论文检索

Miscellaneous Aspects of Machine Learning/Transfer, Multitask and Meta-learning

Hong Chen, Xin Wang, Yuwei Zhou, Yijian Qin, Chaoyu Guan, Wenwu Zhu

Current auxiliary learning methods mainly adopt the methodology of reweighing losses for the manually collected auxiliary data and tasks. However, these methods heavily rely on domain knowledge during data collection, which may be hardly available in reality. Therefore, current methods will become less effective and even do harm to the primary task when unhelpful auxiliary data and tasks are employed. To tackle the problem, we propose a joint data-task generation framework for auxiliary learning (DTG-AuxL), which can bring benefits to the primary task by generating the new auxiliary data and task in a joint manner. The proposed DTG-AuxL framework contains a joint generator and a bi-level optimization strategy. Specifically, the joint generator contains a feature generator and a label generator, which are designed to be applicable and expressive for various auxiliary learning scenarios. The bi-level optimization strategy optimizes the joint generator and the task learning model, where the joint generator is effectively optimized in the upper level via the implicit gradient from the primary loss and the explicit gradient of our proposed instance regularization, while the task learning model is optimized in the lower level by the generated data and task. Extensive experiments show that our proposed DTG-AuxL framework consistently outperforms existing methods in various auxiliary learning scenarios, particularly when the manually collected auxiliary data and tasks are unhelpful.

Deep Learning/Robustness

Damien Teney, Yong Lin, Seong Joon Oh, Ehsan Abbasnejad

Several studies have compared the in-distribution (ID) and out-of-distribution (OOD) performance of models in computer vision and NLP. They report a frequent positive correlation and some surprisingly never even observe an inverse correlation indicative of a necessary trade-off. The possibility of inverse patterns is important to determine whether ID performance can serve as a proxy for OOD generalization capabilities.This paper shows that inverse correlations between ID and OOD performance do happen with multiple real-world datasets, not only in artificial worst-case settings. We explain theoretically how these cases arise and how past studies missed them because of improper methodologies that examined a biased selection of models.Our observations lead to recommendations that contradict those found in much of the current literature.- High OOD performance sometimes requires trading off ID performance.- Focusing on ID performance alone may not lead to optimal OOD performance. It may produce diminishing (eventually negative) returns in OOD performance.- In these cases, studies on OOD generalization that use ID performance for model selection (a common recommended practice) will necessarily miss the best-performing models, making these studies blind to a whole range of phenomena.

Social Aspects/Privacy-preserving Statistics and Machine Learning

Krishna Pillutla, Galen Andrew, Peter Kairouz, H. Brendan McMahan, Alina Oprea, Sewoong Oh

We present a rigorous methodology for auditing differentially private machine learning by adding multiple carefully designed examples called canaries. We take a first principles approach based on three key components. First, we introduce Lifted Differential Privacy (LiDP) that expands the definition of differential privacy to handle randomized datasets. This gives us the freedom to design randomized canaries. Second, we audit LiDP by trying to distinguish between the model trained with $K$ canaries versus $K-1$ canaries in the dataset, leaving one canary out. By drawing the canaries i.i.d., LiDP can leverage the symmetry in the design and reuse each privately trained model to run multiple statistical tests, one for each canary. Third, we introduce novel confidence intervals that take advantage of the multiple test statistics by adapting to the empirical higher-order correlations. Together, this new recipe demonstrates significant improvements in sample complexity, both theoretically and empirically, using synthetic and real data. Further, recent advances in designing stronger canaries can be readily incorporated in the new framework.

