Learning to solve the credit assignment problem

Authors: Benjamin James Lansdell, Prashanth Ravi Prakash, Konrad Paul Kording

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In both feedforward and convolutional networks, we empirically show that our approach learns to approximate the gradient, and can match or the performance of exact gradient-based learning.
Researcher Affiliation Academia Benjamin James Lansdell Department of Bioengineering University of Pennsylvania Pennsylvania, PA 19104 lansdell@seas.upenn.edu Prashanth Ravi Prakash Department of Bioengineering University of Pennsylvania Pennsylvania, PA 19104 Konrad Paul Kording Department of Bioengineering University of Pennsylvania Pennsylvania, PA 19104
Pseudocode No No structured pseudocode or algorithm blocks, labeled as 'Pseudocode' or 'Algorithm', were found in the paper.
Open Source Code Yes Code to reproduce these results can be found at: https://github.com/benlansdell/synthfeedback
Open Datasets Yes MNIST input data
Dataset Splits No The paper mentions using training and test sets (e.g., 'Dashed lines represent training loss, solid lines represent test loss' in Figure 3), and standard datasets like MNIST and CIFAR, but it does not explicitly provide specific percentages, sample counts, or detailed methodologies for training/validation/test dataset splits.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., CPU, GPU models, or cloud computing instances with specifications) used to run the experiments.
Software Dependencies No The paper states 'All code is implemented in Tensor Flow.' but does not specify a version number for TensorFlow or any other software dependencies with their versions.
Experiment Setup Yes Details of each task and parameters are provided here. All code is implemented in Tensor Flow. C.1 FIGURE 2: Networks are 784-50-20-10 with an MSE loss function. A sigmoid non-linearity is used. A batch size of 32 is used. B is updated using synthetic gradient updates with learning rate η = 0.0005, W is updated with learning rate 0.0004, standard deviation of noise is 0.01.