At Applied AI Letters, we noted that, in the field of AI, whilst there is an abundance of places that one can publicise the 'breakthrough' stage, it was much harder to find places to communicate impactful applications and innovations. Additionally, we felt that there was an abundance of potential for impact if we could start a venue in which people across disciplines could communicate their solutions to challenges posed in a wide variety of potential application areas.
We strongly believe that we improve the sustainability of long-term AI research through communicating and celebrating the evolution from idea to real world application. In order to achieve this, there are many additional considerations beyond the theoretical and methodological elements of successfully deploying and delivering an AI application in the real world, and we need a forum to discuss and communicate these successes with new potential users.
In order to deliver a venue which achieves these aims, at Applied AI Letters, we have three distinct types of articles:
1. Emerging applications of Artificial Intelligence - first in kind applications, or examples of early adopter technology, which demonstrate the power of applying advanced AI techniques to impactful problems. Examples of this kind of article could be the development of deep learning models for pharmaceutical applications, development and use of knowledge representation techniques for advanced querying of unstructured databases, or the application of reinforcement learning techniques for advanced control problems.
2. Development and deployment of Artificial Intelligence Systems - description of algorithmic and system engineering required to deploy Artificial Intelligence technologies in a ‘real world’ setting. Examples of this kind of article could be security considerations for AI deployed within a healthcare setting, scalability analysis of AI technologies (for either training or deployment) of deep learning models in a high-velocity data environment such as required by Industry 4.0 strategies, or ethical considerations (such as explainability and transparency) required when deploying AI solutions into application areas such as the law and insurance, or into technologies such as autonomous vehicles.
3. Comments and Challenges for Artificial Intelligence Technologies - these short perspective articles either summarise the current state of the art for a particular AI technology (such as deep learning), highlighting both strengths and challenge areas for future development, or summarise an emerging challenge area (such as genomics) in which application of AI technologies has the potential to initiate a step change in capability.
More Details On Our Aims and Scope
The Importance of understanding algorithms in the real world
There is a difference between the process associated with augmenting the state-of-the-art and the process of turning a prototype algorithm into a deployed system. It is widely acknowledged that invention of a new state-of-the-art algorithms is a creative process. We perceive that deploying algorithms in general, and specifically when those algorithms encode AI, often also demands creativity and is rarely the "simple software engineering" that some might think. Applied AI Letters is a venue for publicising those novel advances and gaining inspiration from others working at this interface between AI and the real world. While some of the focus of any paper will necessarily be specific to the application at hand, we are keen to provide a venue where the community can seek out trends that resonate generically across multiple potentially disparate specific applications.
We therefore welcome papers that explain how algorithms can be redesigned to accommodate constraints imposed by the pre-existing software, hardware and interfaces that pervade real-world systems. We do recognise the value of research that focuses on a simulated setting with clean interfaces to raw data and can be tackled using a state-of-the-art GPU cluster running the latest version of a compiler. However, our interest is in understanding how to achieve similar performance when adhering to pre-existing software interfaces and using legacy hardware, neither of which was designed with AI in mind.
Similarly, while we recognise the value of testing against reference datasets and using general-purpose well-established metrics to quantify performance (eg F-score and RMS errors), we are also interested in papers that describe experiments and/or trials of using AI in the real-world such that the resulting operational benefit of the AI is the metric for success.
We also recognise that adoption is also a key metric of success: while we might be able to show decision makers that they would rationally embrace AI, they don't necessarily all then choose to adopt (e.g. because of an irrational fear of change or simply a lack of belief in the rational argument presented). Applied AI Letters therefore warmly welcomes papers with a focus on how to modify algorithms and define performance metrics that are such that the chance of adoption is maximised.
Challenges in Deploying AI in the Real World
We are facing a global reproducibility crisis in science and research. Artificial intelligence is no different, in part due to the complexity and ever increasing scale of training and testing data, however, we still have many lessons to learn \cite{research}.
As a community, we need to tackle this dangerous lack of transparency head on with real world computational methods that can not only be reproducibly deployed, but also shown to work on real data, at scale.
A fundamental premise of Applied AI Letters is that we feature scientifically significant articles about the actual “application” of modern AI technologies. It is also our understanding that if these applications can’t also be deployed at scale, on modern elastic computational and data infrastructure to solve global sized challenges, then they also do not pass our “litmus test” for having actually been deployed at scale. It is critical that new methods can actually work and scale to run on global sized data challenges, scale must be “baked in” to any new AI algorithms.
Both application of scale and the transparency of methods will need to be identified and described fully in each and every Applied AI Letters article accepted for publication. Modern HPC centers have invested significantly in both training \cite{computing} and providing easy access to scale out computing at national centers or more often in partnership with many major hyperscalers to allow researchers to test and prove out their methods at scale on real world, actionable data.
Modern startups and our global technology industries have long understood that access to large scale in-silico infrastructure is the key business differentiator in many verticals. It is our hope that by combining the detailed work of the academy in concert with industrial scaling techniques and by also partnering on articles together, that Applied AI Letters will be able to further highlight the critical nature of technology transfer in artificial intelligence from universities and into industry, and will subsequently make for extremely compelling reading and solid reference material for everyone.
Example Impact Area: AI in Sciences
It goes without saying that there are a plethora of areas in which the application of AI can have a significant impact. In order to more clearly articulate the kinds of papers we are encouraging, we will take a deep dive on one such area - the application of AI to advance and accelerate scientific discovery.
