Twitter (NYSE:TWTR) recently acquired Magic Pony technologies, a machine learning startup focusing on deep learning in computer vision. In my continuing coverage of the AI market, I want to briefly comment on this transaction.
How magic is this pony?
First, another contributor has already commented and described the company's product (without technical details). He claims:
At its core, Magic Pony uses machine learning to enhance low resolution images and videos. It is quite remarkable when you think about it. We know that a high resolution photo can be easily transformed into low a resolution photo, since we have all the raw materials that we would need in the original photo (i.e. the pixels). What Magic Pony could do is akin to creating something out of nothing. How can a blurry image be transformed into a sharper image? That's where machine learning comes in. Presumably Magic Pony's technology can infer new information from existing data.
Let's clear up a few things here. Magic Pony did not invent image super resolution with deep convolutional networks. In fact, for anyone dealing with deep networks or more generally, generative probabilistic models, this is quite an intuitive idea. For instance, Microsoft (NASDAQ:MSFT) research has published on the topic in 2014. I would go as to say that using deep learning for image enhancement is pretty much in the public domain, e.g. here is a post from Flipboard describing their in-house use of it. Hence, I would say whatever Magic Pony has developed is probably innovative and a specific way of applying this established technique, but not an image processing revolution.
So what did they invent, if anything? Let's look at some patent applications, of which the company has quite a few:
The core expertise of Magic Pony technology is in video compression techniques. Judging from the few descriptions available, I presume the technique for-low compressions streaming works as follows:
Consider a mobile streaming app like Periscope from the client's side. A client streams an event to Twitter's servers. Twitter applies filters to down-sample stream quality - this part is relatively easy, reducing resolution. Users who wish to follow this particular stream get a low-resolution stream. The Magic Pony 'bolt on' library uses the mobile GPU to pass video frames through a pre-trained, compressed convolutional net, which outputs the enhanced representation.
Hence, the Traveling Investor's statement:
Compression could save Twitter some money on bandwidth, but my guess is that it will be offset by the cost of processing the data
is wrong and also somewhat beside the point. It is users who will save bandwidth and retrieve videos with more stability, and in turn they will use their mobile phone's GPUs more, which will probably drain batteries a bit faster. Overall quite useful, since a key problem in delivering video content in many regions of the world is the network, not the end device.
Twitter pays $13 million per machine learning Ph.D.
Apparently, Magic Pony technology has 11 machine learning Ph.D.s, which is a decent enough package for an acqui-hire. Did they overpay? Maybe a bit. See, the machine learning market has been getting hotter and hotter. In 2014, Google (NASDAQ:GOOG) (NASDAQ:GOOGL) paid $500 million for DeepMind. At the time, there were about 50 deep learning experts in the world and DeepMind employed a sizable chunk of them.
Today, the market is somewhat different. Everyone and their mother (including yours truly) are doing machine learning research and in a few years, there will be a large outflow of freshly minted Ph.D.s from universities. Right now, there is still a massive shortage for machine learning experts and if a tech company can buy that expertise in the bundle, that makes things easier. On the other hand, Magic Pony technology is certainly not DeepMind. DeepMind is the bleeding edge of machine learning research. Magic Pony technology has interesting technology, but it is more a strategic move for Twitter trying to establish a larger machine learning team. Or, in other words: Google got the best of the best in absolute terms.
Twitter got the best they could find on a supply-constrained market and actually got some interesting technology to go with it. I expect a continuing slew of machine learning acquisitions from major players in the next few years. Likely candidates are companies like Skymind, which turn state of the art research into enterprise level products. I am somewhat skeptical of companies like Vicarious, who have vague, high-flying goals but no products, seemingly waiting to be acqui-hired.
Overall, Twitter might have made some good use of their large cash position for their focus on video, but it changes little on my overall pessimistic outlook. Still waiting for a turnaround in user-engagement and growth, still not expecting one.
Disclosure: I/we have no positions in any stocks mentioned, and no plans to initiate any positions within the next 72 hours.
I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.