In recent years, Automated Machine Learning (AutoML) has become increasingly important in Computer Science due to the valuable potential it offers. This is testified by the high number of works published in the academic field and the significant efforts made in the industrial sector. However, some problems still need to be resolved. In this paper, we review some Machine Learning (ML) models and methods proposed in the literature to analyze their strengths and weaknesses. Then, we propose their use—alone or in combination with other approaches—to provide possible valid AutoML solutions. We analyze those solutions from a theoretical point of view and evaluate them empirically on three Atari games from the Arcade Learning Environment. Our goal is to identify what, we believe, could be some promising ways to create truly effective AutoML frameworks, therefore able to replace the human expert as much as possible, thereby making easier the process of applying ML approaches to typical problems of specific domains. We hope that the findings of our study will provide useful insights for future research work in AutoML.

Summary Controlling cell fate has great potential for regenerative medicine, drug discovery, and basic research. Although transcription factors are able to promote cell reprogramming and transdifferentiation, methods based on their upregulation often show low efficiency. Small molecules that can facilitate conversion between cell types can ameliorate this problem working through safe, rapid, and reversible mechanisms. Here, we present DECCODE, an unbiased computational method for identification of such molecules based on transcriptional data. DECCODE matches a large collection of drug-induced profiles for drug treatments against a large dataset of primary cell transcriptional profiles to identify drugs that either alone or in combination enhance cell reprogramming and cell conversion. Extensive validation in the context of human induced pluripotent stem cells shows that DECCODE is able to prioritize drugs and drug combinations enhancing cell reprogramming. We also provide predictions for cell conversion with single drugs and drug combinations for 145 different cell types.

Controlling cell fate has great potential for regenerative medicine, drug discovery, and basic research. Although numerous transcription factors have been discovered that are able to promote cell reprogramming and trans-differentiation, methods based on their up-regulation tend to show low efficiency. The identification of small molecules that can facilitate conversion between cell types can ameliorate this problem working through safe, rapid, and reversible mechanisms. Here we present DECCODE, an unbiased computational method for the identification of such molecules solely based on transcriptional data. DECCODE matches the largest available collection of druginduced profiles (the LINCS database) for drug treatments against the largest publicly available dataset of primary cell transcriptional profiles (FANTOM5), to identify drugs that either alone or in combination enhance cell reprogramming and cell conversion. Extensive in silico and in vitro validation of DECCODE in the context of human induced pluripotent stem cells (hIPSCs) generation shows that the method is able to prioritize drugs enhancing cell reprogramming. We also generated predictions for cell conversion with single drugs and drug combinations for 145 different cell types and made them available for further studies.

The State of the Art of the young field of Automated Machine Learning (AutoML) is held by the connectionist approach. Several techniques of such an inspiration have recently shown promising results in automatically designing neural network architectures. However, apart from back-propagation, only a few applications of other learning techniques are used for these purposes. The back-propagation process takes advantage of specific optimization techniques that are best suited to specific application domains (e.g., Computer Vision and Natural Language Processing). Hence, the need for a more general learning approach, namely, a basic algorithm able to make inference in different contexts with distinct properties. In this paper, we deal with the problem from a scientific and epistemological point of view. We believe that this is needed to fully understand the mechanisms and dynamics underlying human learning. To this aim, we define some elementary inference operations and show how modern architectures can be built by a combination of those elementary methods. We analyze each method in different settings and find the best-suited application context for each learning algorithm. Furthermore, we discuss experimental findings and compare them with human learning. The discrepancy is particularly evident between supervised and unsupervised learning. Then, we determine which elementary learning rules are best suited for unsupervised systems, and, finally, we propose some improvements in reinforcement learning architectures.