|
Neural Networks
Summary: Sebastian Seung studies neural networks using mathematical models, computer algorithms, and circuits of biological neurons in vitro. His interests include computational neuroanatomy, the idea of synaptic plasticity as an optimization algorithm, and persistent activity in neural integrators.
By itself, a single neuron is not intelligent. But a vast network of neurons can think, feel, remember, perceive, and generate the many remarkable phenomena that are collectively known as the mind. How does intelligence emerge from the interactions between neurons? This is the central question motivating the study of neural networks.
Computational Neuroanatomy
In the past few decades, there have been many technical advances in measuring the activity of neurons, such as multielectrode extracellular recording, whole-cell patch-clamp intracellular recording, and calcium imaging by two-photon microscopy. These measurement techniques have yielded a wealth of information about brain function.
We may now be witnessing the beginning of a similar revolution in techniques for measuring brain structure. New methods for visualizing neurons and the synaptic connections between them offer the prospect of transforming neuroanatomy into a high-throughput, data-rich science. To analyze unprecedented amounts of data about the nanoscale structure of brain circuitry, the computer will be an indispensable tool. My laboratory has begun to develop new algorithms for this purpose, as part of our venture into a new field that might be called "computational neuroanatomy," or simply "connectomics."
The goal of this field is to take a chunk of brain tissue and generate a "circuit diagram" containing information about the various types of neurons and how they are connected to each other. Progress in understanding the brain has been hampered by our lack of circuit diagrams. Imagine that you were trying to reverse engineer a radio by measuring electrical signals inside it. If someone gave you a circuit diagram, that would simplify the process considerably. Nanoscale resolution is necessary to obtain information about brain circuits, because axons and dendrites are very thin, and synapses are very small.
In particular, we are collaborating with Winfried Denk (Max Planck Institute, Heidelberg), who has invented a new technique called serial block-face scanning electron microscopy. This automated technique yields three-dimensional images of the brain at nanoscale resolution. Our role in this collaboration is to provide algorithms that take the raw images from his microscope and extract information about neuronal morphology and synaptic connectivity. This is an image-processing problem of unprecedented scale. From a cube of tissue that is 300 microns on a side, 1 teravoxel of data is generated. In that data are about 10,000 neurons, which must be identified and segmented as individual objects. There could be about 10,000,000 synapses, which also must be detected. To solve a problem of this scale with high reliability, we are employing a novel approach based on machine learning techniques. Our work on computational neuroanatomy represents a point of convergence for computer science and neuroscience.
By developing new algorithms, we hope to give neuroscientists new tools with which to map the synaptic organization of the brain, and thereby improve our understanding of the relationship between structure and function in neural networks.
Optimizing with Synapses
Many types of biological learning can be regarded as optimizations. For example, operant conditioning can be viewed as a process by which animals adapt their behaviors so as to maximize reward. The adage that "practice makes perfect" refers to the iterative improvement of complex motor skills like playing the piano or serving a tennis ball.
It is widely believed that long-lasting modifications of synaptic connections are responsible at least in part for the changes in behavior called learning. In my laboratory, we are interested in the hypothesis that one function of synaptic plasticity is to perform the computations required to optimize neural circuits.
We have proposed a number of hypothetical synaptic plasticity rules that are driven by the covariance of a global reward signal with various measures of neural activity that are local to the synapse. These rules effectively allow synapses to estimate the gradient of the expected reward, and thereby implement a procedure known in computer science as stochastic gradient learning. Through our modeling, we have shown how these rules can be applied to biophysically realistic, spiking neural network models. Furthermore, we have modeled a number of specific examples of biological learning as arising from our hypothetical synaptic plasticity rules.
Motivated by this theoretical work, we are also doing experiments on synaptic physiology. Our aim is to find synaptic plasticity mechanisms that could serve the computational function of optimization. In particular, we focus on the interaction between neuromodulators and local activity signals in inducing synaptic plasticity. One neuromoduluator of special interest is dopamine, which is thought to play the role of a reward signal in the brain.
Persistent Activity in Neural Integrators
It is an everyday experience to read a telephone number, dial it several seconds later, and then forget it. To execute such short-term memory tasks, the brain must store information during the time delay between sensation and action. Persistent neural activity related to short-term memory has been observed in many brain areas, but the mechanism by which this activity can outlast a transient sensory stimulus is still unknown. In the 1930s, Lorente de Nó hypothesized that such activity was maintained by reverberating circuits composed of synaptic feedback loops. More recently, others have suggested that the mechanism of short-term memory might reside not in the circuit, but in cellular biophysics.
We have been modeling the mechanisms of persistent neural activity. Our research focuses on neural integrators—brain areas that integrate over time in the sense of calculus. Such integrators are key elements of many computational models in neuroscience and cognitive science. For example, all types of eye movements involve a velocity-to-position neural integrator, which transforms angular velocity signals into changes in the angular position of the eyes. Burst inputs from saccadic command neurons are integrated to produce step changes in eye position, and velocity signals from the semicircular canals are integrated to produce changes in eye position appropriate for a vestibulo-ocular reflex that compensates for head movements. Another example of an integrator is the head direction system, which has been studied primarily in rodents. This system integrates angular velocity signals from vestibular and efference copy sources to produce changes in an internal representation of head direction.
In the absence of external input, an integrator maintains its state constant in time. Both the oculomotor and the head direction integrators exhibit persistent neural activity in the absence of velocity inputs. These examples of persistent neural activity are convenient to study experimentally, because they are operational during spontaneous behaviors. In contrast, persistent neural activity was originally studied in awake behaving primates that were extensively trained to perform short-term memory tasks. Furthermore, neural integrators are conceptually simple, because they contain an internal representation of a single scalar variable, such as eye position or head direction.
We have been investigating the hypothesis that neural integrators operate by virtue of tuned positive feedback mediated by recurrent synaptic connections. The hypothesis has been illustrated in a biophysically realistic way using a network of conductance-based model neurons, which exhibits persistent activity patterns that are similar to those seen in the oculomotor integrator. The persistent activity patterns form a continuous family of stable states of the network, or continuous attractor. The main weakness of the model is its sensitivity to mistuning of feedback. We are investigating biophysical mechanisms that might improve robustness to mistuning.
This work is also supported by the National Institutes of Health.
Last updated November 14, 2007
|
