The divide between classical machine learning and deep learning is not a sign that foundational methods are dead, but that conference review culture has become dangerously narrow. When an Area Chair dismisses original work on SVM or logistic regression as a "done and dusted topic," despite reviewers acknowledging innovation, that is not a verdict on the research. It is a failure of the review system to recognize that solid, interpretable methods remain the backbone of production systems and deserve rigorous attention.
This researcher's experience echoes a pattern we see across the field. While deep learning and LLMs dominate headlines, the practical world of data science still relies heavily on classical approaches for everything from credit scoring to medical diagnosis. Consider how Pinterest Fine-Tunes a 4B Language Model to Remap User Journeys represents a massive investment in a single, narrow technique, while the underlying infrastructure that powers most of the web's recommendation systems still leans on logistic regression and SVMs. The irony is stark: conference program committees chase novelty in architecture while ignoring novelty in theory. The fact that only ACL asked for a resubmission with revisions suggests that some venues still value rigorous, well-executed work regardless of the method's age. That is the exception, not the rule.
The practical takeaway for researchers is to stop treating top conferences as the only valid destination for original work. Journals have long been the natural home for theoretical contributions in classical ML, offering space for thorough exposition and multiple rounds of review that conference deadlines simply cannot accommodate. ArXiv, meanwhile, provides immediate dissemination and a permanent citation record. The story of From rejections to interview: one researcher's path to Amazon's applied science team reminds us that industry values real, reproducible contributions over conference acceptance rates. If your ablation studies and experiments satisfy the reviewers who actually read your work, you have already passed the most meaningful test.
The deeper issue is that conference review culture has become a popularity contest for subfields, not a meritocracy for ideas. When reviewers acknowledge innovation and an AC overrides them based on a superficial heuristic about topic age, the system is broken at the decision layer. Researchers should not stop doing classical ML research. They should stop expecting conference ACs to recognize it. Submit to journals, post on ArXiv, and let the community judge your work by its substance, not by its conference badge. The question is not whether classical ML is dead, but whether top conferences have made themselves irrelevant to the researchers who do the foundational work that everything else builds on.