Deep Learning/Self-Supervised Learning

Athanasios Charisoudis, Simon Ekman von Huth, Emil Jansson

Scope of Reproducibility — The Masked Autoencoder (MAE) was recently proposed as aframework for efficient self‐supervised pre‐training in Computer Vision [1]. In this pa‐per, we attempt a replication of the MAE under significant computational constraints.Specifically, we target the claim that masking out a large part of the input image yieldsa nontrivial and meaningful self‐supervisory task, which allows training models thatgeneralize well. We also present the Semantic Masked Autoencoder (SMAE), a novel yetsimple extension of MAE which uses perceptual loss to improve encoder embeddings.Methodology — The datasets and backbones we rely on are significantly smaller than thoseused by [1]. Our main experiments are performed on Tiny ImageNet (TIN) [2] and trans‐fer learning is performed on a low‐resolution version of CUB‐200‐2011 [3]. We use aViT‐Lite [4] as backbone. We also compare the MAE to DINO, an alternative frame‐work for self‐supervised learning [5]. The ViT, MAE, as well as perceptual loss wereimplemented from scratch, without consulting the original authors’ code. Our code isavailable at https://github.com/MLReproHub/SMAE. The computational budget for ourreproduction and extension was approximately 150 GPU hours.Results — This paper successfully reproduces the claim that the MAE poses a nontrivialand meaningful self‐supervisory task. We show that models trained with this frame‐work generalize well to new datasets and conclude that the MAE is reproducible withexception for some hyperparameter choices. We also demonstrate that MAE performswell with smaller backbones and datasets. Finally, our results suggest that the SMAEextension improves the downstream classification accuracy of the MAE on CUB (+5 pp)when coupled with an appropriate masking strategy.What was easy — Given prior experience with a deep learning framework, re‐implementingthe paper was relatively straightforward, with sufficient details given in the paper.What was difficult — We faced challenges implementing efficient patch shuffling and tun‐ing hyperparameters. The hyperparameter choices from [1] did not translate well to asmaller dataset and backbone.Communication with original authors — We have not had contact with the original authors.

Martijn van Raaphorst, Egoitz Gonzalez, Marta Grasa, Paula Antequera Hernández

Scope of ReproducibilityIn this work we reproduce and extend the results presented in “Quantifying Societal Bias Amplification in Image Captioning” by Hirota et al. This paper introduces LIC, a metric to quantify bias amplification by image captioning models, which is tested for gender and racial bias amplification. The original paper claims that this metric is robust, and that all models amplify both gender and racial bias. It also claims that gender bias is more apparent than racial bias, and the Equalizer variation of the NIC+ model increases gender but not racial bias. We repeat the measurements to confirm these claims. We extend the analysis to whether the method can be generalized to other attributes such as bias in age.MethodologyThe authors of the paper provided a repository containing the necessary code. We had to modify it and add several scripts to be able to run all the experiments. The results were reproduced using the same subset of COCO [3] as in the original paper. Additionally, we manually labeled images according to age for our specific experiments. All experiments were ran on GPUs for a total of approximately 100 hours.ResultsAll claims made by the paper seem to hold, as the results we obtained follow the same trends as those presented in the original paper even if they do not match exactly. However, the same cannot always be said of the additional experiments.What was easyThe paper was clear and matched the implementation. The code was well organized and was easy to run using the command interface provided by the authors. This also made it easy to replicate and expand upon it by adding our own new features. The data was also readily available and could be easily downloaded with no need for preprocessing.What was difficultWe had to run several iterations of the same code, using different seeds and models, to get the results with the same conditions as in the original paper, which made use of time and resources. Our own experiments required additional time to hand-annotate data due to lack of data for new features.Communication with original authorsThere was no contact with the authors, since the code and the experiments were clear and did not need any additional explanation.

Deep Learning/Generative Models and Autoencoders

Yuxuan Ding, Chunna Tian, Haoxuan Ding, Lingqiao Liu

The Stable Diffusion model is a prominent text-to-image generation model that relies on a text prompt as its input, which is encoded using the Contrastive Language-Image Pre-Training (CLIP). However, text prompts have limitations when it comes to incorporating implicit information from reference images. Existing methods have attempted to address this limitation by employing expensive training procedures involving millions of training samples for image-to-image generation. In contrast, this paper demonstrates that the CLIP model, as utilized in Stable Diffusion, inherently possesses the ability to instantaneously convert images into text prompts. Such an image-to-prompt conversion can be achieved by utilizing a linear projection matrix that is calculated in a closed form. Moreover, the paper showcases that this capability can be further enhanced by either utilizing a small amount of similar-domain training data (approximately 100 images) or incorporating several online training steps (around 30 iterations) on the reference images. By leveraging these approaches, the proposed method offers a simple and flexible solution to bridge the gap between images and text prompts. This methodology can be applied to various tasks such as image variation and image editing, facilitating more effective and seamless interaction between images and textual prompts.

