It is vital to evaluate the impact of AI on various stakeholders, including academic AI researchers, as the field is undergoing rapid transformation. A recent article by Togelius J. and Yannakakis G.N. titled “Choose Your Weapon: Survival Strategies for Depressed AI Academics” provides profound insight into this area.
The paper’s content explores the difficulties faced by those engaged in theoretical AI research in academic settings, despite the title’s playful narrative suggestion. The main ideas and conclusions of the study will be briefly summarised in this review.
Part 1: The Dilemmas AI Academics Face
1. Scarcity of Computing Resources:
The article underscores the increasing disparity in computing resources available to AI academics and their counterparts in corporate AI departments. A decade ago, local computational setups sufficed for advancing AI research in academia. However, the contemporary scenario has seen a paradigm shift. Significant advancements in AI today often rely on extensive computational power and a series of elaborate experiments. Unfortunately, many academic researchers find themselves without adequate access to such resources.
2. The Challenge of Corporate Dominance:
The concept of competition in the world of scientific research has intensified. Ideally, scientific experiments would represent collaborative endeavors, with due recognition to every contributor. Yet, the corporate realm’s increasing influence has somewhat eclipsed this cooperative spirit. When corporations channel substantial investments into AI research, they tend to dominate the development of promising ideas, often sidelining the original academic contributors. The paper draws a parallel between this situation and the phenomenon where a mega-retailer like Walmart establishes itself near a local family store, overshadowing its business.
The aforementioned challenges, as highlighted by Togelius and Yannakakis, depict a concerning landscape for AI academics. The conditions have led to a certain degree of disillusionment, impacting the morale and productivity of researchers who have dedicated their careers to furthering the field.
The study doesn’t merely identify problems; it also provides survival strategies for those in academia feeling the brunt of these challenges. A subsequent analysis below will delve deeper into the potential solutions proposed by the authors, aiming to offer AI academics tangible paths to navigate this evolving terrain.
Part 2: Strategies for Navigating Challenges
1. Opting for Alternative Publication Avenues:
Researchers are advised to consider publishing in less high-profile journals, focusing on refining technical aspects and exploring niche questions within broader topics.
2. Prioritizing Computing Resources:
An emphasis is placed on allocating a significant portion of research grants for computational resources. However, it’s noted that even substantial grants may not suffice for conducting advanced experiments on par with corporate endeavors.
3. Focusing on Smaller-Scale Experiments:
Researchers can center their efforts on more concise problems, using them to validate theoretical advancements. Several papers, such as those by Shafiullah et al. (2022) and Pearce et al. (2023), successfully employed this approach. Although these methods might initially receive limited attention, their relevance can grow once tested on larger datasets.
4. Leveraging Pretrained Models:
Instead of starting from scratch, using pretrained models can expedite the research process, though it might sometimes limit the depth of findings.
5. In-depth Analysis of Existing Models:
Researchers are encouraged to delve into the intricacies of current models rather than exclusively focusing on creating new ones.
6. Exploring Reinforcement Learning (RL):
RL is proposed as a valuable tool, especially since it doesn’t rely heavily on extensive data sets. However, it’s essential to balance ambition with feasibility.
7. Investigating Minimally Loaded Models:
The paper highlights the rising significance of drawing conclusions using minimally loaded models and a limited dataset, referencing Bayesian methods as an example.
8. Exploring Untapped or Neglected Areas:
Researchers could delve into subjects currently overlooked by the industry or revive previously abandoned methodologies. This approach may offer a window of opportunity before drawing significant attention.
9. Experimenting with Unexpected Methods:
Researchers are prompted to challenge the status quo by testing methods that seem counterintuitive.
10. Navigating Ethical Boundaries:
While corporations might be restricted by ethical guidelines and reputation considerations, academics have slightly more leeway. The authors suggest exploring topics that might be deemed controversial but underscore the importance of abiding by legal regulations.
11. Collaborating with the Industry:
Establishing partnerships with industry stakeholders could provide funding and potentially lead to the inception of start-ups. Yet, it’s essential for the research to align with practical applications.
12. Promoting Inter-University Collaborations:
Building bridges between universities can foster a collaborative environment, though the immediate benefits might appear elusive.
The strategies outlined by Togelius and Yannakakis (2023) represent a roadmap for AI academics navigating the current challenges. While the future of AI academia remains uncertain, these guidelines offer pathways to continue making meaningful contributions to the field. The subsequent articles in this series will further delve into the implications of these recommendations and their potential long-term impact.
Source: mPost
