Deep learning has been the answer to many machine learning problems during the past two decades. However, it comes with two significant constraints: dependency on extensive labeled data and training costs. Transfer learning in deep learning, known as Deep Transfer Learning (DTL), attempts to reduce such reliance and costs by reusing obtained knowledge from a source data/task in training on a target data/task. Most applied DTL techniques are network/model-based approaches. These methods reduce the dependency of deep learning models on extensive training data and drastically decrease training costs. Moreover, the training cost reduction makes DTL viable on edge devices with limited resources. Like any new advancement, DTL methods have their own limitations, and a successful transfer depends on specific adjustments and strategies for different scenarios. This paper reviews the concept, definition, and taxonomy of deep transfer learning and well-known methods. It investigates the DTL approaches by reviewing applied DTL techniques in the past five years and a couple of experimental analyses of DTLs to discover the best practice for using DTL in different scenarios. Moreover, the limitations of DTLs (catastrophic forgetting dilemma and overly biased pre-trained models) are discussed, along with possible solutions and research trends.

This study provides a comprehensive survey of the top-performing research papers in the field of time series prediction, offering insights into the most effective machine learning techniques, including tree-based, deep learning, and hybrid methods. It explores key factors influencing the model performance, such as the type of time series task, dataset size, and the time interval of historical data. Additionally, this study investigates potential biases in model development and weighs the trade-offs between the computational costs and performance. A detailed analysis of the most used error metrics and hyperparameter tuning methods in the reviewed papers is included. Furthermore, this study evaluates the results from prominent forecasting competitions, such as M5 and M6, to enrich the analysis. The findings of this paper highlight that tree-based methods like LightGBM 4.6.0 and deep learning methods like recurrent neural networks deliver the best performance in time series forecasting, with tree-based methods offering a significant advantage in terms of their computational efficiency. This paper concludes with practical recommendations for approaching time series forecasting tasks, offering valuable insights and actionable strategies that can enhance the accuracy and reliability of predictions derived from time series data.

Signaling proteins such as protein kinases adopt a diverse array of conformations to respond to regulatory signals in signaling pathways. Perhaps the most fundamental conformational change of a kinase is the transition between active and inactive states, and defining the conformational features associated with kinase activation is critical for selectively targeting abnormally regulated kinases in diseases. While manual examination of crystal structures have led to the identification of key structural features associated with kinase activation, the large number of kinase crystal structures (~3,500) and extensive conformational diversity displayed by the protein kinase superfamily poses unique challenges in fully defining the conformational features associated with kinase activation. Although some computational approaches have been proposed, they are typically based on a small subset of crystal structures using measurements biased towards the active site geometry. We utilize an unbiased informatics based machine learning approach to classify all eukaryotic protein kinase conformations deposited in the PDB. We show that the orientation of the activation segment, measured by φ, ψ, χ1, and pseudo-dihedral angles more accurately classify kinase crystal conformations than existing methods. We show that the formation of the K-E salt bridge is statistically dependent upon the activation segment orientation and identify evolutionary differences between the activation segment conformation of tyrosine and serine/threonine kinases. We provide evidence that our method can identify conformational changes associated with the binding of allosteric regulatory proteins, and show that the greatest variation in inactive structures comes from kinase group and family specific side chain orientations. We have provided the first comprehensive machine learning based classification of protein kinase active/inactive conformations, taking into account more structures and measurements than any previous classification effort. Further, our unbiased classification of inactive structures reveals residues associated with kinase functional specificity. To enable classification of new crystal structures, we have made our classifier publicly accessible through a stand-alone program housed at https://github.com/esbg/kinconform [DOI: 10.5281/zenodo.249090 ].

Predicting alfalfa biomass and crop yield for livestock feed is important to the daily lives of virtually everyone, and many features of data from this domain combined with corresponding weather data can be used to train machine learning models for yield prediction. In this work, we used yield data of different alfalfa varieties from multiple years in Kentucky and Georgia, and we compared the impact of different feature selection methods on machine learning (ML) models trained to predict alfalfa yield. Linear regression, regression trees, support vector machines, neural networks, Bayesian regression, and nearest neighbors were all developed with cross validation. The features used included weather data, historical yield data, and the sown date. The feature selection methods that were compared included a correlation-based method, the ReliefF method, and a wrapper method. We found that the best method was the correlation-based method, and the feature set it found consisted of the Julian day of the harvest, the number of days between the sown and harvest dates, cumulative solar radiation since the previous harvest, and cumulative rainfall since the previous harvest. Using these features, the k-nearest neighbor and random forest methods achieved an average R value over 0.95, and average mean absolute error less than 200 lbs./acre. Our top R² of 0.90 beats a previous work’s best R² of 0.87. Our primary contribution is the demonstration that ML, with feature selection, shows promise in predicting crop yields even on simple datasets with a handful of features, and that reporting accuracies in R and R² offers an intuitive way to compare results among various crops.

The total boll count from a plant is one of the most important phenotypic traits for cotton breeding and is also an important factor for growers to estimate the final yield. With the recent advances in deep learning, many supervised learning approaches have been implemented to perform phenotypic trait measurement from images for various crops, but few studies have been conducted to count cotton bolls from field images. Supervised learning models require a vast number of annotated images for training, which has become a bottleneck for machine learning model development. The goal of this study is to develop both fully supervised and weakly supervised deep learning models to segment and count cotton bolls from proximal imagery. A total of 290 RGB images of cotton plants from both potted (indoor and outdoor) and in-field settings were taken by consumer-grade cameras and the raw images were divided into 4350 image tiles for further model training and testing. Two supervised models (Mask R-CNN and S-Count) and two weakly supervised approaches (WS-Count and CountSeg) were compared in terms of boll count accuracy and annotation costs. The results revealed that the weakly supervised counting approaches performed well with RMSE values of 1.826 and 1.284 for WS-Count and CountSeg, respectively, whereas the fully supervised models achieve RMSE values of 1.181 and 1.175 for S-Count and Mask R-CNN, respectively, when the number of bolls in an image patch is less than 10. In terms of data annotation costs, the weakly supervised approaches were at least 10 times more cost efficient than the supervised approach for boll counting. In the future, the deep learning models developed in this study can be extended to other plant organs, such as main stalks, nodes, and primary and secondary branches. Both the supervised and weakly supervised deep learning models for boll counting with low-cost RGB images can be used by cotton breeders, physiologists, and growers alike to improve crop breeding and yield estimation.