Deep Learning: Advice on Getting Started with fast.ai - Jeremy Howard | AI Podcast Clips
00:00:00.000 | [MUSIC]
00:00:07.960 | >> What advice do you have for
00:00:11.040 | someone who wants to get started in deep learning?
00:00:15.040 | >> Train lots of models.
00:00:17.240 | That's how you learn it.
00:00:20.280 | So I think, it's not just me,
00:00:24.760 | I think our course is very good,
00:00:26.520 | but also lots of people independently,
00:00:27.960 | is that it's very good.
00:00:28.760 | It recently won the COGX award for
00:00:30.920 | AI courses as being the best in the world.
00:00:33.160 | I'd say come to our course, course.fast.ai.
00:00:36.040 | The thing I keep on hopping on in my lessons is,
00:00:39.040 | train models, print out the inputs to the models,
00:00:42.320 | print out to the outputs to the models,
00:00:44.440 | study, change the inputs a bit,
00:00:48.280 | look at how the outputs vary,
00:00:50.360 | just run lots of experiments to get
00:00:52.080 | an intuitive understanding of what's going on.
00:00:58.120 | >> To get hooked, you mentioned training,
00:01:02.120 | do you think just running the models inference?
00:01:06.000 | If we talk about getting started.
00:01:08.560 | >> No, you've got to fine-tune the models.
00:01:10.600 | So that's the critical thing,
00:01:12.680 | because at that point, you now have
00:01:14.040 | a model that's in your domain area.
00:01:16.320 | So there's no point running
00:01:18.800 | somebody else's model because it's not your model.
00:01:21.280 | It only takes five minutes to fine-tune
00:01:23.280 | a model for the data you care about.
00:01:25.120 | In lesson two of the course,
00:01:26.720 | we teach you how to create your own dataset from
00:01:29.000 | scratch by scripting Google Image Search.
00:01:31.960 | We show you how to actually
00:01:34.000 | create a web application running online.
00:01:36.040 | So I create one in the course that
00:01:37.680 | differentiates between a teddy bear,
00:01:39.440 | a grizzly bear, and a brown bear.
00:01:41.200 | It does it with basically 100 percent accuracy.
00:01:44.040 | Took me about four minutes to scrape
00:01:45.960 | the images from Google Search in the script.
00:01:48.040 | There's a little graphical widgets we
00:01:50.960 | have in the notebook that help you clean up the dataset.
00:01:54.280 | There's other widgets that help you study
00:01:56.680 | the results to see where the errors are happening.
00:01:59.440 | So now we've got over 1,000 replies in our share
00:02:03.280 | your work here thread of students saying,
00:02:05.880 | here's the thing I built.
00:02:07.440 | So there's people who like,
00:02:08.960 | and a lot of them are state of the art.
00:02:10.720 | Like somebody said, "Oh, I tried looking at
00:02:12.280 | Devan Garey characters and I couldn't believe it.
00:02:14.320 | The thing that came out was more accurate than
00:02:16.640 | the best academic paper after lesson one."
00:02:19.600 | Then there's others which are just more fun,
00:02:21.720 | like somebody who's doing Trinidad and Tobago hummingbirds.
00:02:26.280 | She said that's their national bird and she's got
00:02:28.800 | something that can now classify
00:02:30.160 | Trinidad and Tobago hummingbirds.
00:02:31.880 | So yeah, train models,
00:02:33.440 | fine-tune models with your dataset,
00:02:35.640 | and then study their inputs and outputs.
00:02:37.960 | >> How much is Fast.ai courses?
00:02:40.320 | >> Free. Everything we do is free.
00:02:43.400 | We have no revenue sources of any kind.
00:02:45.920 | It's just a service to the community.
00:02:48.160 | >> You're a saint.
00:02:50.120 | Once a person understands the basics,
00:02:53.120 | trains a bunch of models,
00:02:55.320 | if we look at the scale of years,
00:02:59.040 | what advice do you have for someone
00:03:00.440 | wanting to eventually become an expert?
00:03:03.200 | >> Train lots of models.
00:03:05.640 | Specifically, train lots of models in your domain area.
00:03:08.520 | So an expert what?
00:03:09.920 | We don't need more expert,
00:03:12.240 | like create slightly evolutionary research
00:03:18.200 | in areas that everybody's studying.
00:03:19.760 | We need experts at using deep learning to diagnose malaria,
00:03:25.800 | or we need experts at using deep learning to
00:03:29.440 | analyze language to study media bias,
00:03:34.280 | or we need experts in analyzing
00:03:40.680 | fisheries to identify problem areas in the ocean.
00:03:44.880 | That's what we need.
00:03:46.480 | So become the expert in your passion area.
00:03:51.280 | This is a tool which you can use for just about anything,
00:03:54.400 | and you'll be able to do that thing better than other people,
00:03:57.520 | particularly by combining it with your passion and domain expertise.
00:04:00.680 | >> So that's really interesting. Even if you do want to
00:04:02.760 | innovate on transfer learning or active learning,
00:04:05.680 | your thought is, I mean,
00:04:07.440 | it's one I certainly share,
00:04:09.240 | is you also need to find
00:04:11.360 | a domain or a dataset that you actually really care for.
00:04:14.920 | If you're not working on a real problem that you understand,
00:04:18.200 | how do you know if you're doing it any good?
00:04:21.000 | How do you know if your results are good?
00:04:22.520 | How do you know if you're getting bad results?
00:04:24.000 | Why are you getting bad results?
00:04:25.240 | Is it a problem with the data?
00:04:27.280 | How do you know you're doing anything useful?
00:04:30.200 | Yeah, to me, the only really interesting research is,
00:04:33.960 | not the only, but the vast majority of interesting research is
00:04:36.720 | like try and solve an actual problem and solve it really well.
00:04:41.680 | >> Yeah, I think that's really interesting.
00:04:43.320 | I mean, I think that's really interesting.
00:04:44.840 | I think that's really interesting.
00:04:46.120 | I think it's really interesting to see how the world has evolved.
00:04:51.040 | I think it's really interesting to see how the world has evolved.
00:04:53.720 | I think it's really interesting to see how the world has evolved.
00:04:55.920 | I think it's really interesting to see how the world has evolved.
00:04:58.720 | I think it's really interesting to see how the world has evolved.