Marga Don, Satchit Chatterji, Milena Kapralova, Ryan Amaudruz

Scope of Reproducibility — This paper aims to reproduce the study FairCal: Fairness Calibration for Face Verification by Salvador et al., focused on verifying three main claims: FairCal (introduced by the authors) achieves state‐of‐the‐art (i) global accuracy, (ii) fairness-calibrated probabilities and (iii) equality in false positive rates across sensitive attributes (i.e. predictive equality). The sensitive attribute taken into account is ethnicity.Methodology — Salvador et al. provide partial code via a GitHub repository. Additional code to generate image embeddings from three pretrained neural network models were based on existing repositories. All code was refactored to fit our needs, keeping extendability and readability in mind. Two datasets were used, namely, Balanced Faces in the Wild (BFW) and Racial Faces in the Wild (RFW). Additional experiments using Gaussian mixture models instead of K‐means clustering for FairCal validate the use of unsupervised clus‐ tering methods. The code was run on an AMD Ryzen 7 2700X CPU and NVIDIA GeForce GTX1080Ti GPU with a total runtime of around 3 hours for all experiments.Results — In most cases, we were able to reproduce results from the original paper to within 1 standard deviation, and observe similar trends. However, due to missing information about image pre‐processing, we were unable to reproduce the results exactly.What was easy — The original paper is clear and understandable. Furthermore, the authors provided a mostly working version of the code. Though the datasets are not freely available to the public, their authors supplied these to us swiftly after contacting them.What was difficult — While most of the code worked with slight changes, it was assumed there were files containing image embeddings available for both datasets, which the authors neither provided nor gave details about. We therefore pre‐processed and generated embeddings independent of the authors, which makes it more difficult to judge the overall reproducibility of their method. Additionally, we encountered difficulties while improving the efficiency and extendability of the code.Communication with original authors — We emailed the first author of the paper twice. First at the beginning of our undertaking, they were enthusiastic about our attempt, and clarified a few initial doubts about their implementation, the embeddings, and missing files. As per the writing of this paper, they have not responded to the second email.

Social Aspects/Fairness, Equity, Justice and Safety

Valentin Buchner, Philip Schutte, Yassin Ben Allal, Hamed Ahadi

Scope of Reproducibility: The original authors' main contribution is the family of Shifty algorithms, which can guarantee that certain fairness constraints will hold with high confidence even after a demographic shift in the deployment population occurs. They claim that Shifty provides these high-confidence fairness guarantees without a loss in model performance, given enough training data.Methodology: The code provided by the original paper was used, and only some small adjustments needed to be made in order to reproduce the experiments. All model specifications and hyperparameters from the original implementation were used. Extending beyond reproducing the original paper, we investigated the sensibility of Shifty to the size of the bounding intervals limiting the possible demographic shift, and ran shifty with an additional optimization method.Results: Our results approached the results reported in the original paper. They supported the claim that \textit{Shifty} reliably guarantees fairness under demographic shift, but could not verify that Shifty performs at no loss of accuracy. What was easy: The theoretical framework laid out in the original paper was well explained and supported by additional formulas and proofs in the appendix. Further, the authors provided clear instructions on how to run the experiments and provided necessary hyperparameters.What was difficult: While an open-source implementation of Shifty was provided and was debugged with relatively low time investment, the code did not contain extensive documentation and was complex to understand. It was therefore difficult to verify that each part of the code functions as expected and to expand upon the existing experiments. Further, certain hyperparameter and model specifications deviated between the provided code and the original paper, which made it challenging to know which specifications to apply when reproducing.Communication with original authors: The first author of the original paper was contacted, but we have yet to receive a reply.

