Arda Fazla

Arda Fazla, Abolfazl Hashemi

Real-world datasets often contain spurious correlations that are not causally related to the target label. When such correlations dominate the majority of training samples, models tend to rely on them, leading to misclassification of minority samples that do not exhibit the same spurious patterns. While a potential approach is to select subsets of data to better represent the minority samples, this may require access to group labels, which are typically unknown. Furthermore, as we demonstrate, widely used sample scoring functions in the invariant subset or coreset selection literature largely depend on spurious features and therefore fail to accurately capture the importance or difficulty of core, causally relevant features. Accordingly, we propose to mitigate spurious correlations by developing a two-stage sample scoring function that disentangles the learning dynamics of core and spurious features and evaluates their difficulty separately. Based on our proposed metric, we introduce a new algorithm to find and prioritize informative samples both with and without spurious correlations. Extensive experiments demonstrate that a standard ERM model trained on our selected samples achieves superior performance compared to state-of-the-art debiasing techniques, while requiring as little as 10\% of the original training data.

Arda Fazla, Mustafa Enes Aydin, Orhun Tamyigit et al.

We investigate ensemble methods for prediction in an online setting. Unlike all the literature in ensembling, for the first time, we introduce a new approach using a meta learner that effectively combines the base model predictions via using a superset of the features that is the union of the base models' feature vectors instead of the predictions themselves. Here, our model does not use the predictions of the base models as inputs to a machine learning algorithm, but choose the best possible combination at each time step based on the state of the problem. We explore three different constraint spaces for the ensembling of the base learners that linearly combines the base predictions, which are convex combinations where the components of the ensembling vector are all nonnegative and sum up to 1; affine combinations where the weight vector components are required to sum up to 1; and the unconstrained combinations where the components are free to take any real value. The constraints are both theoretically analyzed under known statistics and integrated into the learning procedure of the meta learner as a part of the optimization in an automated manner. To show the practical efficiency of the proposed method, we employ a gradient-boosted decision tree and a multi-layer perceptron separately as the meta learners. Our framework is generic so that one can use other machine learning architectures as the ensembler as long as they allow for a custom differentiable loss for minimization. We demonstrate the learning behavior of our algorithm on synthetic data and the significant performance improvements over the conventional methods over various real life datasets, extensively used in the well-known data competitions. Furthermore, we openly share the source code of the proposed method to facilitate further research and comparison.

Mustafa E. Aydın, Arda Fazla, Suleyman S. Kozat

We investigate nonlinear prediction/regression in an online setting and introduce a hybrid model that effectively mitigates, via a joint mechanism through a state space formulation, the need for domain-specific feature engineering issues of conventional nonlinear prediction models and achieves an efficient mix of nonlinear and linear components. In particular, we use recursive structures to extract features from raw sequential sequences and a traditional linear time series model to deal with the intricacies of the sequential data, e.g., seasonality, trends. The state-of-the-art ensemble or hybrid models typically train the base models in a disjoint manner, which is not only time consuming but also sub-optimal due to the separation of modeling or independent training. In contrast, as the first time in the literature, we jointly optimize an enhanced recurrent neural network (LSTM) for automatic feature extraction from raw data and an ARMA-family time series model (SARIMAX) for effectively addressing peculiarities associated with time series data. We achieve this by introducing novel state space representations for the base models, which are then combined to provide a full state space representation of the hybrid or the ensemble. Hence, we are able to jointly optimize both models in a single pass via particle filtering, for which we also provide the update equations. The introduced architecture is generic so that one can use other recurrent architectures, e.g., GRUs, traditional time series-specific models, e.g., ETS or other optimization methods, e.g., EKF, UKF. Due to such novel combination and joint optimization, we demonstrate significant improvements in widely publicized real life competition datasets. We also openly share our code for further research and replicability of our results.

