In machine learning research, there's basically zero chance of ever getting a paper published where your experiments failed. It doesn't matter if the thing you're testing makes lots of sense and is well motivated and is based on existing effective techniques. In other words, it doesn't matter how obvious the experiments are as a next step. If it doesn't lead to improvements somewhere (time, cost, accuracy, robustness, etc.), then it's almost certainly unpublishable. That's not the case in other areas of science. I went to an ethology conference and was surprised at how many people presented well-motivated experiments where the hypothesis was wrong. During these talks, the general vibe in the room was "ooo, interesting 🤔". But then ethology is a science. Machine learning isn't really science, it's engineering. It's an arms race between all the researchers of the world. The rewards in money and fame go to whoever has the highest numbers on the benchmarks. And none of the journals want to publish work that doesn't continue to juice those numbers. There are clear downsides to this. If an idea is obvious but doesn't lead to improvements on any benchmarks then you get lots of researchers doing the same work over and over, with no one able to get the word out. It also encourages people to test their techniques on weirdo metrics and benchmarks where, for some quirk of statistics, their technique is "state of the art". A culture of clickbait seeps into the literature. And of course almost all of your experiments will fail. That's just how it goes. If we already knew what worked, there would be no need for research. Which means almost all of your work is unpublishable. But if I'm being honest, if journals did publish experiments that failed, I probably wouldn't read them. So maybe it's a good thing that the field is organized this way. It keeps poorly designed, poorly motivated distractions from flooding the net. Signal to noise ratios and all that. It's also probably good for new researchers because it makes the goals clearer. If someone were to say "Ethology has advanced a lot in the last 20 years," we would expect they meant "we have a richer and more nuanced understanding of animal behaviour". But if someone were to say "Machine learning has advanced a lot in the last 20 years," we wouldn't expect the person to mean "we have furthered our understanding of high dimensional manifolds". We would assume they meant "we have developed new techniques that are faster, cheaper, and more accurate". A machine learning journal publishing negative results would be like an ethology journal publishing papers on "all the things I don't know about penguins". Who cares? Still, there must be a place to record all the stuff that didn't work. Maybe blogposts are the best we've got.