Miscellaneous Aspects of Machine Learning/Everything Else

Robin Sasse, Aniek Eijpe, Jona Ruthardt, Elias Dubbeldam

Scope of Reproducibility — CartoonX [1] is a novel explanation method for image classifiers. In this reproducibility study, we examine the claims of the original authors of CartoonX that it: (i) extracts relevant piece‐wise smooth parts of the image, resulting in explanations which are more straightforward to interpret for humans; (ii) achieves lower distortion in the model output, using fewer coefficients than other state-of‐the‐art methods; (iii) is model‐agnostic. Finally, we examine how to reduce the runtime.Methodology — The original authors’ open‐sourced implementation has been used to examine (i). We implemented the code to examine (ii), as there was no public code available for this. We tested claim (iii) by performing the same experiments with a Vision Transformer instead of a CNN. To reduce the runtime, we extended the existing implementation with multiple enhanced initialization techniques. All experiments took approximately 38.4 hours on a single NVIDIA Titan RTX.Results — Our results support the claims made by the original authors. (i) We observe that CartoonX produces piece‐wise smooth explanations. Most of the explanations give valuable insights. (ii) Most experiments, that show how CartoonX achieves lower distortion outputs compared to other methods, have been reproduced. In the cases where exact reproducibility has not been achieved, claim (ii) of the author still holds. (iii) The model‐agnosticism claim still holds as the overall quality of the ViT‐based explanations almost matches that of the CNN‐based explanations. Finally, simple heuristical initializations did not improve the runtime.What was easy — The mathematical background and intuition of CartoonX were clearly explained by the original authors. Moreover, the original author’s code was well structured and documented, which made it straightforward to run and extend.What was difficult — Some hyperparameter settings and implementation details needed to reproduce the experiments were not clear or transparent from the original paper or code. This made it difficult to implement and reproduce these experiments.Communication with original authors — We have been in brief communication with the original authors. They were able to address most of our points, providing us with some additional clarifications about the exact implementation and hyperparameter settings.

Farrukh Baratov, Goksenin Yuksel, Darie Petcu, Jan Bakker

Scope of reproducibility - We study the reproducibility of the paper "Quantifying Societal Bias Amplification in Image Captioning" by Hirota et al. In this paper, the authors propose a new metric to measure bias amplification, called LIC, and evaluate it on multiple image captioning models. Based on this evaluation, they make the following main claims which we aim to verify: (1) all models amplify gender bias, (2) all models amplify racial bias, (3) LIC is robust against encoders, and (4) the NIC+Equalizer model increases gender bias with respect to the baseline. We also extend upon the original work by evaluating LIC for age bias.Methodology - For our reproduction, we were able to run the code provided by the authors without any modifications. For our extension, we automatically labelled the images in the dataset with age annotations and adjusted the code to work with this dataset. In total, 38 GPU hours were needed to perform all experiments.Results - The reproduced results are close to the original results and support all four main claims. Furthermore, our additional results show that only a subset of the models amplifies age bias, while they strengthen the claim that LIC is robust against encoders. However, we acknowledge that our extension to age bias has its limitations.What was easy - The author's code and the data needed to run it are publicly available. The code required no modification to run and the scripts were provided with an extensive argument parser, allowing us to quickly set up our experiments. Moreover, the details of the original experiments were clearly stated in the appendix.What was difficult - We found that it was difficult to interpret the author's code as the provided documentation contained room for improvement. Also, the scripts contained repetitive code. While the authors retrained all image captioning models, they did not share the model weights, making it difficult to extend upon their work.Communication with original authors - No (attempt at) communication with the original authors was performed.

