Weight Perturbation: An Optimal Architecture and Learning Technique for Analog VLSI Feedforward and Recurrent Multilayer Networks

Jabri, Marwan A.; Flower, B.

doi:10.1162/neco.1991.3.4.546

Cited by 58 publications

(52 citation statements)

References 3 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In weight perturbation (Jabri and Flower, 1991;Alspector et al, 1993;Flower and Jabri, 1993;Kirk et al, 1993;Cauwenberghs, 1993) the gradient ' is approximated using only the globally broadcast result of the computation of E(© ). This is done by adding a random zero-mean perturbation vector ∆© to © repeatedly and approxmating the resulting change in error by…”

Section: Weight Perturbationmentioning

confidence: 99%

Fast Exact Multiplication by the Hessian

1994

View full text Add to dashboard Cite

Just storing the Hessian (the matrix of second derivatives ∂ 2 E¡ ∂w i ∂w j of the error E with respect to each pair of weights) of a large neural network is difficult. Since a common use of a large matrix like is to compute its product with various vectors, we derive a technique that directly calculates £ ¢, which takes about as much computation, and is about as local, as a gradient evaluation. We then apply the technique to a one pass gradient calculation algorithm (backpropagation), a relaxation gradient calculation algorithm (recurrent backpropagation), and two stochastic gradient calculation algorithms (Boltzmann Machines and weight perturbation). Finally, we show that this technique can be used at the heart of many iterative techniques for computing various properties of , obviating any need to calculate the full Hessian.

show abstract

Section: Weight Perturbationmentioning

confidence: 99%

Fast Exact Multiplication by the Hessian

1994

View full text Add to dashboard Cite

show abstract

“…Spike timing dependent plasticity [77] is a recent learning scheme that exploits the precise timing between pre and post synaptic firing events in the learning rule. Numerous adaptations of these major learning rules have been proposed, as well as hardware-friendly rules, such as weight perturbation [78].…”

Section: Learningmentioning

confidence: 99%

Neuromorphic microelectronics from devices to hardware systems and applications

Schmid

2016

NOLTA

View full text Add to dashboard Cite

Neuromorphic systems aiming at mimicking some characteristics of the nervous systems of living humans or animals have been developed since the late 1980s', taking benefit of intrinsic properties and increasing performances of the successive silicon fabrication technologies. A regain of interest has been observed in the middle of the 2010s', which manifests itself from the emergence of large-scale projects integrating various computational and hardware perspectives, by the increased interest and involvement of industry and the growth of the volume of scientific publications. This paper reviews research directions and methods of neuromorphic microelectronics hardware, the developed hardware and its performance, and discusses current issues and potential future developments.

show abstract

“…The stochastic gradient descent algorithm, also called the weight perturbation algorithm (Jabri and Flower, 1992), is a simple method for descending the gradient of a noisy objective function. The algorithm proceeds as follows.…”

Section: Risk-sensitive Stochastic Gradient Descentmentioning

confidence: 99%

Variable risk control via stochastic optimization

Kuindersma

Grupen

Barto

2013

The International Journal of Robotics Research

View full text Add to dashboard Cite

We present new global and local policy search algorithms suitable for problems with policy-dependent cost variance (or risk ), a property present in many robot control tasks. These algorithms exploit new techniques in nonparameteric heteroscedastic regression to directly model the policy-dependent distribution of cost. For local search, the learned cost model can be used as a critic for performing risk-sensitive gradient descent. Alternatively, decision-theoretic criteria can be applied to globally select policies to balance exploration and exploitation in a principled way, or to perform greedy minimization with respect to various risk-sensitive criteria. This separation of learning and policy selection permits variable risk control, where risk sensitivity can be flexibly adjusted and appropriate policies can be selected at runtime without relearning. We describe experiments in dynamic stabilization and manipulation with a mobile manipulator that demonstrate learning of flexible, risk-sensitive policies in very few trials.

show abstract

Weight Perturbation: An Optimal Architecture and Learning Technique for Analog VLSI Feedforward and Recurrent Multilayer Networks

Cited by 58 publications

References 3 publications

Fast Exact Multiplication by the Hessian

Fast Exact Multiplication by the Hessian

Neuromorphic microelectronics from devices to hardware systems and applications

Variable risk control via stochastic optimization

Contact Info

Product

Resources

About