Learning and control of next generation deep neural technologies
Abstract
In recent years, methods from machine learning have make great advances to solve difficult problems in artificialintelligence. In p""articular, (deep) feed-forward neural networks have reestablished themselves as one of the mostpowerful learning architectures. Thi""s has led to spectacular applications in the area of computer vision, speechrecognition and game playing. The current successes are"" pushing the technology to larger and larger applications,pushing the limits of computation and energy consumption.The learning a""lgorithms for deep networks have much improved over the last years along various directions. However, good performance is still hard"" to achieve because of either vanishing gradients in deterministic networks, or unreliable gradients due to sampling difficulty and"" large mixing times in stochastic networks. For recurrent neural networks, there are no good algorithms for learning general (no det""ailed balance) stochastic networks. Furthermore, there is no good algorithms to learn with low precision synapses.In this proposal"", we aim to contribute to advance algorithms and theoretical understanding of learning in artificial neuralnetworks by formulating" the neural networks as Markov processes and treat the learning as a stochastic optimalcontrol problem. This approach also provides a theoretical framework for learning low precision synapses. We analyzethe properties of these networks using replica analysis. We aim to design novel learning methods for deep andrecurrent stochastic networks.The control approach to learning can be seen as a finite temperature generalization of maximum likelihood learningthat reduces to maximum likelihood method at zero temperature. Lea"rning in this framework can be shown to be a form of importance sampling. Sampling efficiency, and therefore learning efficiency can"" be significantly accelerated by introducing a trust region approach, resulting in an optimized annealing schedule.The replica app"roach has led to a novel learning method using physical replicas of the neural network. The methodhas been shown to learn large per"ceptrons in cases that traditional energy based methods fail.In this project, we integrate these research lines with the objective" to develop novel neural network learning rules fordeep or recurrent stochastic neural networks.The learning rules from this project may also provide the algorithms for on-chip learning on future neuro-morphichardware devices. An important limitation for the g"rowth of future generations of computers, and thus of neuralnetworks, is the excessive amount of heat that they produce. Devices th""at allow for on-chip learning and that replace double precision synapses with unreliable stochastic bits, can potentially save many" orders of magnitude in energy consumption and heat dissipation and thus scale-up neural network computation. The stochastic approach to learning proposed in this project naturally yields learning rules that are more local than traditional back-propagation based l"earning, and are promising candidates for on chip learning.This basic research effort is directly relevant to ONR Autonomy and Inf"ormation Dominance focus areas.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 09, 2017
- Source ID
- N000141712569
Entities
People
- Bert Kappen
Organizations
- Office of Naval Research
- Stichting Katholieke Universiteit
- United States Navy