A Simple Neural Attentive Meta-Learner

Authors: Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, Pieter Abbeel

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In the most extensive set of meta-learning experiments to date, we evaluate the resulting Simple Neural Attent Ive Learner (or SNAIL) on several heavily-benchmarked tasks. On all tasks, in both supervised and reinforcement learning, SNAIL attains state-of-the-art performance by significant margins.
Researcher Affiliation Collaboration UC Berkeley, Department of Electrical Engineering and Computer Science Embodied Intelligence {nmishra, rohaninejadm, c.xi, pabbeel}@berkeley.edu Accepted as a conference paper at ICLR 2018 Authors contributed equally and are listed in alphabetical order. Part of this work was done at Open AI.
Pseudocode Yes 1: function DENSEBLOCK(inputs, dilation rate R, number of filters D): 2: xf, xg = Causal Conv(inputs, R, D), Causal Conv(inputs, R, D) 3: activations = tanh(xf) * sigmoid(xg) 4: return concat(inputs, activations)
Open Source Code No Some video results can be found at https://sites.google.com/view/snail-iclr-2018/. (This link is for video results, not source code.)
Open Datasets Yes Introduced by Lake et al. (2011), Omniglot consists of black-and-white images of handwritten characters gathered from 50 languages, for a total of 1632 different classes with 20 instances per class. ... Mini-Image Net is a more difficult benchmark; a subset of the well-known Image Net dataset, it consists of 84 84 color images from 100 different classes with 600 instances per class. We used the split released by Ravi & Larochelle (2017) and used by a number of other works
Dataset Splits Yes Mini-Image Net ... We used the split released by Ravi & Larochelle (2017) and used by a number of other works, with 64 classes for training, 16 for validation, and 20 for testing.
Hardware Specification No The paper does not specify any particular hardware (e.g., GPU models, CPU types, or cloud computing instances) used for running the experiments.
Software Dependencies No The paper mentions optimizers like 'Adam (Kingma & Ba, 2015)' and methods like 'trust region policy optimization with generalized advantage estimation (TRPO with GAE; Schulman et al. (2015; 2016))', but does not list specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions or other libraries).
Experiment Setup Yes The hyperparameters are listed in Table 7. For multi-armed bandits, tabular MDPs, and visual navigation, we used the same hyperparamters as Duan et al. (2016) to make our results directly comparable; additional tuning could potentially improve SNAIL s performance. Table 7: The TRPO + GAE hyperparameters we used in our RL experiments. Hyperparameter: Batch Size (timesteps), Discount, GAE λ, Mean KL