Last summer (2018), I was fortunate enough to be invited to attend the Lindau Nobel Laureate Meeting. It was an incredible experience. While there, I was interviewed by Michael Kisselgof, cofounder of the biotech company IKU. We spoke about a range of topics, including deep neural networks and science communication. Take a listen below.
When children learn to read, they are taught to “sound out” words, or read aloud letter-by-letter. In the 1980s, Terrence Sejnowski and Charles Rosenberg sought to model this process by building a neural network called NETtalk that learned to convert written text to speech1. In other words, they taught a computer2 to read. The authors concluded that NETtalk was too simple to serve as a complete model for human learning, but it does have some important implications for connectomics. Connectomics is the study of wiring-diagrams of the nervous system, or the patterns of connections between neurons. The hope is that by mapping the connectome, we will gain fundamental insights into the function, and dysfunction, of the brain in health and disease. A quick summary of the inner-workings of NETtalk is necessary before we can understand the implications for connectomics.
NETtalk is a simple neural network that consists of 3 layers of nodes: an input layer, a hidden layer, and an output layer (Figure 1). To train the network, written text is fed into the input layer, is then propagated through the hidden layer, and is finally mapped onto a phoneme in the output layer. The letter-to-phoneme correspondence is determined by the connections between nodes, and the network learns the correct correspondence by readjusting the weights of connections between layers. A teacher unit provides feedback, and the network uses this feedback through many iterations of training to adjust its connections.
This is a column I wrote for the Georgetown Interdisciplinary Program in Neuroscience newsletter about computational psychiatry. My aim was to write a broad overview of the field for a lay audience, so it should be pretty accessible. Here is a link to the newsletter.
Fifty percent of Americans will develop at least one form of mental illness in their lifetime1, a staggering statistic. Despite such high disease burden, our understanding of psychiatric disorders significantly lags behind that of other medical conditions. Psychiatric investigation is difficult: the brain is an incredibly complex system embedded in an unpredictable environment. Equipped with rapidly advancing technologies, basic neuroscience has recently begun to overcome this complexity to reveal fundamental
mechanisms of brain function. A bottleneck exists, however, in translating discoveries in the lab into meaningful clinical improvements for patients. The nascent field of computational psychiatry, the confluence of computational, statistical, and clinical investigation, may change that. Computational psychiatry crosses disciplinary borders to imagine a future in which clinicians, aided by powerful algorithms, are able to use a picture of a patient’s brain and its connections to better detect and treat psychiatric disease.
One reason for the paucity of success stories in psychiatry is that diagnosing mental illness is difficult. Physicians are trained to integrate observations from patint interviews, physical exams, and laboratory tests to make a diagnosis. Medical students are taught that some clinical signs are pathognomonic, or characteristic for a particular disease. In psychiatry, diagnosis is far more subtle and complex. First, laboratory tests and physical exams, indispensible tools for clinicians, are less useful in the evaluation of mental illness. Instead, diagnostic classification is primarily symptom-based. The Diagnostic and Statistical Manual of Mental Disorders (DSM), often referred to as the “Bible” of the field, is used to diagnose patients on the basis of clusters of symptoms, while remaining agnostic to the underlying pathophysiology. Very few symptoms in psychiatry are pathognomonic, however; most symptoms are characteristic of multiple disorders. Individuals often suffer from comorbidity, or the co-occurrence of several diseases, further complicating diagnostic classification. Clinicians need new tools to deal with this complexity.
Machine learning (ML), a subfield of artificial intelligence, is a system of algorithms that learns patterns in data. You encounter these algorithms every day: Siri on your iPhone and Facebook’s image recognition software are both powered by ML. Recently, researchers have begun applying ML algorithms to neuroimaging data of psychiatric patients with the goal of discovering potential biomarkers for disease diagnosis and progression. Here’s how it works: we begin with a cohort of individuals with schizophrenia, and a group of healthy controls. We collect data using functional magnetic resonance imaging (fMRI), a technique that measures blood flow in the brain as a proxy for neuronal activity. Using a statistical technique known as independent component analysis, we can characterize the functional activity of various networks of brain regions, each of which carries clinically useful information. These data are far too large and complex for psychiatrists to parse, however. Rather, we feed these datasets, or examples, into a type of ML algorithm know as a classifier, which scours millions of data points to learn the relevant characteristics, or features, that predict the class of each example. In other words, the classifier is a data-driven technique for classifying individuals with and without schizophrenia on the basis of pictures of their brain.
What follows is a review I wrote with Ajay Pillai and Sule Tinaz as a post-baccalaureate fellow in Mark Hallett’s lab at NIH, which we submitted as a journal club review to The Journal of Neuroscience. It wasn’t accepted, so I thought I would post it here. In this review, we summarize the findings of a 2013 J neuro paper by Baraduc and colleagues, and relate their study to recent developments in our understanding of bradykinesia, or slowness of movement, one of the cardinal symptoms of Parkinson’s disease. Further, we extend the authors’ discussion by relating their model and findings to another feature of bradykinesia known as the “sequence effect.” Finally, we discuss another model of motor control, Dynamical Systems Theory, and its potential use in elucidating the pathophysiology of the sequence effect and bradykinesia.
Why does “the battery run down”? Computational Modeling in Parkinson’s Disease: Towards an Understanding of the Pathophysiology of Bradykinesia
Patrick Malone, Ajay Pillai, and Sule Tinaz
How organisms coordinate multiple end effectors, muscles, and neural activity to produce purposeful motor behavior has been the central question of motor control theories. The Optimal Control Theory (OCT) provides a computational framework to understand motor control. The theory posits that organisms learn to generate goal-directed movements that optimize cost (e.g., speed, accuracy). The models employ a two-factor cost function that encodes both the feature to be optimized (e.g., speed) and another factor that needs to be regularized (e.g., motor command representing the neural input to the system). More recent versions of OCT-based approaches also incorporate feedback into the model in which the current state of the system is informed by its previous states (Diedrichsen et al., 2010).
Research in this area has focused on computational characterization of visually guided reaching movements (Shadmehr and Krakauer, 2008). It has long been known that movement speed inversely scales with the accuracy requirements of the task (Fitts, 1954). Bradykinesia, the cardinal motor feature of Parkinson’s disease (PD), is characterized by slowness of movement caused by problems in scaling speed to movement distance (Hallett and Khoshbin, 1980). Scaling deficits could be due to a higher accuracy cost for these patients, however this view has been challenged in recent years. In response, the energetic cost has been introduced as another variable in the cost/benefit models and is thought to be a major determinant of movement speed (Mazzoni et al., 2007).