Experimentation is an intrinsic part of research in artificial intelligence since it allows for collecting quantitative observations, validating hypotheses, and providing evidence for their reformulation. For that reason, experimentation must be coherent with the purposes of the research, properly addressing the relevant questions in each case. Unfortunately, the literature is full of works whose experimentation is neither rigorous nor convincing, oftentimes designed to support prior beliefs rather than answering the relevant research questions. In this paper, we focus on the field of metaheuristic optimization, since it is our main field of work, and it is where we have observed the misconduct that has motivated this letter. Even if we limit the focus of this manuscript to the experimental part of the research, our main goal is to sew the seed of sincere critical assessment of our work, sparking a reflection process both at the individual and the community level. Such a reflection process is too complex and extensive to be tackled as a whole. Therefore, to bring our feet to the ground, we will include in this document our reflections about the role of experimentation in our work, discussing topics such as the use of benchmark instances vs instance generators, or the statistical assessment of empirical results. That is, all the statements included in this document are personal views and opinions, which can be shared by others or not. Certainly, having different points of view is the basis to establish a good discussion process.
Tokenization, the division of input text into input tokens, is an often overlooked aspect of the large language model (LLM) pipeline and could be the source of useful or harmful inductive biases. Historically, LLMs have relied on byte pair encoding, without care to specific input domains. With the increased use of LLMs for reasoning, various number-specific tokenization schemes have been adopted, with popular models like LLaMa and PaLM opting for single-digit tokenization while GPT-3.5 and GPT-4 have separate tokens for each 1-, 2-, and 3-digit numbers. In this work, we study the effect this choice has on numerical reasoning through the use of arithmetic tasks. We consider left-to-right and right-to-left tokenization for GPT-3.5 and -4, finding that right-to-left tokenization (enforced by comma separating numbers at inference time) leads to largely improved performance. Furthermore, we find that model errors when using standard left-to-right tokenization follow stereotyped error patterns, suggesting that model computations are systematic rather than approximate. We show that the model is able to convert between tokenizations easily, thus allowing chain-of-thought-inspired approaches to recover performance on left-to-right tokenized inputs. We also find the gap between tokenization directions decreases when models are scaled, possibly indicating that larger models are better able to override this tokenization-dependent inductive bias. In summary, our work performs the first study of how number tokenization choices lead to differences in model performance on arithmetic tasks, accompanied by a thorough analysis of error patterns. We hope this work inspires practitioners to more carefully ablate number tokenization-related choices when working towards general models of numerical reasoning.
In deep learning, stochastic gradient descent often yields functionally similar yet widely scattered solutions in the weight space even under the same initialization, causing barriers in the Linear Mode Connectivity (LMC) landscape. Overcoming these barriers is crucial for understanding deep learning dynamics and enhancing model-fusion algorithms. Previous studies highlight the role of permutation symmetry in reducing post-training barriers through network permutation. However, these post-hoc methods, demanding extra computations, are less effective for larger, complex models (e.g., ViT, LLM) due to numerous permutation matrices. Thus, in this paper, we study training-time neuron alignment. Our hypothesis suggests that training-time permutation subspace can reduce LMC barriers for free. We find that pruning at initialization supports this. Beyond pruning, we introduce TNA-PFN, a simple yet lossless algorithm using a partial gradient mask during training. TNA-PFN is theoretically and empirically validated for reducing LMC barriers. It excels in wide model fusion applications, especially in federated learning, two algorithms based on TNA-FPN that are proposed to show its prospects even under heterogeneous datasets. Moreover, TNA-PFN can enhance the generalization of model soup for vision transformers and ColD fusion for pretrained language models.
The attention mechanism, a cornerstone of state-of-the-art neural models, faces computational hurdles in processing long sequences due to its quadratic complexity. Consequently, research efforts in the last few years focused on finding more efficient alternatives. Among them, Hyena (Poli et al., 2023) stands out for achieving competitive results in both language modeling and image classification, while offering sub-quadratic memory and computational complexity. Building on these promising results, we propose ConfHyena, a Conformer whose encoder self-attentions are replaced with an adaptation of Hyena for speech processing, where the long input sequences cause high computational costs. Through experiments in automatic speech recognition (for English) and translation (from English into 8 target languages), we show that our best ConfHyena model significantly reduces the training time by 27%, at the cost of minimal quality degradation (~1%), which, in most cases, is not statistically significant.
We consider distributed kernel bandits where $N$ agents aim to collaboratively maximize an unknown reward function that lies in a reproducing kernel Hilbert space. Each agent sequentially queries the function to obtain noisy observations at the query points. Agents can share information through a central server, with the objective of minimizing regret that is accumulating over time $T$ and aggregating over agents. We develop the first algorithm that achieves the optimal regret order (as defined by centralized learning) with a communication cost that is sublinear in both $N$ and $T$. The key features of the proposed algorithm are the uniform exploration at the local agents and shared randomness with the central server. Working together with the sparse approximation of the GP model, these two key components make it possible to preserve the learning rate of the centralized setting at a diminishing rate of communication.
