The term "stochastic resonance" was first introduced in 1981 by Benzi as a mechanism by which random perturbations in Earth’s climate together with the eccentricity of its orbit cause the climate to switch from warm to cool phases in a 100,000 year cycle. While this explanation for the oscillations between ice ages and warm periods is still unproven, stochastic resonance has since been well-established in a huge variety of natural phenomena and nonlinear systems.
Learning algorithms based on reinforcement learning try to optimize reward. There exists a large literature modelling both dopaminergic reward-based learning and hippocampal memory formation using variants of reinforcement learning. Neurophysiologically inspired robotic cognitive control using reinforcement learning has been used to successfully handle environmental uncertainty. The DeepMind algorithm is directly inspired by hippocampal memory replay, which is thought to involve the sequential reactivation of hippocampal place cells that represent previously experienced behavioral trajectories. Interestingly, there are several models of dopamine function and hippocampal CA1 neurons that find evidence of stochastic resonance improving signal detection. Computational modelling has shown that dopaminergic neurons and of hippocampal neurons likely benefit from the presence of noise, either endogenously or from external sources.
The brain is still far better at many learning tasks compared to computers. If we can understand how the brain exploits stochastic resonance we may be able to improve machine learning. And conversely if we can use stochastic resonance to improve artificial learning algorithms we will be in a better theoretical position to solve how the brain uses noise. A convergence in these two lines of research, noise benefits and deep or reinforcement learning, could yield advances in both artificial intelligence and neuroscience. In fact, the DeepMind reinforcement learning algorithm introduces a novel feature that randomizes over the data, thereby removing correlations in the observation sequence and smoothing over changes in the data distribution. This kind of additive noise benefit might be further exploited by explicitly using the well-established mathematical concepts of stochastic resonance together with reinforcement learning. Future research should focus on synthesizing these lines of research to develop novel algorithms that are robust to uncertainty and take advantage of nature’s most plentiful untapped resource: noise.
- More Here
Learning algorithms based on reinforcement learning try to optimize reward. There exists a large literature modelling both dopaminergic reward-based learning and hippocampal memory formation using variants of reinforcement learning. Neurophysiologically inspired robotic cognitive control using reinforcement learning has been used to successfully handle environmental uncertainty. The DeepMind algorithm is directly inspired by hippocampal memory replay, which is thought to involve the sequential reactivation of hippocampal place cells that represent previously experienced behavioral trajectories. Interestingly, there are several models of dopamine function and hippocampal CA1 neurons that find evidence of stochastic resonance improving signal detection. Computational modelling has shown that dopaminergic neurons and of hippocampal neurons likely benefit from the presence of noise, either endogenously or from external sources.
The brain is still far better at many learning tasks compared to computers. If we can understand how the brain exploits stochastic resonance we may be able to improve machine learning. And conversely if we can use stochastic resonance to improve artificial learning algorithms we will be in a better theoretical position to solve how the brain uses noise. A convergence in these two lines of research, noise benefits and deep or reinforcement learning, could yield advances in both artificial intelligence and neuroscience. In fact, the DeepMind reinforcement learning algorithm introduces a novel feature that randomizes over the data, thereby removing correlations in the observation sequence and smoothing over changes in the data distribution. This kind of additive noise benefit might be further exploited by explicitly using the well-established mathematical concepts of stochastic resonance together with reinforcement learning. Future research should focus on synthesizing these lines of research to develop novel algorithms that are robust to uncertainty and take advantage of nature’s most plentiful untapped resource: noise.
- More Here
No comments:
Post a Comment