Antesh Upadhyay, Arda Fazla, Abolfazl Hashemi

We study nonconvex stochastic optimization under the Blum-Gladyshev ($\mathsf{BG}$-0) noise model, where the stochastic gradient variance grows quadratically with the distance from the initialization. We consider this problem under both standard smoothness and the symmetric generalized-smoothness framework, which captures objectives whose local curvature can scale with the gradient norm. We prove that normalized stochastic gradient descent with momentum, using only one stochastic gradient per iteration, converges under $\mathsf{BG}$-0 noise with oracle complexity $O(\varepsilon^{-6})$. This rate holds both for standard smoothness and for $α$-symmetric generalized smoothness, showing that generalized smoothness is rate-neutral for normalized momentum in this setting. We then study a variance-reduced normalized STORM method. Under mean-square smoothness and sharp initialization, the method achieves the minimax optimal $O(\varepsilon^{-4})$ complexity, matching the lower bound. Under expected $α$-symmetric generalized smoothness, the STORM recursion couples gradient-dependent smoothness with distance-dependent noise, leading to complexity $O(\varepsilon^{-(4+α)})$ for $α\in(0,1)$ and $O(\varepsilon^{-5})$ for $α=1$. When the distance-growth parameter in the noise model vanishes, our guarantees recover the standard bounded-variance rates: $O(\varepsilon^{-4})$ for momentum, $O(\varepsilon^{-3})$ for variance reduction, and $O(\varepsilon^{-2})$ in the deterministic case. To our knowledge, these are the first convergence guarantees for normalized methods in non-convex stochastic optimization under $\mathsf{BG}$-0 noise without bounded domains, increasing batch sizes, or explicit anchoring, covering both standard and generalized smoothness regimes.

Arda Fazla, Ege C. Kaya, Antesh Upadhyay et al.

Analysis of Stochastic Gradient Descent (SGD) and its variants typically relies on the assumption of uniformly bounded variance, a condition that frequently fails in practical non-convex settings, such as neural network training, as well as in several elementary optimization settings. While several relaxations are explored in the literature, the Blum-Gladyshev (BG-0) condition, which permits the variance to grow quadratically with distance has recently been shown to be the weakest condition. However, the study of the oracle complexity of stochastic first-order non-convex optimization under BG-0 has remained underexplored. In this paper, we address this gap and establish information-theoretic lower bounds, proving that finding an $ε$-stationary point requires $Ω(ε^{-6})$ stochastic BG-0 oracle queries for smooth functions and $Ω(ε^{-4})$ queries under mean-square smoothness. These limits demonstrate an unavoidable degradation from classical bounded-variance complexities, i.e., $Ω(ε^{-4})$ and $Ω(ε^{-3})$ for smooth and mean-square smooth cases, respectively. To match these lower bounds, we consider Proximally Anchored STochastic Approximation (PASTA), a unified algorithmic framework that couples Halpern anchoring with Tikhonov regularization to dynamically mitigate the extra variance explosion term permitted by the BG-0 oracle. We prove that PASTA achieves minimax optimal complexities across numerous non-convex regimes, including standard smooth, mean-square smooth, weakly convex, star-convex, and Polyak-Lojasiewicz functions, entirely under an unbounded domain and unbounded stochastic gradients.

Selim Furkan Tekin, Arda Fazla, Suleyman Serdar Kozat

Numerical weather forecasting using high-resolution physical models often requires extensive computational resources on supercomputers, which diminishes their wide usage in most real-life applications. As a remedy, applying deep learning methods has revealed innovative solutions within this field. To this end, we introduce a novel deep learning architecture for forecasting high-resolution spatio-temporal weather data. Our approach extends the conventional encoder-decoder structure by integrating Convolutional Long-short Term Memory and Convolutional Neural Networks. In addition, we incorporate attention and context matcher mechanisms into the model architecture. Our Weather Model achieves significant performance improvements compared to baseline deep learning models, including ConvLSTM, TrajGRU, and U-Net. Our experimental evaluation involves high-scale, real-world benchmark numerical weather datasets, namely the ERA5 hourly dataset on pressure levels and WeatherBench. Our results demonstrate substantial improvements in identifying spatial and temporal correlations with attention matrices focusing on distinct parts of the input series to model atmospheric circulations. We also compare our model with high-resolution physical models using the benchmark metrics and show that our Weather Model is accurate and easy to interpret.

Kutalmis Gokalp Ince, Aybora Koksal, Arda Fazla et al.

We propose a semi-automatic bounding box annotation method for visual object tracking by utilizing temporal information with a tracking-by-detection approach. For detection, we use an off-the-shelf object detector which is trained iteratively with the annotations generated by the proposed method, and we perform object detection on each frame independently. We employ Multiple Hypothesis Tracking (MHT) to exploit temporal information and to reduce the number of false-positives which makes it possible to use lower objectness thresholds for detection to increase recall. The tracklets formed by MHT are evaluated by human operators to enlarge the training set. This novel incremental learning approach helps to perform annotation iteratively. The experiments performed on AUTH Multidrone Dataset reveal that the annotation workload can be reduced up to 96% by the proposed approach.