This paper introduces a Nonlinear Model Predictive Control (N-MPC) framework exploiting a Deep Neural Network for processing onboard-captured depth images for collision avoidance in trajectory-tracking tasks with UAVs. The network is trained on simulated depth images to output a collision score for queried 3D points within the sensor field of view. Then, this network is translated into an algebraic symbolic equation and included in the N-MPC, explicitly constraining predicted positions to be collision-free throughout the receding horizon. The N-MPC achieves real time control of a UAV with a control frequency of 100Hz. The proposed framework is validated through statistical analysis of the collision classifier network, as well as Gazebo simulations and real experiments to assess the resulting capabilities of the N-MPC to effectively avoid collisions in cluttered environments. The associated code is released open-source along with the training images.
When solving a task with limited labelled data, researchers can either use a general large language model without further update, or use the few examples to tune a specialised smaller model. When enough labels are available, the specialised models outperform the general ones on many NLP tasks. In this work, we aim to investigate how many labelled samples are required for the specialised models to achieve this superior performance, while taking the results variance into consideration. Observing the behaviour of prompting, in-context learning, fine-tuning and instruction-tuning, identifying their break-even points when increasing number of labelled training samples across three tasks of varying complexity, we find that the specialised models often need only few samples ($100-1000$) to be on par or better than the general ones. At the same time, the amount of required labelled data strongly depends on the task complexity and results variance.
Marine mammal communication is a complex field, hindered by the diversity of vocalizations and environmental factors. The Watkins Marine Mammal Sound Database (WMMD) is an extensive labeled dataset used in machine learning applications. However, the methods for data preparation, preprocessing, and classification found in the literature are quite disparate. This study first focuses on a brief review of the state-of-the-art benchmarks on the dataset, with an emphasis on clarifying data preparation and preprocessing methods. Subsequently, we propose the application of the Wavelet Scattering Transform (WST) in place of standard methods based on the Short-Time Fourier Transform (STFT). The study also tackles a classification task using an ad-hoc deep architecture with residual layers. We outperform the existing classification architecture by $6\%$ in accuracy using WST and $8\%$ using Mel spectrogram preprocessing, effectively reducing by half the number of misclassified samples, and reaching a top accuracy of $96\%$.
We present a new approach to direct depth estimation for Spatial Augmented Reality (SAR) applications using event cameras. These dynamic vision sensors are a great fit to be paired with laser projectors for depth estimation in a structured light approach. Our key contributions involve a conversion of the projector time map into a rectified X-map, capturing x-axis correspondences for incoming events and enabling direct disparity lookup without any additional search. Compared to previous implementations, this significantly simplifies depth estimation, making it more efficient, while the accuracy is similar to the time map-based process. Moreover, we compensate non-linear temporal behavior of cheap laser projectors by a simple time map calibration, resulting in improved performance and increased depth estimation accuracy. Since depth estimation is executed by two lookups only, it can be executed almost instantly (less than 3 ms per frame with a Python implementation) for incoming events. This allows for real-time interactivity and responsiveness, which makes our approach especially suitable for SAR experiences where low latency, high frame rates and direct feedback are crucial. We present valuable insights gained into data transformed into X-maps and evaluate our depth from disparity estimation against the state of the art time map-based results. Additional results and code are available on our project page: https://fraunhoferhhi.github.io/X-maps/
3D Gaussian splatting (3D-GS) is a new rendering approach that outperforms the neural radiance field (NeRF) in terms of both speed and image quality. 3D-GS represents 3D scenes by utilizing millions of 3D Gaussians and projects these Gaussians onto the 2D image plane for rendering. However, during the rendering process, a substantial number of unnecessary 3D Gaussians exist for the current view direction, resulting in significant computation costs associated with their identification. In this paper, we propose a computational reduction technique that quickly identifies unnecessary 3D Gaussians in real-time for rendering the current view without compromising image quality. This is accomplished through the offline clustering of 3D Gaussians that are close in distance, followed by the projection of these clusters onto a 2D image plane during runtime. Additionally, we analyze the bottleneck associated with the proposed technique when executed on GPUs and propose an efficient hardware architecture that seamlessly supports the proposed scheme. For the Mip-NeRF360 dataset, the proposed technique excludes 63% of 3D Gaussians on average before the 2D image projection, which reduces the overall rendering computation by almost 38.3% without sacrificing peak-signal-to-noise-ratio (PSNR). The proposed accelerator also achieves a speedup of 10.7x compared to a GPU.