Artificial intelligence (AI) is still a vast field with many unknowns, but it is not difficult for us to predict that some of the things that will or should happen in 2019 are inseparable from deep learning...
The use of language has become more and more sloppy. It is certainly convenient to use artificial intelligence (AI) as a shorthand for deep learning, and it is also easier to appear in media headlines. However, recent general AI (machines can learn on their own, just like curious people visiting a bookstore) is still more like science fiction, not science.
Recently, the deep neural network (DNN) has spread rapidly on the Internet like wildfire. DNN is a special case of AI, and it is usually based on a process initiated by people. Deep learning technology supports the ability to recognize images, speech and other domain models, which are usually faster than humans, thus opening up a new direction of computing. As for its long-term future, no one can say for sure.
Obviously, in the past one or two years, everyone has been rushing to take the AI train, no matter where it is headed. As for what value it can bring, it is not difficult to guess the next few stops of the AI train.
1. Accelerated chips will show more momentum
As we reported in September, at least 4 new accelerators for training deep learning neural networks are being sampled. For a long time, many network giants in the industry have been looking forward to the emergence of these chips. As Greg Diamos, a senior researcher at Baidu’s Silicon Valley AI Lab, said at the end of 2016, the task of training machine learning models is "limited by computing power. If you have a faster processor, you can Implement a larger model."
Therefore, 2019 is expected to see some of the top data center operators begin to buy these chips in large quantities. However, it would be too unrealistic to expect that the AI startups that have been set up in a swarm will be eliminated. I think we should see some early market winners gain market share and actual gains.
2. Strictly examine the value of AI
Some startups that have gained market momentum in the field of deep learning accelerators are also expected to receive large sums of capital. As investment companies start to wonder how much return on investment (ROI) they can get, it is expected that the investment boom that started this fall will gradually cool down in 2019.
This wave of deep learning craze has so far attracted investments in about 50 start-up companies and more than tens of millions of dollars. In the past few weeks, there has also been another wave of new investment boom.
The Israeli startup Habana Labs completed a $75 million Series B fundraising in November, increasing its total fundraising to $120 million. Wave Computing raised US$86 million this month. So far, it has accumulated a total of about US$200 million. A part of the funds has been used to acquire MIPS, and recently announced its open source core plan.
The British startup Graphcore recently announced the completion of a $200 million round D fundraising, with a total of $312 million raised so far. Graphcore chips have also recently been used in Dell's newly designed systems. It is expected that there will be a lot of investment in the field of deep learning, but as business managers begin to calculate their actual benefits, there will definitely be many "hard landing" and "soft landing" investment strategies.
3. Inferred performance is based on running points
When it comes to numbers, in addition to the MLPerf test benchmark used to train deep learning networks, it is expected that other benchmarks will appear in 2019. The goal is to release a set of inference task benchmarks based on cloud and embedded systems.
I'm not sure if this counts as a prediction. But according to the MLPerf organization, this is their plan in 2019. Therefore, in 2019, I expect that all the enthusiasm for training will be transferred to the larger inference chip market.
4. Chip vendors embrace benchmarking
This is not really a prediction, it's more like a "command"? Chip suppliers must accept emerging deep learning benchmarks. Customers who can’t wait and generous investors should ask for benchmarking so that this emerging market can obtain some much-needed key analysis and guidelines-after all, we can’t rely on market hype forever!
So far, only Google, Intel (Intel) and Huida (Nvidia) have used the early MLPerf version 0.5 training benchmarks and released the results for a few systems. Many companies need to publish more benchmark results for various configurations and workloads, so that related fields can see where they are, and calibrate the parts they need.
5. AI software platforms will quickly become popular
This may have already happened. In the past few weeks, I have received information released by various AI software platforms almost every day. Due to the simplification of application development and the market pressure of AI platforms, I highly doubt what value these products bring.
In the next few years, this AI software platform jungle will become denser and more diverse. End users and investors will begin to evaluate carefully.
6. Deep learning will encounter bottlenecks
It stands to reason that this is already happening, but no one has connected everything together. For example, during the vacation, I once carefully browsed all the playlists on Pandora radio station, and then clicked on the track suggestion button that may be added on the Bach App. As a result, the recommendation engine brought me back to the screen where I browsed all the tracks at the beginning.
Pandora is not the only web app with insufficient features. I expect there will be several storms of consumer rebound in 2019. I hope that programmers and marketers can exercise restraint and don't be driven crazy by headlines like "artificial stupidity." Even if you have a great core technology, you need a good sense of human cooperation.
7. Rise of interest in pervasive AI
The industry's enthusiasm for deep learning and funding has also aroused interest in researching pervasive AI. I am not an expert in this field myself, but I noticed that Numenta, founded by Palm Pilot designer Jeff Hawkins, reportedly made progress this year in developing a general theory of how the new cortex works.
No one really knows how the human brain accomplishes some incredible things. No one can explain why deep learning can effectively match results and patterns, even for very narrow and artificially controlled AI omens.
In 2019, I expect that smarter people will start to doubt some of the bigger problems. I hope this will lead to some interesting discussions, and maybe even some important developments that no one has predicted.
