learning
をテンプレートにして作成
[
トップ
] [
新規
|
一覧
|
単語検索
|
最終更新
|
ヘルプ
|
ログイン
]
開始行:
[[Neuromorphic]]
COLOR(red){Q.Where and when should training/learning for a neuromorphic system take place?}
A major technical and detailed consideration, especially for trained and learning neuromorphic computers, is where training/learning takes place and when training/learning takes place.
We define the location of learning as either “on-chip” or “off-chip.” We use these terms for convenience even though the actual device or computer acting as a neuromorphic computer may not be a chip.
On-chip training/learning is when the training/learning algorithm is implemented as part of the operation of the neuromorphic computer. Many existing neuromorphic computers do not include any notion of on-chip
training or learning.
Off-chip training/learning is when the training/learning algorithm does not take place on the neuromorphic computer (though the neuromorphic computer may take part of the algorithm or a simulation may be used in its place). We define the time of learning as either “on-line” or “off-line.”
On-line training/learning occurs when the algorithm makes real-time updates in a real situation. Usually on-line refers to a learning instance, which the neuromorphic chip is making decisions on how to update itself dynamically without feedback. However, training algorithms may also fit into this paradigm as part of a framework such as reinforcement learning, where the user may be providing minimal feedback.
Offline training/learning usually occurs when the training knowledge is available earlier than when the device is used in the environment. One possibility for implementing hybrid on-line/off-line learning system is a system that can operate in real-time using an off-line trained model but store sampled input statistics, network outputs, and implications and update itself off-line at a later point.
This approach is arguably what the cortex/hippo-campus interaction provides to brains; our cortical “models” update very slowly (potentially even off-line during sleep), whereas the hippocampus is continuously learning and storing what occurs in the world and eventually using this info to update the cortical model. Neuromorphic devices can contain combinations of on-line, off-line, on-chip, and off-chip training/learning.
On-line training/learning methods are usually by necessity on-chip. Off-line training/learning mechanisms may be on-chip or off-chip. Even neuromorphic computers that rely heavily on on-line learning will almost certainly include off-line pre-training and/or programming by the user in which some structure and parameters are defined and then refined as part of the on-line learning process[[[ref>http://ornlcda.github.io/neuromorphic2016/files/ORNLNeuromorphicComputingWorkshop2016Report.pdf]]].
終了行:
[[Neuromorphic]]
COLOR(red){Q.Where and when should training/learning for a neuromorphic system take place?}
A major technical and detailed consideration, especially for trained and learning neuromorphic computers, is where training/learning takes place and when training/learning takes place.
We define the location of learning as either “on-chip” or “off-chip.” We use these terms for convenience even though the actual device or computer acting as a neuromorphic computer may not be a chip.
On-chip training/learning is when the training/learning algorithm is implemented as part of the operation of the neuromorphic computer. Many existing neuromorphic computers do not include any notion of on-chip
training or learning.
Off-chip training/learning is when the training/learning algorithm does not take place on the neuromorphic computer (though the neuromorphic computer may take part of the algorithm or a simulation may be used in its place). We define the time of learning as either “on-line” or “off-line.”
On-line training/learning occurs when the algorithm makes real-time updates in a real situation. Usually on-line refers to a learning instance, which the neuromorphic chip is making decisions on how to update itself dynamically without feedback. However, training algorithms may also fit into this paradigm as part of a framework such as reinforcement learning, where the user may be providing minimal feedback.
Offline training/learning usually occurs when the training knowledge is available earlier than when the device is used in the environment. One possibility for implementing hybrid on-line/off-line learning system is a system that can operate in real-time using an off-line trained model but store sampled input statistics, network outputs, and implications and update itself off-line at a later point.
This approach is arguably what the cortex/hippo-campus interaction provides to brains; our cortical “models” update very slowly (potentially even off-line during sleep), whereas the hippocampus is continuously learning and storing what occurs in the world and eventually using this info to update the cortical model. Neuromorphic devices can contain combinations of on-line, off-line, on-chip, and off-chip training/learning.
On-line training/learning methods are usually by necessity on-chip. Off-line training/learning mechanisms may be on-chip or off-chip. Even neuromorphic computers that rely heavily on on-line learning will almost certainly include off-line pre-training and/or programming by the user in which some structure and parameters are defined and then refined as part of the on-line learning process[[[ref>http://ornlcda.github.io/neuromorphic2016/files/ORNLNeuromorphicComputingWorkshop2016Report.pdf]]].
ページ名: