Daphne Bavelier
Richard Aslin
Daniel Kersten
Steven Hillyard
Wayne Gray
Josh Tenenbaum
Alexander Pouget
Recent evidence suggests that the framework of Bayesian decision making may provide a good first-order model of how subjects learn to make optimal decisions in dynamic complex tasks. Optimal in this context is defined in terms of attaining the task goal, while minimizing loss and maximizing gains. Although this research is suggestive, it is critical to test this hypothesis under conditions of increasing, realistic complexity if we are to understand the conditions under which optimal learning of complex dynamic tasks can be induced.
Our previous research shows that the simple act of playing action video games for a mere 10 hours can enhance the ability of young adults to search their visual environment for a pre-specified target, to monitor moving objects in a complex visual scene, or to process a fast-paced stream of visual information. Thus, limits on attentional capacity previously hypothesized to be fixed can be modified by appropriate training. We will extend this research and document in detail which aspects of sensory and motor skills (see Kersten), attention, pattern learning and recognition (see Aslin), working memory, and decision-making processes (see Pouget) are modified by gaming. We will also characterize associated neural changes using human brain imaging (fMRI see RCBI, ERPs see Hillyard and Optical Imaging see Aslin).
Professor Aslin, in collaboration with Professor Newport, will be conducting research on the basic mechanisms of sequence learning. Gathering information from sequences of events is of critical importance for optimal performance in several domains, such as listening to instructions and making decisions under extreme time pressure, or monitoring visual events and making motor responses. A key aspect of sequence learning is transfer: how do learners acquire both specific and general information so that they can appropriately generalize beyond the training set? The Aslin/Newport lab will conduct studies of sequence learning using novel miniature languages (in the auditory modality) and using visuomotor reaction-time tasks on a touch-screen to determine the training needed for optimal learning and for robust transfer to novel sequences. Understanding generalization—when it occurs and what cues learners use to determine the optical balance between generalizing and retaining specificity—is a focus of this research.
Our aim is to understand the factors that promote learning, in collaboration with Paul Schrater and Yuhong Jiang. We will develop a small number of prototype video games allowing us to incrementally add task elements in ways that are amenable to laboratory study as well as normative computational analysis of learning. We are prototyping several games using Virtools - a high-end game development package. In one example, "the jet-fighter game" , players fly through a simulated 3D environment and shoot or do not shoot approaching targets that may be "friend" or "foe". The environment can be manipulated to be "rich" (e.g., complex backgrounds, with a large immersive 3D view) or "poor" (simple backgrounds with a flattened abstract view). The characteristic features or cues that determine a foe, and distinguish it from the background, or an ally, are inherently statistical, and we can simulate degrees of uncertainty and risk. So for example, the validity of the visual cues for distinguishing friend from foe can be probabilistically manipulated; thus, we can embed the so-called "multi-armed bandit" problem in the context of a video game requiring visual-motor skill.
We will test elements that foster learning in the context of various versions of the game. Among such elements will be:
Demo Links:
http://www.socsci.umn.edu/~pete/Virtools/jet_game.html
http://www.socsci.umn.edu/~pete/Virtools/driving_game.html
The High Frequency SSVEP is Enhanced by Spatial Attention
We plan to investigate the neural bases of improved performance in video gamers by recording event-related brain potentials (ERPs) from video game players and comparing their brain activity with that of non-players. ERPs represent the summated electric field potentials generated by the neural networks of the cerebral cortex as they process information in the brain. These potentials can be recorded non-invasively from the scalp and provide a detailed picture of the spatiotemporal patterns of brain activity that underlie sensory processing, perception and cognition. In the proposed research we will use ERPs to elucidate the cortical mechanisms by which video gaming enhances attention and perception in a broad range of tasks. In particular, we will investigate whether video game players:
Our modeling framework aims to capture several key features of complex human learning that have not previously been present in a single approach, and that spans cognitive and neural levels of analysis. Our model learner should be able to learn flexible symbolic knowledge and strategies, and to use structured prior knowledge to constrain and accelerate learning from data. Yet our model should also be sensitive to the fine-grained statistical structure of specific tasks and environments. It should be able to adaptively control its own information gathering and processing, dynamically allocating limited attention and working memory resources to different information sources and tasks in order to improve total performance.
It should be able to make adaptive decisions, trading off uncertainty, risk, and potential for reward. It should be able to learn through rational and analytically transparent means, so it can be useful as a tool for understanding why some learning tasks are harder or more natural than others, characterizing how training should be decomposed to promote learning when a new set of tasks are considered, and discovering new ways of training those cognitive processes that have so far been hard to modify. Such a model should also provide an upper bound on the type of performance enhancement we can expect from a human learner.
In collaboration with Marc Destefano, we will ask how action video game players construct deep symbolic representations of a task, especially representations that can be transferred into other domains (see Tenenbaum). We will be using advances in Cognitive Task Analysis (CTA) to predict when a game player will choose one particular low-level interactive routine (spanning from about 300 to 3000 milliseconds) over another. The task of choice for this research is Revised Space Fortress – a fast-paced, real-time interactive video game with intensive multitasking demands typical of action games.
Space Fortress is tractable in terms of task elements, but at the same time it contains many of the elements of the first-person action video game known to enhance perceptual and cognitive skills. We will build a cognitive model that is able to play the game at an expert level, with the ultimate aim of building a cognitive model that can learn the game through reinforcement and repetition alone, with no prior strategic knowledge. This will yield new insights as to how implicit knowledge can become consciously accessible.
We will develop models that integrate optimal principles of statistical learning, inference, planning and decision-making under uncertainty, with hierarchically structured and symbolic forms of representations. Our basic framework for modeling learning in complex tasks will be based on POMDPs (Partially Observable Markov Decision Problems), a powerful language for formulating sequential decision problems in operations research and artificial intelligence. A POMDP describes a task environment in terms of a set of states, a set of possible actions, and three probabilistic functions defined on state-action pairs:
POMDPs have various advantages as a computational formalism:
We will develop a neural theory of learning which is consistent with the POMDPs framework (see Tenenbaum). By doing so, we will provide a viable approach for modeling human learning in complex cognitive tasks not only at the behavioral but also at the neural level. Our starting point is our past work on decision making which indicates that networks of spiking neurons are implementing POMDPs. We will develop this framework further and use it to explore the neural basis of learning in complex tasks.
There have been numerous attempts at modeling the neural basis of decision making. Many of these models are based on the framework of sequential decision making and build from a strong tradition in mathematical psychology where classical models of decision making such as diffusion to bound, leaky accumulator, and random walk models were first developed. These models, and their neural implementations, are all limited to two choices, and are optimal only in situations in which the statistics of the environment do not change over time. These are severe limitations. For instance, deciding where a predator is hiding is not a binary task, and the quality of the image on which this task is based can greatly vary since it is much harder to spot a sniper in a cluttered urban environment than in a desert oasis with only two palm trees. Yet, none of the existing models tell us how the nervous system deals optimally with such situations. The POMDP framework is ideal for this purpose and will allow us to develop theories of optimal decision making for any number of discrete choices, or decisions over continuous variables, even in situations in which the quality of the data changes over time.