Have you ever wished that you'd taken that latest selfie in summer not winter? Or that your photo of a horse was actually a cool and exotic zebra? Good news - the latest research from UC Berkley is looking to fix exactly these problems.

Most image translation today uses large training sets of aligned pairs as a starting point, but often these training sets aren't available or would be hard to create. This research is looking into tackling exactly this problem: how can AI tackle translation tasks without a good starting point? The results for changing colour or texture within an image are great, but so far the results for changing shape are not so good - there's a particularly awful example of an attempt to change a cat into a dog.

More generally, I think this is a great example of how flexibility is becoming an increasingly valuable commodity in AI. It's not just about who can create the most accurate algorithm, but also who can do it the fastest, with least training data and covering the broadest range. As deep learning algorithms move from labs and competitions to the business world, the real winners won't be the perfect algorithms - it'll be first to market with a 95% solution.