Meta-Learning via Learned Loss

Bechtle, Sarah; Molchanov, Artem; Chebotar, Yevgen; Grefenstette, Edward; Righetti, Ludovic; Sukhatme, Gaurav S.; Meier, Franziska

doi:10.48550/arxiv.1906.05374

Cited by 14 publications

(40 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In this regard, a recent research direction is concerned with loss function meta-learning, with diverse applications in supervised and reinforcement learning [18][19][20][21][22][23][24][25][26][27][28][29][30][31]. Although different works utilize different meta-learning techniques and have different goals, it has been shown that loss functions obtained via meta-learning can lead to an improved convergence of the gradient-descent-based optimization.…”

Section: Related Work and Motivationmentioning

confidence: 99%

“…• We extend the loss function meta-learning technique of [27] by considering alternative loss parametrizations and various algorithm design options.…”

Section: Summary Of Innovative Claimsmentioning

confidence: 99%

“…Indicatively, can be represented with a feed-forward NN (FFN) and thus η represents the weights and biases of the FFN. Such a parametrization in conjunction with Algorithm 1 has been proposed in [27] for supervised learning and reinforcement learning problems. Alternatively, can be the adaptive loss function proposed in [12] (see Eqs.…”

Section: An Algorithm For Meta-learning Pinn Loss Functionsmentioning

confidence: 99%

“…3.4.1.2 NN-parametrized loss function. An alternative parametrization based on FFN has been proposed in [27]. Specifically, for the one-dimensional supervised learning regression problem considered in [27], the most expressive representation η = ˆ η (û θ , u) of Fig.…”

Section: Adaptive Loss Functionmentioning

confidence: 99%

“…In this regard, we note that ensuring positivity of the loss function does not affect NN parameter optimization; i.e., the softplus output activation function affects the results of [27] only through its nonlinearity and not by dictating positive loss outputs. Furthermore, instead of randomly initializing NN parameters η, one can alternatively initialize them so that η approximates a known loss function such as the squared 2 -norm.…”

Section: Adaptive Loss Functionmentioning

confidence: 99%

See 4 more Smart Citations

Meta-learning PINN loss functions

Psaros,

Kawaguchi,

Karniadakis

2021

Preprint

View full text Add to dashboard Cite

We propose a meta-learning technique for offline discovery of physics-informed neural network (PINN) loss functions. We extend earlier works on meta-learning, and develop a gradient-based meta-learning algorithm for addressing diverse task distributions based on parametrized partial differential equations (PDEs) that are solved with PINNs. Furthermore, based on new theory we identify two desirable properties of meta-learned losses in PINN problems, which we enforce by proposing a new regularization method or using a specific parametrization of the loss function. In the computational examples, the meta-learned losses are employed at test time for addressing regression and PDE task distributions. Our results indicate that significant performance improvement can be achieved by using a shared-among-tasks offline-learned loss function even for out-of-distribution meta-testing. In this case, we solve for test tasks that do not belong to the task distribution used in meta-training, and we also employ PINN architectures that are different from the PINN architecture used in meta-training. To better understand the capabilities and limitations of the proposed method, we consider various parametrizations of the loss function and describe different algorithm design options and how they may affect meta-learning performance.

show abstract

Section: Related Work and Motivationmentioning

confidence: 99%

“…• We extend the loss function meta-learning technique of [27] by considering alternative loss parametrizations and various algorithm design options.…”

Section: Summary Of Innovative Claimsmentioning

confidence: 99%

Section: An Algorithm For Meta-learning Pinn Loss Functionsmentioning

confidence: 99%

Section: Adaptive Loss Functionmentioning

confidence: 99%

Section: Adaptive Loss Functionmentioning

confidence: 99%

See 3 more Smart Citations

Meta-learning PINN loss functions

Psaros,

Kawaguchi,

Karniadakis

2021

Preprint

View full text Add to dashboard Cite

show abstract

Meta-learning based passenger flow prediction for newly-operated stations

Han,

Zhang,

Tian

et al. 2023

Geoinformatica

View full text Add to dashboard Cite

Task-Agnostic Exploration via Policy Gradient of a Non-Parametric State Entropy Estimate

Mutti

Pratissoli

Restelli

2021

AAAI

View full text Add to dashboard Cite

In a reward-free environment, what is a suitable intrinsic objective for an agent to pursue so that it can learn an optimal task-agnostic exploration policy? In this paper, we argue that the entropy of the state distribution induced by finite-horizon trajectories is a sensible target. Especially, we present a novel and practical policy-search algorithm, Maximum Entropy POLicy optimization (MEPOL), to learn a policy that maximizes a non-parametric, $k$-nearest neighbors estimate of the state distribution entropy. In contrast to known methods, MEPOL is completely model-free as it requires neither to estimate the state distribution of any policy nor to model transition dynamics. Then, we empirically show that MEPOL allows learning a maximum-entropy exploration policy in high-dimensional, continuous-control domains, and how this policy facilitates learning meaningful reward-based tasks downstream.

show abstract

Meta-Learning via Learned Loss

Cited by 14 publications

References 0 publications

Meta-learning PINN loss functions

Meta-learning PINN loss functions

Meta-learning based passenger flow prediction for newly-operated stations

Task-Agnostic Exploration via Policy Gradient of a Non-Parametric State Entropy Estimate

Contact Info

Product

Resources

About