Differentiable Programs with Neural Libraries

Authors: Alexander L. Gaunt, Marc Brockschmidt, Nate Kushman, Daniel Tarlow

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we show that this leads to lifelong learning systems that transfer knowledge to new tasks more effectively than baselines. The experimental results are presented in Sec. 6.
Researcher Affiliation Industry 1Microsoft Research, Cambridge, UK 2Google Brain, Montréal, Canada (work done while at Microsoft). Correspondence to: Alexander L. Gaunt <algaunt@microsoft.com>.
Pseudocode Yes Figure 1 shows 'Learned operations' and an 'Instruction Set Declaration & initialization' block, including variable declarations, loops, and conditional statements, formatted like pseudocode.
Open Source Code No The paper provides a link for the Neural GPU baseline ('available at https://github.com/tensorflow/models/tree/master/neural_gpu') but does not state that the code for their own method (NTPT) is open-source or publicly available.
Open Datasets Yes ADD2X2 scenario: The first scenario in Fig. 2(a) uses of a 2x2 grid of MNIST digits.
Dataset Splits No We detect convergence by a rapid increase in the accuracy on a validation set (typically occurring after around 30k training examples). However, the paper does not specify the size or exact splitting methodology of this validation set.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper mentions 'TensorFlow (Abadi et al., 2015)' but does not provide a specific version number for TensorFlow or any other software dependencies.
Experiment Setup Yes We have a separate learning rate for the perceptual networks in both the MTNN baseline and NTPT which is 100 fold smaller than the learning rate for the task-specific parts.