From the recent progress in this field and the models we use in developing softwares 90% are traditional ML models and we hardly use deep learning to solve real problems. moreover, the work we do it's not different from a data analyst. Sometimes it feels like bullshit work with the sole aim of increasing accuracy by delta amount without even having real impact in the product. I agree there are some extreme cases where the products have integrated with ML in good use cases but I am talking about the general sentiment and use case of the tech.
If by increasing accuracy by delta amount you mean the accuracy improvement is insignificant, for plenty of problems that's not true at all. Speech, vision, NLP rely heavily on deep learning, you won't get close to the best results without it (or if you do you should publish)
One is the use case they portray speech and visions have. But is there a significant need for all of that? Agreed that without these model we wouldn't have good recommendation and search. But given the investment that is currently made and the technology that is already present is there a need for such?
We're in the infancy of AI/ML, it's going to get much more interesting between now and 2030
Compared to what will be possible, it's still very early days.
what will be possible?
It’s early, but we will soon slip into the Trough of Disillusionment once people realize that we’re still quite a ways off from what we’ve tried to sell people.
One could argue that the trough already happened in the 30 years after neural networks first got hyped
ML is underused, but that could also be what makes a bubble be a bubble: not capitalizing on opportunities even though they're there. Two things come to mind that make good ML hard in practice. First, data acquisition and access protocols in large companies are in their infancy despite data now being a hot commodity. Second, the majority of people who work on ML in products aren't the applied scientists, but are regular engineers who took Ng's class and know little more stats than basic probability.
Then comes the issue of privacy, models being unbiased. These are limitations of ML models. How do you solve that?
Those aren't much of a concern for most ML applications. Not enough for them to create a serious enough obstacle to claim a bubble
If your mom hears about AI/ML, it is a bubble.
AI/ML just started and will only continue to accelerate. We are in the baby steps phase and low on the S curve in this tech. Once the algs develop and can take disparate data sets with no apparent relation to each other and start to churn out useful predictive data showing just how interrelated they have been all along, it’s going to become even more interesting.
Thanks so much for your insight, Butthugs. I wonder in Facebook, where ML is being used? Maybe in ads relevance?
Going to increase for sure.