Martin Špendl, Klementina Pirc

REPRODUCIBILITY SUMMARYScope of ReproducibilityIn this work, we study the reproducibility of the paper: Pre-Train Your Loss: Easy Bayesian Transfer Learning with Informative Priors. The paper proposes a three-step pipeline for replacing standard transfer learning with a pre-trained prior. The first step is training a prior, the second is re-scaling of a prior, and the third is inference.The authors claim that increasing the rank and the scaling factor improves performance on the downstream task. They also argue that using Bayesian learning with informative prior leads to a more data-efficient and improved performance compared to standard SGD transfer learning or using non-informative prior. We reproduce the main claims on one of the four data sets in the paper.MethodologyWe used a combination of the authors' and our code. The authors provided a training pipeline for the user but not the code to fully reproduce the paper. We modified the training pipeline to suit our needs and created a testing pipeline to evaluate the models. We reproduced the results for the Oxford-102-Flowers data set on an Nvidia RTX 3070 GPU using approximately 310 GPU hours for the main results.ResultsOur results confirm most of the claims tested, although we could not achieve the exact same accuracy due to missing hyper-parameters. We reproduced the trend in how scaling the prior impacts the performance and how a learned prior outperforms a non-learned prior.On contrary, we could not reproduce the effect of rank in low-rank covariance approximation on model performance, as well as the beneficial boost in performance of Bayesian learning compared to the standard SGD.What was easyThe authors' implementation provides various training and logging parameters. It is also helpful that the authors provided both the learned priors and scripts for the download, split and pre-processing of the data sets used in the study.What was difficultSetting the environment for the used packages to work correctly was difficult. Although many parameters are available for running the pipeline, their descriptions are misguiding, therefore a lot of time went into clarifying the parameter function and debugging different settings. The training also took a while, especially when training 5 models per data point. Communication with original authorsWe contacted the authors via e-mail about their pipeline and their use of hyper-parameters but did not hear back.

Deep Learning/Recurrent Networks

Sean McLeish, Long Tran-Thanh

Scope of Reproducibility:In this report, we aim to validate the claims of Bansal et al. These are that the recurrent architecture presented, with skip connections and a progressive loss function, prevent the original problem being forgotten or corrupted during processing allowing for the recurrent module to be applied indefinitely and that this architecture avoids the overthinking trap. We use both code released by the authors and newly developed to recreate many results presented in the paper. Additionally, we present analysis of the newly introduced alpha hyperparameter and investigate interesting perturbation behaviour of prefix sums models. Further, we conduct a hyperparameter search and provide an analysis of the Asymptotic Alignment scores of the models presented.Methodology:We use the PyTorch code released by the authors to replicate accuracy experiments. We then, independently, develop our own PyTorchFI to replicate perturbation experiments presented by Bansal et al. Overall, providing a replication of all results shown in the main body of the paper. We then extend these results, providing an analysis of the alpha hyperparameter, analysis of perturbation recovery, Asymptotic Alignment scores and a hyperparameter search. We used both a Nvidia RTX 2080Ti GPU and sets of three NVIDIA Quadro RTX6000 GPUs, taking a total of 982.2 GPU hours to create all results presented in the main body of this report.Results:We verify the authors' claims by replicating the experiments presented by Bansal et al. All of our experiments show identical results to the ones presented in the paper, apart from perturbation testing for which we provide an additional in depth analysis. We also provide an analysis of the new alpha hyperparameter and a hyperparameter search.What was easy:The code provided by Bansal et al gave clear instructions on how to use it, along with pretrained models being available for all problems.What was difficult:Chess models required a considerable amount of time to train, putting a drain on resources. Also, code for reproducing perturbation results was not available so this had to developed from scratch.Communication with original authors:We had good communication with the original authors, both emailing and meeting online.

