Using deep neural networks as a guide for modeling human planning

When developing models in cognitive science, researchers typically start with their own intuitions about human behavior in a given task and then build in mechanisms that explain additional aspects of the data. This refinement step is often hindered by how difficult it is to distinguish the unpredictable randomness of people's decisions from meaningful deviations between those decisions and the model. One solution for this problem is to compare the model against deep neural networks trained on behavioral data, which can detect almost any pattern given sufficient data. Here, we apply this method to the domain of planning with a heuristic search model for human play in 4-in-a-row, a combinatorial game where participants think multiple steps into the future. Using a data set consisting of 10,874,547 games, we train deep neural networks to predict human moves and find that they accurately do so while capturing meaningful patterns in the data. Thus, deviations between the model and the best network allow us to identify opportunities for model improvement despite starting with a model that has undergone substantial testing in previous work. Based on this analysis, we add three extensions to the model that range from a simple opening bias to specific adjustments regarding endgame planning. Overall, our work demonstrates the advantages of model comparison with a high-performance deep neural network as well as the feasibility of scaling cognitive models to massive data sets for systematically investigating the processes underlying human sequential decision-making.

Statistical inference on representational geometries

Neuroscience has recently made much progress, expanding the complexity of both neural activity measurements and brain-computational models. However, we lack robust methods for connecting theory and experiment by evaluating our new big models with our new big data. Here, we introduce new inference methods enabling researchers to evaluate and compare models based on the accuracy of their predictions of representational geometries: A good model should accurately predict the distances among the neural population representations (e.g. of a set of stimuli). Our inference methods combine novel 2-factor extensions of crossvalidation (to prevent overfitting to either subjects or conditions from inflating our estimates of model accuracy) and bootstrapping (to enable inferential model comparison with simultaneous generalization to both new subjects and new conditions). We validate the inference methods on data where the ground-truth model is known, by simulating data with deep neural networks and by resampling of calcium-imaging and functional MRI data. Results demonstrate that the methods are valid and conclusions generalize correctly. These data analysis methods are available in an open-source Python toolbox (rsatoolbox.readthedocs.io).

Point estimate observers: A new class of models for perceptual decision making

Bayesian optimal inference is often heralded as a principled, general framework for human perception. However, optimal inference requires integration over all possible world states, which quickly becomes intractable in complex real-world settings. Additionally, deviations from optimal inference have been observed in human decisions. A number of approximation methods have previously been suggested, such as sampling methods. In this study, we additionally propose point estimate observers, which evaluate only a single best estimate of the world state per response category. We compare the predicted behavior of these model observers to human decisions in five perceptual categorization tasks. Compared to the Bayesian observer, the point estimate observer loses decisively in one task, ties in two and wins in two tasks. Two sampling observers also improve upon the Bayesian observer, but in a different set of tasks. Thus, none of the existing general observer models appears to fit human perceptual decisions in all situations, but the point estimate observer is competitive with other observer models and may provide another stepping stone for future model development. (PsycInfo Database Record (c) 2023 APA, all rights reserved).

Deep neural models for color classification and color constancy

Color constancy is our ability to perceive constant colors across varying illuminations. Here, we trained deep neural networks to be color constant and evaluated their performance with varying cues. Inputs to the networks consisted of two-dimensional images of simulated cone excitations derived from three-dimensional (3D) rendered scenes of 2,115 different 3D shapes, with spectral reflectances of 1,600 different Munsell chips, illuminated under 278 different natural illuminations. The models were trained to classify the reflectance of the objects. Testing was done with four new illuminations with equally spaced CIEL*a*b* chromaticities, two along the daylight locus and two orthogonal to it. High levels of color constancy were achieved with different deep neural networks, and constancy was higher along the daylight locus. When gradually removing cues from the scene, constancy decreased. Both ResNets and classical ConvNets of varying degrees of complexity performed well. However, DeepCC, our simplest sequential convolutional network, represented colors along the three color dimensions of human color vision, while ResNets showed a more complex representation.

Disentangling bottom-up versus top-down and low-level versus high-level influences on eye movements over time

Bottom-up and top-down as well as low-level and high-level factors influence where we fixate when viewing natural scenes. However, the importance of each of these factors and how they interact remains a matter of debate. Here, we disentangle these factors by analyzing their influence over time. For this purpose, we develop a saliency model that is based on the internal representation of a recent early spatial vision model to measure the low-level, bottom-up factor. To measure the influence of high-level, bottom-up features, we use a recent deep neural network-based saliency model. To account for top-down influences, we evaluate the models on two large data sets with different tasks: first, a memorization task and, second, a search task. Our results lend support to a separation of visual scene exploration into three phases: the first saccade, an initial guided exploration characterized by a gradual broadening of the fixation density, and a steady state that is reached after roughly 10 fixations. Saccade-target selection during the initial exploration and in the steady state is related to similar areas of interest, which are better predicted when including high-level features. In the search data set, fixation locations are determined predominantly by top-down processes. In contrast, the first fixation follows a different fixation density and contains a strong central fixation bias. Nonetheless, first fixations are guided strongly by image properties, and as early as 200 ms after image onset, fixations are better predicted by high-level information. We conclude that any low-level, bottom-up factors are mainly limited to the generation of the first saccade. All saccades are better explained when high-level features are considered, and later, this high-level, bottom-up control can be overruled by top-down influences.

