Things I Learnt from fast.ai v3(towardsdatascience.com)
towardsdatascience.com
Things I Learnt from fast.ai v3
https://towardsdatascience.com/10-new-things-i-learnt-from-fast-ai-v3-4d79c1f07e33
7 comments
As someone who learned ML the "traditional way" I both like and dislike it.
I like that I can quickly learn about advanced models since I already know about / can infer the intuition / mathematical details myself most of the time.
I dislike it because it's almost all handwaving. The way things actually work aren't really explained (other than at a high level) so either you're stuck with a wishy washy understanding or you will read the paper yourself.
But for some of us "in between", we're not satisfied with the layman's explanation, yet the paper is too formal to digest. I think there's an element of survivorship bias there, where some people just give up because they're not getting the explanations they're looking for.
Those who are smart enough to understand papers on their own can do fine, because they'll take the course as-is and just use it as a guide for what's new and fresh. Those who don't know what the heck is going on (and don't care) are happy with "plain English" explanations without looking any further.
I like that I can quickly learn about advanced models since I already know about / can infer the intuition / mathematical details myself most of the time.
I dislike it because it's almost all handwaving. The way things actually work aren't really explained (other than at a high level) so either you're stuck with a wishy washy understanding or you will read the paper yourself.
But for some of us "in between", we're not satisfied with the layman's explanation, yet the paper is too formal to digest. I think there's an element of survivorship bias there, where some people just give up because they're not getting the explanations they're looking for.
Those who are smart enough to understand papers on their own can do fine, because they'll take the course as-is and just use it as a guide for what's new and fresh. Those who don't know what the heck is going on (and don't care) are happy with "plain English" explanations without looking any further.
I started the course last weekend and I really appreciate it so far. I don’t have a background in it, I don’t have a college degree, I didn’t even make it very far in high school math! But I’m quite competent with code and I learn best by doing things, then tinkering with them and diving deep to understand how they work so I can do it better. Fast.ai’s course seems to be perfectly tailored for someone like me. I’m sure that the explanations are fascinating but it’s completely alien to me, I need a friendlier entry point. If I can become comfortable with the material in their course, I’m confident that I’ll be more motivated to and capable of understanding at least some of the underlying science. It seems likely that there are many people like me out there.
Think of it this way - the fast.ai curriculum is intended for training electricians. People interested in physics may consider some of the simplifications similarly handwavy, and they'd be correct, but it misses the larger context.
You know, I know web devs who went to university and bootcamps and understand what's happening at every level of the application as it runs, and they don't seem much better
off than the tweens I knew who obsessed over their Neopets fansites. The only real differential I've noticed between learners is movitavion, and I think fast ai provides early successes and fun that lead to motivation to dig deeper.
No you are not, a "from scratch" approach works better even for ML in my opinion.
But if you don't know anything about it, fastai is a good generic introduction to basic concept and jargon (that you'll learn somewhere else).
I like the fastai pedagogical approach. It follows the 'tutorial style', where everyone should be able to make it to the end, and you should get a reward to motivate you to continue.
They can't do everything, but what is most obviously missing is the "80% of data science" - data preparation/feature-engineering. The data scientists we encounter that are self-trained with courses like fastai are mostly wholly unprepared for dealing with enterprise data platforms, and, in particular, engineering data at scale.
They can't do everything, but what is most obviously missing is the "80% of data science" - data preparation/feature-engineering. The data scientists we encounter that are self-trained with courses like fastai are mostly wholly unprepared for dealing with enterprise data platforms, and, in particular, engineering data at scale.
At times it's just a little bit too much handwaving for my taste.