Historically, the terms “discovery” and “serendipity” have been closely linked with many discoveries of important materials, chemical products or medicines. In many domains, the scientific discoveries are still associated with laborious and time-consuming “probe and error” experiments or the so-called Edisonian approach to innovation. The combination of big data and AI is often referred to as the fourth industrial revolution.\cite{httpwwweconomistcomnewsspecial-report21700761-after-many-false-starts-artificial-intelligence-has-taken-will-it-cause-mass} AI is revolutionizing topics across medicine, diagnostics, computer vision and language processing as they accumulate more and more data.
What are the general ways in which scientists apply AI methods in scientific settings?
- Prediction: Arguably, the simplest, most straightforward way to apply AI is to use it to tackle prediction problems: mapping data to predicted numerical outputs. This is where machine learning is typically used to extract complex patterns and correlations from these data.
- Understanding: Here the emphasis is of AI on understanding the underlying problem. In many cases an accurate prediction is not enough. Instead, we want to gain interpretable insights into what properties of the data or the process led to the observed outcome.
- Explainability: Explainable AI (XAI) will go even further, complementing predictive models with logical reasoning and explanations of their actions, to ensure that researchers are getting the right answer for the right reasons .
- Discovery: A cutting-edge frontier of research. Can AI find new laws and phenomena?
Some areas of science, have already started to realise some of those benefits. However, there still remains a large role for the applied AI and scientific communities to play to ensure that the best in class algorithms match up with the most impactful scientific problems of our time. Within this theme we feel that some particularly relevant areas for papers would include:
- The combination of machine learning and physics and emergence of physics-aware artificial intelligence (PAI)
- Explainability and uncertainty in AI systems and applications
- The use of machine learning to build powerful yet trustworthy models for prediction of complex phenomena
- The use of natural language processing and knowledge representation to extract scalable insights from the scientific literature
- The use of generative models to aid creativity and discovery
There is a large appetite in this area for the application of bleeding-edge methods to further push the boundaries of the possible, and we would strongly encourage papers which are the result of the collaboration of domain specialists and those whose major expertise is in the development of AI methodologies.
As we have stated throughout this editorial, a key metric of success is uptake beyond the initial method development. We therefore encourage labs, both academic and industrial, to submit papers demonstrating the real-world improvements which have been garnered through the application of AI to their mission-critical problems.
Our attitude to Publishing
Openness
Applied AI Letters is launched with a commitment to openness in the scientific and publishing process. Where possible and where appropriate, we want to give authors the ability to share their research as widely and as transparently as possible.
To that end, Applied AI Letters will be an open access journal, giving free and immediate access to critical research to all who want to read it. This greatly increases the reach and impact of the journal and ensures that our authors are able to share their work with the widest possible audience. Without the need to pay for a license, we are hopeful that our readership will include many in the industrial world who are often locked out of current research. We particularly envision those working in the diverse world of AI start-ups, who have often lacked a venue to prove their technology in a scientific manner, and to distinguish themselves from the unfortunate amount 'vapour ware' which infects this market.
Our commitment to openness extends beyond open access publishing of articles. We know that it will not be possible for all our authors to share their data. However, we do ask that authors publish a data availability statement with their paper to indicate whether the data associated with their article is freely available online.
A core offering of Applied AI Letters is the integrated preprint server. At the point of submission, authors are asked whether they would like to simultaneously upload their manuscript to the Applied AI Letters preprint collection. In doing so, authors are able to gain all the benefits of preprinting including a DOI and, in the event of acceptance, linking with the final published article.
The interpretation and implementation of openness in research is sure to change over time. As we begin our journey, we are constantly looking for the best ways to meet the needs of our authors and readers. We welcome feedback and ideas: we want to hear from our community. We do not seek to enforce a philosophy but rather to engage in an ongoing conversation.
Rapid Publication
Aligned with a future-looking mindset on openness, we also commit to offering authors a peer review and publication experience defined by ease and speed. As indicated by the title, the journal will publish primarily short-format letters. This decision was made to decrease the burden on authors, reviewers and readers. At every stage of the process, from writing to reviewing, revising, proofing and even reading, we hope that the shorter format will engender a more rapid turnaround.
Fostering Fairness and Diversity
We fundamentally believe that diversity promotes excellence and innovation. Our goal at Applied AI Letters will be to foster a culture of inclusion at all levels and in all interactions. As we set out on this mission, we cannot help but be aware of the lack of diversity in the author list of this inaugural editorial. We hope that our recognition of this fact functions as a first step towards realizing our mission. Moving forward, we will be reaching out to researchers engaged in the same movement to help us do more.
Conclusion
AI has the potential to revolutionise many areas of research. At Applied AI Letters, we aim to be a world-leading venue for the demonstration of, and dissusion about, the application of cutting edge AI technologies to the most impactful problems of today. We belive that openness, fairness and diversity are key to delivering on this promise, and want to foster a culture which embraces these key principles. With applied AI evolving at such a rapid rate, we appreciate that speed is important for the timely communication of your breakthroughs, but we also recognise that this must not come at the expense of rigourous peer review. Combining all of this may seem a lofty goal, but we believe that we, as both an editorial board and as a community of researchers, are up to the task. We look forward to receiving your papers for many years to come.