3 ms·

This piece makes assumptions that turn this into a very boring discussion with little depth. How effective is preliminary learning vs on-demand learning? If I

This piece makes assumptions that turn this into a very boring discussion with little depth. How effective is preliminary learning vs on-demand learning? If I can have 80% of the conceptual understanding in 1/10 of the time is that a good trade-off? For preliminary learning, how do you know which knowledge will be useful in the future and which will be abstracted over? How fast is the field moving and what do you want to focus on? For example, someone asked on the ML subreddit about theoretical ML research and many comments said theoretical research was unnecessary: "But, regarding the value of pure ML theory research e.g. convergence bounds, versus practical ML research e.g. quantization methods, my personal feeling for quite some time has been that purely theoretical ML research has been predominantly bunk. Machine Learning is so high dimensional that things that can't be proven universal can be nearly guaranteed probabilistically, and things that can be shown to be possible can be staggeringly unlikely; for example, just because the No Free Lunch theorem exists, doesn't mean that Adam won't work in the vast, vast majority of cases." https://www.reddit.com/r/MachineLearning/comments/101qbfl/d_life_advice_to_relatively_late_bloomer_ml/j2rplc1/ https://www.reddit.com/r/MachineLearning/comments/101qbfl/d_... This topic is extremely grey, and blunt arguments like "on-demand learning is ruining the tech industry" are not interesting. I recommend you watch this talk by Laurie Voss where he (They?) explains that fundamentals are always shifting. What you consider "conceptual understanding" now was a tool many years ago. Tools solidify into the new generation of concepts over time. One controversial point he makes is that at some point, new developers will not know HTML and that's ok! https://twitter.com/seldo/status/1075031181685592064 https://twitter.com/seldo/status/1075031181685592064 https://www.youtube.com/live/hWjT_OOBdOc?feature=share https://www.youtube.com/live/hWjT_OOBdOc?feature=share

psyklic 4y ago
Mature fields ideally are driven by theory. Newer fields are driven by experiment, until the theory catches up. I've learned that it is a waste of time (and brainpower) to blindly tweak code without understanding the algorithm, hoping it will magically work. Yet, I've observed many newcomers to machine learning essentially doing this. It might be somewhat useful at this moment, but if it is easy to do then it will likely become automated.

psyklic 4y ago

Mature fields ideally are driven by theory. Newer fields are driven by experiment, until the theory catches up. I've learned that it is a waste of time (and brainpower) to blindly tweak code without understanding the algorithm, hoping it will magically work. Yet, I've observed many newcomers to machine learning essentially doing this. It might be somewhat useful at this moment, but if it is easy to do then it will likely become automated.