Organizations of all sizes, across all industries and domains are leveraging artificial intelligence (AI) technologies to solve some of their biggest challenges around operations, customer experience, and much more. However, due to the probabilistic nature of AI, the risks associated with it are far greater than traditional technologies. Research has shown that these risks can range anywhere from regulatory, compliance, reputational, and user trust, to financial and even societal risks. Depending on the nature and size of the organization, AI technologies can pose a significant risk, if not used in a responsible way. This position paper seeks to present a brief introduction to AI governance, which is a framework designed to oversee the responsible use of AI with the goal of preventing and mitigating risks. Having such a framework will not only manage risks but also gain maximum value out of AI projects and develop consistency for organization-wide adoption of AI.

This manuscript outlines a viable approach for training and evaluating machine learning systems for high-stakes, human-centered, or regulated applications using common Python programming tools. The accuracy and intrinsic interpretability of two types of constrained models, monotonic gradient boosting machines and explainable neural networks, a deep learning architecture well-suited for structured data, are assessed on simulated data and publicly available mortgage data. For maximum transparency and the potential generation of personalized adverse action notices, the constrained models are analyzed using post-hoc explanation techniques including plots of partial dependence and individual conditional expectation and with global and local Shapley feature importance. The constrained model predictions are also tested for disparate impact and other types of discrimination using measures with long-standing legal precedents, adverse impact ratio, marginal effect, and standardized mean difference, along with straightforward group fairness measures. By combining interpretable models, post-hoc explanations, and discrimination testing with accessible software tools, this text aims to provide a template workflow for machine learning applications that require high accuracy and interpretability and that mitigate risks of discrimination.

Explainable machine learning (ML) enables human learning from ML, human appeal of automated model decisions, regulatory compliance, and security audits of ML models. Explainable ML (i.e. explainable artificial intelligence or XAI) has been implemented in numerous open source and commercial packages and explainable ML is also an important, mandatory, or embedded aspect of commercial predictive modeling in industries like financial services. However, like many technologies, explainable ML can be misused, particularly as a faulty safeguard for harmful black-boxes, e.g. fairwashing or scaffolding, and for other malevolent purposes like stealing models and sensitive training data. To promote best-practice discussions for this already in-flight technology, this short text presents internal definitions and a few examples before covering the proposed guidelines. This text concludes with a seemingly natural argument for the use of interpretable models and explanatory, debugging, and disparate impact testing methods in life- or mission-critical ML systems.

The black-box Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) criminal risk assessment instrument (RAI) is analyzed for confounding racial bias and a novel procedure is proposed for remediating bias from individual criminal risk predictions. A repeatable global versus local analysis motif is introduced in which global and local model behavior are compared to debug and diagnose unwanted bias in a black-box prediction system using tools such as surrogate models, gradient boosting machine feature importance, leave-onecovariate-out (LOCO) feature importance, partial dependence plots, and individual conditional expectation (ICE) plots. LOCO-derived feature importance is also used to remove prediction contributions from bias-inducing input features. The proposed global versus local approach and remediation strategy can be applied to many blackbox and machine learning (ML) decision-making systems.

Our attentional focus is constantly shifting: In one moment, our attention may be intently concentrated on a specific spot, whereas in another moment we might spread our attention more broadly. Although much is known about the mechanisms by which we shift our visual attention from place to place, relatively little is known about how we shift the aperture of attention from more narrowly to more broadly focused. Here we introduce a novel attentional distribution task to examine the neural mechanisms underlying this process. In this task, participants are presented with an informative cue that indicates the location of an upcoming target. This cue can be perfectly predictive of the exact target location, or it can indicate-with varying degrees of certainty-approximately where the target might appear. This cue is followed by a preparatory period in which there is nothing on the screen except a central fixation cross. Using scalp EEG, we examined neural activity during this preparatory period. We find that, with decreasing certainty regarding the precise location of the impending target, participant RTs increased whereas target identification accuracy decreased. Additionally, the multivariate pattern of preparatory period visual cortical alpha (8-12 Hz) activity encoded attentional distribution. This alpha encoding was predictive of behavioral accuracy and RT nearly 1 sec later. These results offer insight into the neural mechanisms underlying how we use information to guide our attentional distribution and how that influences behavior.

INTRODUCTION. One of the most fascinating phenomena in stereopsis is the profound hysteresis reported by Fender and Julesz (1967), in which the depth percept with increasing disparity persisted long past the point of depth recovery with decreasing disparity. To control retinal disparity without vergence eye movements, they stabilized the stimuli on the retinas with an eye tracker. We now report that stereo hysteresis can be observed simply by rotating the binocular stereogram image. As the image rotates, the horizontal disparities rotate to become vertical, then horizontal with inverted sign, then vertical again before returning to the original orientation. The depth shows an interesting popout effect, almost as though the depth was rapidly switching on and off, despite the inherently sinusoidal change in the horizontal disparity vector. METHODS. This stimulus was set up electronically in a circular format so that the random-dot field could be dynamically replaced, eliminating any cue to cyclorotation. Noise density was proportional eccentricity to fade the stimulus near the zero-disparity fixation target, allowing us to verify that fixation was held accurately at zero disparity. RESULTS AND DISCUSSION. For both the invariant and the dynamic noise, profound hysteresis of many seconds delay was found in eight observers for both the onset and offset of the perceived depth surface. This hysteresis was far longer than the <1000 ms reaction time to respond to changes in disparity. A similar hysteresis was obtained for depth popout from vertical disparity modulation of a fixed horizontal disparity. Conversely, sinusoidal modulation of the horizontal disparity to match the horizontal vector component of the disparity rotation did not show the popout effect, which thus seems to be a function of the interaction between horizontal and vertical disparities and is attributable to the time course of surface interpolation processes for the perceived depth structure. Meeting abstract presented at VSS 2012

Disparity vergence is commonly viewed as being controlled by at least two mechanisms, an open-loop vergence-specific burst mechanism analogous to the ballistic drive of saccades, and a closed-loop feedback mechanism controlled by the disparity error. We show that human vergence dynamics for disparity jumps of a large textured field have a typical time course consistent with predominant control by the open-loop vergence-specific burst mechanism, although various subgroups of the population show radically different vergence behaviors. Some individuals show markedly slow divergence responses, others slow convergence responses, others slow responses in both vergence directions, implying that the two vergence directions have separate control mechanisms. The faster time courses usually had time-symmetric velocity waveforms implying open-loop burst control, while the slow response waveforms were usually time-asymmetric implying closed-loop feedback control. A further type of behavior seen in a distinct subpopulation was a compound anomalous divergence response consisting of an initial convergence movement followed by a large corrective divergence movement with time courses implying closed-loop feedback control. This analysis of the variety of human vergence responses thus contributes substantially to the understanding of the oculomotor control mechanisms underlying the generation of vergence movements [corrected].