Does anyone else miss the old conference ecosystem? [D]

The recent Reddit post lamenting the decline of specialized AI conferences – BMVC, ACCV, FG, ICIP, and ICASSP – resonates deeply with many in the field. It's a familiar feeling: the fragmentation of a once vibrant ecosystem into a few increasingly monolithic events. The author’s concern about the potential for good research to be lost amidst exploding submission numbers and inconsistent review processes is valid, and echoes sentiments shared in discussions surrounding the high costs of attending these events, as highlighted in Why is ECCV so insanely expensive for students presenting papers?. The shift towards a handful of flagship conferences, while seemingly efficient on the surface, risks homogenizing research directions and marginalizing specialized communities. We’ve seen similar concerns arise regarding the prioritization of certain methodologies, as evidenced in discussions around optimal training parameters for smaller datasets, which some have argued are often overlooked – a point underscored by the debate surrounding qlora defaults, detailed in The qlora 2e-4 default is wrong under 10k samples and nobody talks about it.

The loss of focused communities is perhaps the most significant consequence. Each of the conferences mentioned fostered a unique environment where researchers in specific domains could deeply engage with each other’s work, receive targeted feedback, and forge lasting collaborations. FG, for instance, was *the* place to be for face analysis, ICASSP dominated signal processing, and BMVC/ACCV consistently showcased rigorous research. This concentrated expertise is difficult to replicate at larger, more generalized conferences, where presentations can easily get lost in the shuffle. The current trend towards prioritizing high-profile venues creates a pressure to conform to broader trends, potentially stifling innovation in niche areas. The resulting impact on journal quality, as some have questioned regarding TACL, also speaks to the broader challenges of maintaining rigorous standards within an increasingly concentrated academic landscape TACL journal doubts.

This concentration isn’t solely driven by a desire for efficiency. The economics of academic conferences play a significant role. Hosting large, prestigious events attracts sponsors, generates revenue, and enhances institutional reputations. The pressure to maximize attendance and impact inevitably leads to a consolidation of resources and a prioritization of flagship conferences. Furthermore, the sheer volume of research being produced in AI necessitates some form of curation, but the current system seems to be inadvertently favoring breadth over depth. The rise of arXiv as a primary dissemination channel, while democratizing access to research, also contributes to the problem. Without the filtering and validation of a peer-reviewed conference presentation, it becomes even more challenging to identify truly impactful work. This creates a cycle where researchers feel compelled to publish on arXiv to maintain visibility, further contributing to the overload of information.

Looking ahead, the challenge lies in finding ways to revitalize specialized communities and ensure that valuable research isn't lost in the shuffle. This might involve fostering smaller, more focused workshops and satellite events alongside the major conferences. It could also mean exploring alternative publishing models that prioritize quality and peer review over sheer volume. Ultimately, a more decentralized and diverse research ecosystem, one that values both breadth and depth, will be crucial for sustaining the long-term vitality of AI research. The question we should be asking is not just *how* to manage the explosion of research, but *how* to ensure that the most important discoveries – those that may not fit neatly into the dominant narratives – are brought to light and have the opportunity to shape the future.