Evolving AI from Research to Real Life - Some Challenges and Suggestions

Sandya Mannarswamy, Shourya RoyIJCAI 2018 · DOI

TL;DR

This paper argues that while AI has made significant strides, its transition from research to real-world applications faces critical challenges like explainability and fragility. The authors propose leveraging learnings from psychology and medicine to develop more robust, trustworthy, and deployable AI systems.

What Problem Does This Paper Address?

The paper addresses the significant gap between AI research and its successful, widespread deployment in real-life applications. While AI has moved beyond science fiction to impact areas like autonomous vehicles and medical diagnostics, the authors contend that several inherent challenges hinder its full integration into society. These challenges include:

  • Explainability: AI decisions, especially in critical domains like loan applications or medical diagnoses, need to be understandable and justifiable to users.
  • Fragility: AI systems can be surprisingly brittle, failing in unexpected ways when presented with minor variations in input data, leading to potentially disastrous consequences.
  • Research Practices: Issues like poor reproducibility, publication bias towards positive results, and the race for state-of-the-art numbers can sacrifice scientific rigor.
  • Limited Generalization: Current AI models often struggle to transfer knowledge across different tasks or domains, limiting their applicability.

The paper aims to highlight these issues and suggest practical ways the AI community can learn from other disciplines to overcome them, ensuring AI becomes a true "real-life presence" rather than just a "cool technology."

Methods

This paper is primarily a position paper and a call to action, rather than a technical paper detailing novel algorithms or experiments. The authors' "method" involves:

  • Identifying and Categorizing Challenges: They meticulously outline and discuss key challenges AI faces in its evolution from research to real life, categorizing them into AI-specific issues (explainability, fragility, limited generalization) and common research/deployment issues (research practices, deployment hurdles).
  • Drawing Analogies and Leveraging Cross-Disciplinary Learnings: A core aspect of their approach is to look at how other scientific disciplines, particularly psychology and medicine, have tackled similar problems during their own evolution. They highlight specific mechanisms like "Registered Reports" in psychology and "translational medicine" principles as valuable lessons for AI.
  • Proposing Solutions and Frameworks: Based on these cross-disciplinary learnings, they suggest concrete steps for the AI community, such as forming a "translational AI workforce," adopting phased deployment strategies, establishing federated advisory boards, and developing AI-specific "Health Technology Assessments."

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • The "Dickensian" Duality of AI: The paper opens by likening the current state of AI to Charles Dickens' "It was the best of times, it was the worst of times," reflecting the simultaneous excitement and apprehension surrounding AI's rapid progress and potential impact.
  • The "AI Winter" Threat: The authors warn that if current challenges aren't addressed, AI could face another "AI winter," similar to past periods of reduced funding and interest.
  • Explainability is Crucial for Trust: For AI to be accepted in high-stakes applications, its decisions must be transparent and justifiable. The paper emphasizes that simply achieving high accuracy isn't enough; understanding why a decision was made is paramount.
  • Fragility Undermines Real-World Reliability: The brittleness of current AI systems, where minor input changes can lead to drastic errors, poses a significant risk for safety-critical applications.
  • The "Publish or Perish" Culture Hinders Reproducibility: The pressure to publish novel results can lead to rushed work, insufficient documentation, and a lack of incentive for replication, creating a "replication crisis" in AI.
  • Learning from "Bench to Bedside": The authors strongly advocate for emulating the "translational medicine" approach, which emphasizes carefully controlled deployment and staged translation of research findings into clinical practice. This principle is seen as vital for mitigating risks in AI deployment.
  • AI Can Be Part of the Solution: The paper suggests that AI itself can be used to address some of its own challenges, such as developing tools to automate code verification and improve replication.

Key Results

As a position paper, this work does not present quantitative experimental results. Instead, its "results" are the insights and proposed solutions:

  • Identification of Key Challenges: The paper clearly articulates and categorizes the primary hurdles AI faces in transitioning from research to real-world applications.
  • Proposed Cross-Disciplinary Solutions: The authors offer a framework for AI to learn from psychology and medicine, suggesting specific mechanisms like "Registered Reports" and "phased deployment" as actionable steps.
  • Call for Community Action: The paper serves as a catalyst, urging the AI community to engage in self-reflection, collaboration, and the development of new practices and guidelines for responsible AI development and deployment.

Limitations and Future Work

Limitations:

  • Not Exhaustive: The authors acknowledge that their list of challenges is not exhaustive and that other issues, such as safety, ethical considerations, and cultural impacts, also require attention.
  • Generalizability of Solutions: While drawing parallels, the authors recognize that the translational cycle for AI might differ from medicine, and some phases may need omission or refinement.
  • Potential for Stifling Innovation: The paper acknowledges the need to devise control and regulation measures without stifling innovation and enabling faster translation.

Future Work:

  • Developing Translational AI Workforce: The paper calls for the creation of specialized teams focused on translating AI research into real-world applications.
  • Establishing Federated Deployment Advisory Boards: The authors suggest forming interdisciplinary boards to review large-scale AI deployments.
  • Implementing Comprehensive AI Technology Assessments: Similar to Health Technology Assessments, these would evaluate AI applications before deployment.
  • Active Inclusion of End Consumers: AI development should actively incorporate feedback from its end-users.
  • Creating an "AI Cochrane Report": The paper suggests a mechanism for evidence-based translation of AI research into practice.
  • Developing Mechanisms for Controlled Release: Tailored phased trial approaches are needed for different AI application domains.
  • AI-Driven Verification Tools: The paper highlights the need for AI tools to help verify code and experimental results, addressing the reproducibility crisis.

Summary updated

BibTeX
@inproceedings{mannarswamy2018evolving,
  author = {Sandya Mannarswamy and Shourya Roy},
  title = {{Evolving AI from Research to Real Life - Some Challenges and Suggestions}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/717},
}
Figure