Learning to Optimize
Authors: Ke Li, Jitendra Malik
ICLR 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We learn optimization algorithms for various convex and non-convex classes of objective functions that correspond to loss functions for different machine learning models. We learn an optimizer for logistic regression, robust linear regression using the Geman-Mc Clure M-estimator and a twolayer neural net classifier with Re LU activation units. The geometry of the error surface becomes progressively more complex: the loss for logistic regression is convex, the loss for robust linear regression is non-convex, and the loss for the neural net has many local minima. |
| Researcher Affiliation | Academia | Ke Li & Jitendra Malik Department of Electrical Engineering and Computer Sciences University of California, Berkeley Berkeley, CA 94720 United States {ke.li,malik}@eecs.berkeley.edu |
| Pseudocode | Yes | Algorithm 1 General structure of unconstrained optimization algorithms |
| Open Source Code | No | The paper does not provide any links to source code or explicitly state that source code is available. |
| Open Datasets | No | The paper describes generating synthetic datasets for experiments ('randomly generate a dataset of 100 instances for each function', 'drawing 25 random samples from each one of four multivariate Gaussians'), but does not provide access information (links, DOIs, citations) to these datasets or the code used to generate them. |
| Dataset Splits | No | The paper specifies training and test sets ('we divide the dataset of objective functions into training and test sets'), but does not explicitly describe a separate validation set or provide specific details on its size or use for hyperparameter tuning in the context of their learned optimizer. |
| Hardware Specification | No | This research used the Savio computational cluster resource provided by the Berkeley Research Computing program at the University of California, Berkeley (supported by the UC Berkeley Chancellor, Vice Chancellor for Research, and Chief Information Officer). This mentions a cluster resource but lacks specific hardware details such as CPU/GPU models, memory, or number of nodes. |
| Software Dependencies | No | The paper mentions using a 'neural net' and 'guided policy search (GPS)' but does not specify any software platforms, libraries, or their version numbers used for implementation or experimentation. |
| Experiment Setup | Yes | We store the current iterate, previous gradients and improvements in the objective value from previous iterations in the state. We keep track of only the information pertaining to the previous H time steps and use H = 25 in our experiments. We use a small neural net with a single hidden layer of 50 hidden units to model the mean of π. Softplus activation units are used at the hidden layer and linear activation units are used at the output layer. We initialize the weights of the neural net randomly and do not regularize the magnitude of weights. For training, we sample 20 trajectories with a length of 40 time steps for each objective function in the training set. |