21st Century Science Initiative Grant: Understanding Human Cognition

Significance and Potential Impact to Neurosciences Research: High-speed simulations of brain injury will allow prediction of disease progression and the future effects of treatments, and detailed understanding of the interaction between brain injury and resulting disorders of movement, perception, and behavior. In addition to disorders such as stroke, traumatic brain injury, and neuro-degenerative diseases, the greatest potential impact will be for those studying developmental brain disorders due to cerebral palsy (CP), with over 150,000 children and 350,000 adults affected in the United States. CP is particularly relevant because it is caused by early brain injury, the disability persists and may worsen years after the injury, and the cause of the late worsening is unknown. If disease progression is predictable, treatment interventions can occur much earlier and include customization to each patient’s predicted disease course. With early intervention, it may be possible to attenuate or block the natural progression of their disease. With over 4.5% of the child and adult US population affected by developmental brain disorders, acquired brain injury, and progressive brain disease, the applicability and potential impact of an early intervention and disease prediction technique is significant.

Knowledge Gap—Repairing the Software: The brain is an organ of information processing (software) whose behavior depends on both its cellular structure (hardware) and the information stored (memory) in that structure (figure 1). When the brain is injured during development, there is an effect not only on immediate brain function but also on the future ability to acquire and store information. While much research has focused on repair or prevention of injury to brain structures, we have almost no understanding of how structural injury affects current and future information processing and storage and how degradation of information processing and storage leads to ongoing deficits of neurological function. If such an understanding could be achieved, it would permit a completely new avenue for the treatment of neurological disease by preventing or reversing the behavioral expression of cellular injury. For example, consider a child with injury to central somatosensory systems due to a prenatal stroke near primary sensory cortex. All motor skill learning for that child will take place in an environment of reduced sensation, which can reduce detailed knowledge of the state of the body or the results of movement. Therefore, this child may learn abnormal patterns of movement and may not develop age-appropriate skills [1]. However, if we can understand the relationship between the original injury and motor learning, we may be able to manipulate the environment using virtual reality, biofeedback, or specific training tasks in order to store information that results in improved motor function. The fundamental paradigm shift that we seek requires recognition of a tremendous gap in our knowledge: we need a quantitative understanding of the relationship between injury and future brain function, where brain function depends on both the cellular consequences of injury as well as the informational consequences of injury. There is currently no theory, nor even the recognition of the lack of a theory, that links injury to specific neuronal types or brain regions to failed acquisition of information. We propose a broad application of computational models of learning and information storage to the prediction and treatment of the time course of neurological disease.

Approach: We do not believe that quantitative predictions of normal brain function from normal brain structure will be possible in the foreseeable future. However, our task is much simpler. We need only to predict the change in brain function that results from a change in brain structure due to injury, or a change due to treatment of injury. This is simpler because we only need to model the few dimensions of behavior that are of direct relevance to the disease process. For Figure 1: Simple schematic of the dependence of information processing on the integrity of brain structure. In this view, the structure is the foundation upon which information is processed and stored in the brain. Instead of focusing only on the mechanisms of injury to brain structure, we propose to study the effect of structural injury on information processing and storage. example, a two-neuron model of the monosynaptic stretch reflex may be sufficient to predict some aspects of spasticity, even though many other spinal and supraspinal cells may play an important role in normal behavior [2]. A binocular receptive field model of cells in primary visual cortex may be sufficient to predict some aspects of amblyopia, even though other cortical visual areas may play an important role and may have reciprocal connections with primary visual cortex.

The use of much simpler models of structural injury in neurological disease has a long history. One example is the interpretation of the multiple waves of the brainstem auditory evoked response (BAER). Individual waves are attributed to the cochlea, auditory nerve, cranial nerve nucleus, and other elements in the pathway from sound to auditory cortex. In this case, the model is used to relate changes in the observed electrophysiological phenomena (the waves of the BAER) to structural changes at the presumed anatomic location of injury, even though the precise mechanisms that generate the BAER are likely to be considerably more complex. Similarly, we propose to use simulations of computational models to relate changes in observed electrophysiological and behavioral phenomena to changes in stored information and future information processing.

FPGA’s and Likelihood Calculus: To be clinically useful, we require several properties for a model of the time-course of neurological disease: (1) Extremely high speed, (2) large scale, and (3) physiological predictions in terms of measurable physiological variables such as the electroencephalogram, the electromyogram, extracellular electrode recordings, motion kinematics, perceptual deficits, and perceptual threshold. Existing large-scale simulations run considerably slower than real-time or are unable to model the spike behavior of large collections of individual neurons [3, 4]. Therefore, we will use a fundamentally new approach combining two technologies: Field-Programmable Gate Arrays (FPGA’s) and the mathematical theory of Likelihood Calculus. FPGA’s are a low-cost and extremely fast hardware base on which parallel simulations are implemented. Likelihood Calculus is what allows the efficient parallel simulation of very large collections of neurons on the FPGA. Significantly, the information processing can be understood using probability theory and information theory. This differs from attempts at biologically-realistic multi-neuron simulations that provide detailed observational data but which do not attempt to explain function in terms of well-understood mathematics or information theory [3]. Our technical approach solves a fundamental problem of connectivity. To simulate one million neurons with complete connectivity requires at least 1 trillion connections, and storage and recall of the activity at each synapse is not possible in real time, even for computer clusters or massively parallel computers. Likelihood Calculus permits such a network to be approximated using a set of much smaller parallel sub-networks. This structure mimics the hypercolumn structure of cerebral cortex or the repeating parallel subunit structure of cerebellar cortex. Its great advantage is the ease of implementation using inexpensive FPGA hardware. Thus likelihood calculus is the fundamental enabling technology that allows the extremely high-speed simulation of large networks necessary to test the model.