What is Synthetic Data? No, It's Not 'Fake' Data

Synthetic data is artificially generated data versus data based on actual events, but it's not "fake" data. It replicates the properties of real data without the troubles of capturing it, such as confidentiality, low-volume, or expensive-to-validate. With synthetic data, it's easier and less costly to train AI models, however, it's not a panacea. For example, synthetic data may not fully represent the unexpected events that happen in the real world. In this video, Martin Keen explains what synthetic data is, its uses, benefits, and challenges; he wraps up his presentation by explain how it's generated.

#datascience #businesssolutions #lightboard #ibm #computerscience #data #machinelearning

[danielmaciel]

I am amazed how this dude can write backwards so perfectly

[tmastana]

Amazing series and very classical and engrossing style of explanation... keep up the good work

[amazingwarrior]

What is very interesting about this concept is the validity and reliability of them. Why they don't talk about it! it's essential when we talk about mathematical set's of any data!

[bejxtyn]

Really love these IBM mini lectures, they are very insightful. Helped me during my college days, and are also helpful for learning as a hobby. Thanks!

[anandkalhore]

Can synthetic data be as effective as real data? Wouldn’t model getting trained with synthetic data be giving false results when used against real data?

[rickharold]

Yes, cool stuff. We use synthetic data for tracking trucks in the field. By taking existing labeled data and transforming the truck in three dimensions to get the additional data for the model.

[lozanojavier]

I find it difficult to stop thinking about Martin Keen, and his prediction about Southampton's future in the Premier League. It's quite remarkable that both Southampton and Leicester will be battling it out in the Championship to regain their positions in the top tier in 2025. A great example of the problems with synthetic data.

[talalrahim]

You are a very good teacher. Do you have a full course on this?

[HoustonKhanyile]

I think this video might have jinxed Southampton. Instead of winning the Premier league they are now getting relegated.😢

[mthoko]

Great series from IBM in general and this instructor specifically . Slightly hopeful on the Southampton bit but if you can't dream, what's the point of it all😃

[karengomez]

Takeaway:
Made up data can be used to deal with biased real word data and can be obtained from data sources or transforming existing data by adding noise or using GANs.

[seanrrr]

Synthetic data has been very useful in my field (gene regulatory networks; maps of interactions that affect gene expression within cells). We can't manually test the interactions of tens of thousands of genes, especially across tens/hundreds of thousands of species, so we predict them using large molecular datasets. 

The problem is, how can you evaluate the accuracy of a prediction algorithm if you don't know what's true or false? Synthetic data is super useful, since you can generate data with known interactions that you can compare to. Algorithms can then be ranked on how close their predictions match the synthetic dataset. A great example is the GNW DREAM Network Inference Challenge, if you want to see how they use this!

[quantumpotential]

What kind of transparent white board is he using to write on? Very cool. Have not quite seen this before.

[KNOT-zdwh]

Can we add regional human corruption to make synthetic data more reliable one also and should it be under noise?

[nagkumar]

Why is it not called a fake message that is not clear in the video..

[almor]

How is this not basing later models on copies of copies of potentially incorrect data? Won't we end up with piles of structurally sound, true seeming noise eventually?

[ianoldfield]

Interesting, if rather simplistic. Having spent the past 5/6 years developing a synthetic police-data model, it is not easy or cheap (if time is factored in). Rows and rows of financial transactions might be easy to generate, less so, complex family groups, locations, incidents and crimes, vehicles, organisations, where these are interlinked, related and reflect real-world scenarios. Whilst IBM has some excellent tools such as i2 and Watson, the real data in those systems would be unlikely to be made available for sythesising.

[nicoles_handle]

using the prem was the perfect hook icl

[watipasokamanga]

nice, now I can generate data for my HIV viral load detector model at no cost

[itdataandprocessanalysis]

Thanks for the video.

May I ask... is this British accent?