Ermin Omeragic, Vuk Đuranović

Scope of ReproducibilityThis paper presents a novel augmentation method that can be used for graph classification tasks: $\mathcal{G}$-Mixup. Our goal is to reproduce eight claims that the authors make in their paper. The first two claims relate to the properties of graphons estimated from graphs, which are the main components of the method. Claims three to eight relate to the superior performance of the method compared to other augmentation strategies.MethodologyTo reproduce the results, we use the open-source implementation of the method provided by the authors, with a few modifications. We write from scratch all the experiments and pipelines needed to defend the claims of the paper. Additionally, we implement three out of four baseline augmentation methods that are compared to the novel method. For one part of the experiments, we use a local computer and run the experiments on a CPU, with a total of 31.7 CPU hours, while for other more demanding experiments, we use a GPU-accelerated machine for a total of 157.3 GPU hours.ResultsDue to many missing implementation details, we were not able to reproduce all of the original results. Some claims can be supported by our results, but most results are very vague. Even though the new method outperforms the baselines in certain scenarios, we find that the superiority of the method is not as strong as presented in the original paper.What was easyThe novel augmentation method and its theoretical justification are presented intuitively in the paper, and it was easy to grasp the main ideas of the paper.What was difficultWhile the code with the method implementation was given, the reproduction of all the results required much more code and details than what was provided in the paper. This means that we had to make a lot of educated guesses about the experimental settings, choices of hyperparameters, and details about the models used. Communication with original authorsWe contacted the authors on two occasions. The first time was very early in the reproduction process, when we asked about the details of the graph estimation methods that they used and which weren't explained in the paper, to which they responded swiftly. On the second occasion, we asked about the experimental settings and hyperparameter details, but we did not receive a reply.

Miscellaneous Aspects of Machine Learning/Representation Learning

Xuanjie Liu, Daniel Chin, Yichen Huang, Gus Xia

We have recently seen great progress in learning interpretable music representations, ranging from basic factors, such as pitch and timbre, to high-level concepts, such as chord and texture. However, most methods rely heavily on music domain knowledge. It remains an open question what general computational principles *give rise to* interpretable representations, especially low-dim factors that agree with human perception. In this study, we take inspiration from modern physics and use *physical symmetry* as a self-consistency constraint for the latent space. Specifically, it requires the prior model that characterises the dynamics of the latent states to be *equivariant* with respect to certain group transformations. We show that physical symmetry leads the model to learn a *linear* pitch factor from unlabelled monophonic music audio in a self-supervised fashion. In addition, the same methodology can be applied to computer vision, learning a 3D Cartesian space from videos of a simple moving object without labels. Furthermore, physical symmetry naturally leads to *counterfactual representation augmentation*, a new technique which improves sample efficiency.

Deep Learning/Generative Models and Autoencoders

Volodymyr Kyrylov, Navdeep Singh Bedi, Qianbo Zang

Scope of Reproducibility — Exponential family variational autoencoders struggle with reconstruction when encoders output limited information. We reproduce two experiments in which we first train the decoder and encoder separately. Then, we train both modules jointly using ELBO and observe the degradation of reconstruction. We verify how the theoretical insight into the design of the approximate posterior and decoder distributions for a discrete VAE in a semantic hashing application influences the choice of input features to improve overall performance.Methodology — We implement and train a synthetic experiment from scratch on a laptop. We use a mix of authors’ code and publicly available code to reproduce a GAN reinterpreted as a VAE. We consult authors’ code to reproduce semantic hashing experiment and make our own implementation. We train models the USI HPC cluster on machines with GTX 1080 Ti or A100 GPUs and 128 GiB of RAM. We spend under 100 GPU hours for all experiments.Results — We observe expected qualitative behavior predicted by the theory on all experiments. We improve the best semantic hashing model’s test performance by 5 nats by using a modern method for gradient estimation of discrete random variables.What was easy — Following experiment recipes was easy once we worked out the theory.What was difficult — The paper enables verification of exponential family distributions VAE designs of arbitrary complexity, which require probabilistic modeling skills. We contribute elaboration on the implementation details of the synthetic experiment and provide code.Communication with original authors — We are extremely grateful to the authors for discussing the paper, encouraging us to implement experiments on our own, and suggesting directions for improving results over e‐mail.