Searchers adjust their eye-movement dynamics to target characteristics in natural scenes

When searching a target in a natural scene, it has been shown that both the target's visual properties and similarity to the background influence whether and how fast humans are able to find it. So far, it was unclear whether searchers adjust the dynamics of their eye movements (e.g., fixation durations, saccade amplitudes) to the target they search for. In our experiment, participants searched natural scenes for six artificial targets with different spatial frequency content throughout eight consecutive sessions. High-spatial frequency targets led to smaller saccade amplitudes and shorter fixation durations than low-spatial frequency targets if target identity was known. If a saccade was programmed in the same direction as the previous saccade, fixation durations and successive saccade amplitudes were not influenced by target type. Visual saliency and empirical fixation density at the endpoints of saccades which maintain direction were comparatively low, indicating that these saccades were less selective. Our results suggest that searchers adjust their eye movement dynamics to the search target efficiently, since previous research has shown that low-spatial frequencies are visible farther into the periphery than high-spatial frequencies. We interpret the saccade direction specificity of our effects as an underlying separation into a default scanning mechanism and a selective, target-dependent mechanism.

Temporal evolution of the central fixation bias in scene viewing

When watching the image of a natural scene on a computer screen, observers initially move their eyes toward the center of the image-a reliable experimental finding termed central fixation bias. This systematic tendency in eye guidance likely masks attentional selection driven by image properties and top-down cognitive processes. Here, we show that the central fixation bias can be reduced by delaying the initial saccade relative to image onset. In four scene-viewing experiments we manipulated observers' initial gaze position and delayed their first saccade by a specific time interval relative to the onset of an image. We analyzed the distance to image center over time and show that the central fixation bias of initial fixations was significantly reduced after delayed saccade onsets. We additionally show that selection of the initial saccade target strongly depended on the first saccade latency. A previously published model of saccade generation was extended with a central activation map on the initial fixation whose influence declined with increasing saccade latency. This extension was sufficient to replicate the central fixation bias from our experiments. Our results suggest that the central fixation bias is generated by default activation as a response to the sudden image onset and that this default activation pattern decreases over time. Thus, it may often be preferable to use a modified version of the scene viewing paradigm that decouples image onset from the start signal for scene exploration to explicitly reduce the central fixation bias.

An image-computable psychophysical spatial vision model

A large part of classical visual psychophysics was concerned with the fundamental question of how pattern information is initially encoded in the human visual system. From these studies a relatively standard model of early spatial vision emerged, based on spatial frequency and orientation-specific channels followed by an accelerating nonlinearity and divisive normalization: contrast gain-control. Here we implement such a model in an image-computable way, allowing it to take arbitrary luminance images as input. Testing our implementation on classical psychophysical data, we find that it explains contrast detection data including the ModelFest data, contrast discrimination data, and oblique masking data, using a single set of parameters. Leveraging the advantage of an image-computable model, we test our model against a recent dataset using natural images as masks. We find that the model explains these data reasonably well, too. To explain data obtained at different presentation durations, our model requires different parameters to achieve an acceptable fit. In addition, we show that contrast gain-control with the fitted parameters results in a very sparse encoding of luminance information, in line with notions from efficient coding. Translating the standard early spatial vision model to be image-computable resulted in two further insights: First, the nonlinear processing requires a denser sampling of spatial frequency and orientation than optimal coding suggests. Second, the normalization needs to be fairly local in space to fit the data obtained with natural image masks. Finally, our image-computable model can serve as tool in future quantitative analyses: It allows optimized stimuli to be used to test the model and variants of it, with potential applications as an image-quality metric. In addition, it may serve as a building block for models of higher level processing.

Likelihood-based parameter estimation and comparison of dynamical cognitive models

Dynamical models of cognition play an increasingly important role in driving theoretical and experimental research in psychology. Therefore, parameter estimation, model analysis and comparison of dynamical models are of essential importance. In this article, we propose a maximum likelihood approach for model analysis in a fully dynamical framework that includes time-ordered experimental data. Our methods can be applied to dynamical models for the prediction of discrete behavior (e.g., movement onsets); in particular, we use a dynamical model of saccade generation in scene viewing as a case study for our approach. For this model, the likelihood function can be computed directly by numerical simulation, which enables more efficient parameter estimation including Bayesian inference to obtain reliable estimates and corresponding credible intervals. Using hierarchical models inference is even possible for individual observers. Furthermore, our likelihood approach can be used to compare different models. In our example, the dynamical framework is shown to outperform nondynamical statistical models. Additionally, the likelihood based evaluation differentiates model variants, which produced indistinguishable predictions on hitherto used statistics. Our results indicate that the likelihood approach is a promising framework for dynamical cognitive models. (PsycINFO Database Record

Perception of light source distance from shading patterns

Varying the distance of a light source from an object alters both the intensity and spatial distribution of surface shading patterns. We tested whether observers can use such cues to infer light source distance. Participants viewed stereoscopic renderings of rough objects with diffuse and glossy surfaces, which were illuminated by a point source at a range of distances. In one task, they adjusted the position of a small probe dot in three dimensions to report the apparent location of the light in the scene. In a second task, they adjusted the shading on one object (by moving an invisible light source) until it appeared to be illuminated from the same distance as another object. Participants' responses increased linearly with the true light source distance, suggesting that they have clear intuitions about how light source distance affects shading patterns for a variety of different surfaces. However, there were also systematic errors: Subjects overestimated light source distance in the probe adjustment task, and in both experiments, roughness and glossiness affected responses. We find the pattern of results is predicted surprisingly well by a simplistic model based only on the area of the image that exceeds a certain intensity threshold. Thus, although subjects can report light source distance, they may rely on simple--sometimes erroneous--heuristics to do so.