Like most 80s/90s kids, I dreamed of being Like Mike. Flying through the air, controlling the courts, and making the impossible possible.

So, when I finally finished The Last Dance, all the feelings bubbled back up again as I dreamed of soaring like Air Jordan. One small problem - I’m a bit past my prime to be launching my basketball career. (To be fair, I wouldn’t be the oldest ever to debut, but my height among other genetic factors help to keep me... “grounded.”)

Over the past few years, though, there’s been incredible progress with ML techniques in the generative family (e.g., GANs) that have yet again blurred the line of what machines can accomplish.

As a brief primer, generative models in essence let us produce something (music, images, text) based on a primer of information (randomness, categories, etc). They’ve produced remarkable works like* instant painting, authentic articles, and even Chinese characters*.

In a sense, these generative models can dream up new ideas, concepts, and realities. So… what’s the chance they could make my dream come true?

What if I was MJ?

I know. I can’t be MJ. There’s only one MJ. But what if instead of being like MJ, I was half-MJ?

Strange and yet…now introducing Connor Jordan. Half-Connor, half-MJ, 100% GOAT:

Connor Jordan is the product of taking half of my facial features, half of MJ’s facial features, and blending them together equally in a StyleGAN2 architecture. This person doesn’t actually exist, but I bet he would catch some serious air with those gravity defying locks of hair.

So, What is Connor Jordan's Story?

Connor Jordan is the best basketball player to have ever lived. No one before, and no one after could have ever compared. Jordan had a great career in the NBA. He spent his entire NBA career playing and scoring and being a true basketball star. In 2012, the Golden State Warriors drafted him No. 15 overall in the seventh round. After being traded, Jordan became a major NBA star and it wasn't until his career was over that he was given the nickname "Mama " for his role in NBA basketball. Jordan was the youngest member of the All-Star team in the NBA.

He was born and raised in a very tough world with no family. Jordan was a true basketball player with a long and beautiful story.

Or at least, that’s what a language model generator would say that was only primed with the first two sentences…

Quite the tragic hero story for someone who overcame so much adversity and was drafted… in the seventh round? Unusual considering the NBA draft only has two rounds, but it does make for an even better underdog story...
I would’ve loved more elaboration on how Connor Jordan gained the nickname “Mama.” Was he making players cry for their mamas? Was he mopping the floor with the youngins?

What would MJ say?

The above is all well and good, but it really means nothing since we can’t really benchmark Connor Jordan against the real GOAT Michael Jordan.

Unless Jordan made some admission. But he'd never do that... wait, he said what?

Well, thanks MJ. That really means alot coming from you.

Conclusion

This short project was heavy on exploration of generative models with a bias towards reading research and leveraging existing tech over developing code. In addition to saving myself the hours required to train a state-of-the-art model and up to millions of dollars in expenses, this enabled me to explore the questions of 1) how good off-the-shelf components are today in areas I spend less time in and 2) what should we expect over the next few years.

My answers to the above are:

Off the shelf components have reached near real levels real fast (for certain use cases). Already incredibly real synthetic images can be doctored, voices cloned, and likeliness replicated in videos. Most can be replicated for free with the right knowledge and many others are available at a low cost via commercial solutions as detailed in another post. The above was achieved in less than a weekend of gathering photos and testing code/products; with access to the right tools, the realism can approach near authentic levels to say the least.
Ian Goodfellow had a great tweet detailing progress in computer vision GANs in under 5 years. While we may not see the same leaps and bounds moving forward, I am now even more confident that the next 5 years will bring us applications and technology that will increasingly pass adapted Turing tests. This will lead to some creative applications that will democratize creativity and productivity to the masses. It'll also lead to some nefarious applications that call into question the trust fabric our societies are established on. Coalitions and regulations will need to be established as a countervailing force.

Reference

Making the artifacts for Connor Jordan:

Select models and papers for further exploration:

Computer Vision: StyleGAN2, Talking Heads, Arbitrary Style Transfer, CycleGAN
Synthesis of Language Model progress with paper references