Domen Mohorčič, David Ocepek

Scope of Reproducibility: The paper presents a novel post-hoc regularization technique for tree-based models, called Hierarchical Shrinkage (Agarwal 2022). Our main goal is to confirm the claims that it substantially increases the predictive performance of both decision trees and random forests, that it is faster than other regularization techniques, and that it makes the interpretation of random forests simpler. Methodology: In our reproduction, we used the Hierarchical Shrinkage, provided by the authors in the Python package imodels. We also used their function for obtaining pre-cleaned data sets. While the algorithm code and clean datasets were provided we re-implemented the experiments as well as added additional experiments to further test the validity of the claims. The results were tested by applying Hierarchical Shrinkage to different tree models and comparing them to the authors' results. Results: We managed to reproduce most of the results the authors get. The method works well and most of the claims are supported. The method does increase the predictive performance of tree-based models most of the time, but not always. When compared to other regularization techniques the Hierarchical Shrinkage outperforms them when used on decision trees, but not when used on random forests. Since the method is applied after learning, it is extremely fast. And it does simplify decision boundaries for random forests, making them easier to interpret. What was easy: The use of the official code for Hierarchical Shrinkage was straightforward and used the same function naming convention as other machine learning libraries. The function for acquiring already clean data sets saved a lot of time. What was difficult: The authors also provided the code for their experiments in a separate library, but the code did not run out of the box and we had no success reproducing the results with it. The code was inconsistent with the paper methodology. We had the most problems with hyperparameter tuning. The authors did not specify how they tuned the hyperparameters for the used RF regularizers. Communication with original authors: We did not contact the authors of the original paper

Applications/Neuroscience, Cognitive Science

Jingyuan Li, Leo Scholl, Trung Le, Pavithra Rajeswaran, Amy Orsborn, Eli Shlizerman

Latent Variable Models (LVMs) propose to model the dynamics of neural populations by capturing low-dimensional structures that represent features involved in neural activity. Recent LVMs are based on deep learning methodology where a deep neural network is trained to reconstruct the same neural activity given as input and as a result to build the latent representation. Without taking past or future activity into account such a task is non-causal. In contrast, the task of forecasting neural activity based on given input extends the reconstruction task. LVMs that are trained on such a task could potentially capture temporal causality constraints within its latent representation. Forecasting has received less attention than reconstruction due to recording challenges such as limited neural measurements and trials. In this work, we address modeling neural population dynamics via the forecasting task and improve forecasting performance by including a prior, which consists of pairwise neural unit interaction as a multivariate dynamic system. Our proposed model---Additive, Multiplicative, and Adaptive Graph Neural Network (AMAG)---leverages additive and multiplicative message-passing operations analogous to the interactions in neuronal systems and adaptively learns the interaction among neural units to forecast their future activity. We demonstrate the advantage of AMAG compared to non-GNN based methods on synthetic data and multiple modalities of neural recordings (field potentials from penetrating electrodes or surface-level micro-electrocorticography) from four rhesus macaques. Our results show the ability of AMAG to recover ground truth spatial interactions and yield estimation for future dynamics of the neural population.

Gerson de Kleuver, David Bikker, Wenhua Hu, Bram Veenman

Scope of Reproducibility — This paper analyses the reproducibility of the study Proto2Proto: Can you recognize the car, the way I do? The main contributions and claims of the study are: 1) Using Proto2Proto, a shallower student model is more faithful to the teacher in terms of interpretability than a baseline student model while also showing the same or better accuracy; 2) Global Explanation loss forces student prototypes to be close to teacher prototypes; 3) Patch‐Prototype Correspondence loss enforces the local representations of the student to be similar to those of the teacher; 4) The proposed evaluation metrics determine the faithfulness of the student to the teacher in terms of interpretability.Methodology — A public code repository was available for the paper, which provided a working but incomplete and minimally documented codebase. With some modifications we were able to carry out the experiments that were best supported by the codebase. We spent a total of 60 computational GPU hours on reproduction.Results — The results we were able to produce support claim 1, albeit weakly. Further results are in line with the paper, but we found them to go against claim 3. In addition, we carried out a theoretical analysis which provides support for claim 4. Finally, we were unable to carry out our intended experiment to verify claim 2. What was easy — The original paper was clearly structured and understandable. The experiments for which configurations were provided were simple to conduct.What was difficult — The public codebase contained minimal documentation. Moreover, the use of variable names did not correspond between the code and the paper. Furthermore, the codebase lacked elements vital to reproducing some experiments. Another significant constraint were the computational requirements needed to reproduce the original experiments. Finally, the code required to reproduce one of the visualizations was not provided.Communication with original authors — We contacted the authors to ask for trained model weights and missing hyperparameters for several experiments. We did not receive aresponse.

