Implicit Bias and Interpretability in Artificial Neural Networks

Dear Colleagues,

Despite the considerable recent progress and achievements of deep learning architectures in a variety of fields, the reasons underlying their effectiveness remain opaque. State of art artificial neural networks algorithms are neither  interpretable nor explainable in terms of human features or computations. Moreover, simply by choosing different optimization algorithms, initialization schemes or architectures, we can dramatically change the set of features learned by the model, a phenomenon known as implicit bias.

The general purpose of this Special Issue is twofold:

1. From a mathematical perspective, we encourage papers that deal with mathematically grounded notions of interpretability or explainability. Of particular interest will be those that connect interpretable features with implicit bias.  
2. From the perspective of applications, we are also interested in work that introduces novel interpretable deep models or explicitly elucidates the non-interpretability of existing successful models with new techniques.

Specific topics of interest include (but are not limited to) the following:

• Mathematically precise notions of interpretability or explainability
• Designing and initializing networks with sets of features that incorporate prior-knowledge (e.g. Physics informed networks)
• Feature importance scores for neural networks
• Rigorous statistical hypothesis testing for neural network regression and classification models
• Compositionality and interpretable features
• Characterizing and Detecting shortcut features
• Theory and Algorithms for inferring properties of the inductive bias of state-of-the-art feedforward neural network architectures, initializations and learning algorithms
• Characterizing and Interpreting the inductive bias of (simple) recurrent neural networks to build foundations
• Characterizing and Interpreting the Inductive Bias of non-gradient-based learning algorithms e.g. evolutionary learning algorithms or global dynamic programming based algorithms
• Domain knowledge captured in the form of coarse-grained causal generative models.

Dr. Ankit B. Patel
Dr. Fabio Anselmi
Guest Editors

Manuscript Submission Information

Manuscripts should be submitted online at www.mdpi.com by registering and logging in to this website. Once you are registered, click here to go to the submission form. Manuscripts can be submitted until the deadline. All submissions that pass pre-check are peer-reviewed. Accepted papers will be published continuously in the journal (as soon as accepted) and will be listed together on the special issue website. Research articles, review articles as well as short communications are invited. For planned papers, a title and short abstract (about 100 words) can be sent to the Editorial Office for announcement on this website.

Submitted manuscripts should not have been published previously, nor be under consideration for publication elsewhere (except conference proceedings papers). All manuscripts are thoroughly refereed through a single-blind peer-review process. A guide for authors and other relevant information for submission of manuscripts is available on the Instructions for Authors page. Mathematics is an international peer-reviewed open access semimonthly journal published by MDPI.

Please visit the Instructions for Authors page before submitting a manuscript. The Article Processing Charge (APC) for publication in this open access journal is 2600 CHF (Swiss Francs). Submitted papers should be well formatted and use good English. Authors may use MDPI's English editing service prior to publication or during author revisions.

• artificial neural networks
• interpretability
• implicit bias
• inductive bias