Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Learning with Fenchel-Young losses

Authors: Mathieu Blondel, André F.T. Martins, Vlad Niculae

JMLR 2020 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 9. Experiments In this section, we demonstrate one of the key features of Fenchel-Young losses: their ability to induce sparse probability distributions. We focus on two tasks: label proportion estimation ( 9.1) and dependency parsing ( 9.2). [...] 9.1. Label proportion estimation experiments [...] Table 3: Dataset statistics [...] Table 4: Test-set performance of Tsallis losses [...] Figure 10: Jensen-Shannon divergence between predicted and true label proportions, when varying document length, of various losses generated by a Tsallis entropy.
Researcher Affiliation Collaboration Mathieu Blondel EMAIL NTT Communication Science Laboratories Kyoto, Japan Andre F. T. Martins EMAIL Unbabel & Instituto de Telecomunicações Lisbon, Portugal Vlad Niculae EMAIL Instituto de Telecomunicações Lisbon, Portugal
Pseudocode Yes Algorithm 1 Bisection for byΩ(θ) = Ω (θ) Input: θ Rd, Ω(p) = I d + P i g(pi) p(τ) := (g ) 1(max{θ τ, g (0)}) φ(τ) := p(τ), 1 1 τmin max(θ) g (1); τmax max(θ) g (1/d) τ (τmin + τmax)/2 while |φ(τ)| > ϵ if φ(τ) < 0 τmax τ else τmin τ τ (τmin + τmax)/2 Output: byΩ(θ) p(τ)
Open Source Code No The text is ambiguous or lacks a clear, affirmative statement of release. No specific link to a code repository or explicit statement about the release of their implementation code was found.
Open Datasets Yes We ran experiments on 7 standard multi-label benchmark datasets see Table 3 for dataset characteristics1. 1. The datasets can be downloaded from http://mulan.sourceforge.net/datasets-mlc.html and https: //www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/.
Dataset Splits Yes We use 1200 samples as training set, 200 samples as validation set and 1000 samples as test set. [...] Table 3: Dataset statistics Dataset Type Train Dev Test Features Classes Avg. labels Birds Audio 134 45 172 [...] Yeast Micro-array 1,125 375 917
Hardware Specification No No specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments were found in the paper.
Software Dependencies No No specific software versions (e.g., Python 3.8, PyTorch 1.9, CUDA 11.1) for key components were provided. The paper mentions algorithms like L-BFGS and Adam, but not the specific software implementations or their versions.
Experiment Setup Yes We chose λ {10 4, 10 3, . . . , 104} and α {1, 1.1, . . . , 2} against the validation set. [...] Parameters are trained using Adam (Kingma and Ba, 2015), tuning the learning rate on the grid {.5, 1, 2, 4, 8} 10 3, expanded by a factor of 2 if the best model is at either end.