Social Aspects/Accountability, Transparency and Interpretability

Kyosuke Morita

Scope of ReproducibilityThis study aims to reproduce the results of the paper 'FOCUS: Flexible Optimizable Counterfactual Explanations for Tree Ensembles' by Lucic et al.The main claims of the original paper are that FOCUS is able to (i) generate counterfactual explanations for all the instances in a dataset; and (ii) find counterfactual explanations that are closer to the original input for tree-based algorithms than existing methods.MethodologyThis study replicates the original experiments using the code, data, and models provided by the authors. Additionally, this study re-implements code and retrains the models to evaluate the robustness and generality of FOCUS.All the experiments were conducted on a personal laptop with a quad-core CPU with 8GB of RAM and it approximately took 33 hours in total.ResultsThis study was able to replicate the results of the original paper in terms of finding counterfactual explanations for all instances in datasets. Additional experiments were conducted to validate the robustness and generality of the conclusion. While there were slight deviations in terms of generating smaller mean distances, half of the models still outperformed the results of the existing method.What was easyThe implementation of the original paper is publicly available on GitHub. The repository contains the models and data used in the original experiments. Also, the authors provided a technical appendix, which includes all the hyperparameters that were used for the experiments for reproduction upon request.What was difficultAlthough the implementation code was available, it employs outdated packages and the code structure is complex. Also, the comments in the functions and the documentation of the code are sparse or nonexistent, which made it difficult to follow the code.Communication with original authorsI reached out to the authors to obtain the hyperparameters used in the experiments. The authors responded promptly with a detailed technical appendix of the original paper.

Priyanka Bose, Chandra Shekhar Pandey, Fraida Fund

Reproducibility SummaryScope of Reproducibility — We aim to reproduce a result from the paper “Exploring the Role of Grammar and Word Choice in Bias Toward African American English (AAE) in Hate Speech Classification” [1]. Our study is restricted specifically to the claim that the use of swear words impacts hate speech classification of AAE text. We were able to broadly validate the claim of the paper, however, the magnitude of the effect was dependent on the word replacement strategy, which was somewhat ambiguous in the original paper.Methodology — The authors did not publish source code. Therefore, we reproduce the experiments by following the methodology described in the paper. We train BERT models from TensorFlow Hub [2] to classify hate speech using the DWMW17[3] and FDCL18[4]Twitter datasets. Then, we compile a dictionary of swear words and replacement words with comparable meaning, and we use this to create “censored” versions of samples in Blodgett et al.’s[5] AAE Twitter dataset. Using the BERT models, we evaluate the hate speech classification of the original data and the censored data. Our experiments are conducted on an open‐access research testbed, Chameleon [6], and we make available both our code and instructions for reproducing the result on the shared facility.Results — Our results are consistent with the claim that the censored text (without swear words) is less often classified as hate speech, offensive, or abusive than the same text with swear words. However, we find the classification is very sensitive to the word replacement dictionary being used.What was easy — The authors used three well known datasets which were easy to obtain. They also used well‐known widely available models, BERT and Word2Vec.What was difficult — Some of the details of model training were not fully specified. Also, we were not able to exactly re‐create a comparable dictionary of swear words and their replacement terms of similar meanings, following their methodology.Communication with original authors — We reached out to the authors, but were not able to communicate with them before the submission of this report.