depth rather precisely from motion parallax and that extraretinal signals may be used to correctly perceive the sign of depth (near vs far) when binocular and pictorial depth cues are absent or weak. However, the neural basis for this capacity remains poorly understood. We have shown previously that neurons in the macaque middle temporal (MT) area combine retinal image motion with smooth eye movement command signals to signal depth sign from motion parallax. However, those studies were performed in animals that were required simply to track a visual target, thus precluding direct comparisons between neural activity and behavior. Here, we examine the activity of MT neurons in rhesus monkeys that were trained to discriminate depth sign based on motion parallax, in the absence of binocular disparity and pictorial depth cues. We find that the most sensitive MT neurons approach behavioral sensitivity, whereas the average neuron is twofold to threefold less sensitive than the animal. We also find that MT responses are predictive of perceptual decisions (independent of the visual stimulus), consistent with a role for MT in providing sensory signals for this behavior. Our findings suggest that, in addition to its established roles in processing stereoscopic depth, area MT is well suited to contribute to perception of depth based on motion parallax.

A novel role for visual perspective cues in the neural computation of depth.

As we explore a scene, our eye movements add global patterns of motion to the retinal image, complicating visual motion produced by self-motion or moving objects. Conventionally, it has been assumed that extraretinal signals, such as efference copy of smooth pursuit commands, are required to compensate for the visual consequences of eye rotations. We consider an alternative possibility: namely, that the visual system can infer eye rotations from global patterns of image motion. We visually simulated combinations of eye translation and rotation, including perspective distortions that change dynamically over time. We found that incorporating these 'dynamic perspective' cues allowed the visual system to generate selectivity for depth sign from motion parallax in macaque cortical area MT, a computation that was previously thought to require extraretinal signals regarding eye velocity. Our findings suggest neural mechanisms that analyze global patterns of visual motion to perform computations that require knowledge of eye rotations.

Joint representation of depth from motion parallax and binocular disparity cues in macaque area MT.

Perception of depth is based on a variety of cues, with binocular disparity and motion parallax generally providing more precise depth information than pictorial cues. Much is known about how neurons in visual cortex represent depth from binocular disparity or motion parallax, but little is known about the joint neural representation of these depth cues. We recently described neurons in the middle temporal (MT) area that signal depth sign (near vs far) from motion parallax; here, we examine whether and how these neurons also signal depth from binocular disparity. We find that most MT neurons in rhesus monkeys (Macaca Mulatta) are selective for depth sign based on both disparity and motion parallax cues. However, the depth-sign preferences (near or far) are not always aligned: 56% of MT neurons have matched depth-sign preferences ("congruent" cells) whereas the remaining 44% of neurons prefer near depth from motion parallax and far depth from disparity, or vice versa ("opposite" cells). For congruent cells, depth-sign selectivity increases when disparity cues are added to motion parallax, but this enhancement does not occur for opposite cells. This suggests that congruent cells might contribute to perceptual integration of depth cues. We also found that neurons are clustered in MT according to their depth tuning based on motion parallax, similar to the known clustering of MT neurons for binocular disparity. Together, these findings suggest that area MT is involved in constructing a representation of 3D scene structure that takes advantage of multiple depth cues available to mobile observers.

Modulation of V1 spike response by temporal interval of spatiotemporal stimulus sequence.

The spike activity of single neurons of the primary visual cortex (V1) becomes more selective and reliable in response to wide-field natural scenes compared to smaller stimuli confined to the classical receptive field (RF). However, it is largely unknown what aspects of natural scenes increase the selectivity of V1 neurons. One hypothesis is that modulation by surround interaction is highly sensitive to small changes in spatiotemporal aspects of RF surround. Such a fine-tuned modulation would enable single neurons to hold information about spatiotemporal sequences of oriented stimuli, which extends the role of V1 neurons as a simple spatiotemporal filter confined to the RF. In the current study, we examined the hypothesis in the V1 of awake behaving monkeys, by testing whether the spike response of single V1 neurons is modulated by temporal interval of spatiotemporal stimulus sequence encompassing inside and outside the RF. We used two identical Gabor stimuli that were sequentially presented with a variable stimulus onset asynchrony (SOA): the preceding one (S1) outside the RF and the following one (S2) in the RF. This stimulus configuration enabled us to examine the spatiotemporal selectivity of response modulation from a focal surround region. Although S1 alone did not evoke spike responses, visual response to S2 was modulated for SOA in the range of tens of milliseconds. These results suggest that V1 neurons participate in processing spatiotemporal sequences of oriented stimuli extending outside the RF.

Trial-to-trial variability of spike response of V1 and saccadic response time.

Single neurons in the primary visual cortex (V1) show variability in spike activity in response to an identical visual stimulus. In the current study, we examined the behavioral significance of the variability in spike activity of V1 neurons for visually guided saccades. We recorded single-cell activity from V1 of monkeys trained to detect and make saccades toward visual targets of varying contrast and analyzed trial-to-trial covariation between the onset time or firing rate of neural response and saccadic response time (RT). Neural latency (NL, the time of the first spike of neural response) was correlated with RT, whereas firing rate (FR) was not. When FR was computed with respect to target onset ignoring NL, a "false" correlation between FR and RT emerged. Multiple regression and partial correlation analyses on NL and FR for predictability of RT variability, as well as a simulation with artificial Poisson spike trains, supported the conclusion that the correlation between FR with respect to target onset and RT was mediated by a correlation between NL and RT, emphasizing the role of trial-to-trial variability of NL for extracting RT-related signals. We attempted to examine laminar differences in RT-related activity. Neurons recorded in the superficial layers tended to show a higher sensitivity to stimulus contrast and a lower correlation with RT compared with those in the lower layers, suggesting a sensory-to-motor transformation within V1 that follows the order of known anatomical connections. These results demonstrate that the trial-to-trial variability of neural response in V1 propagates to the stage of saccade execution, resulting in trial-to-trial variability of RT of a visually guided saccade.