The author introduces: commercial technology review, public address (biztechreview)
Not long ago, according to The Register, there were internal sources in IBM that the Watson health department had to lay off about 50% to 70% of its employees.
This is very bad news for IBM, which has been heavily promoted by IBM over the years.
As the first enterprise to push AI technology to the market, IBM can take a look at some of the pain points of AI landing from the failure of Watson health.
1. Lack of effective evaluation criteria
We can see a variety of comments about Watson health online, and some people say that
Of course, a large part of this is because IBM confuses attention by means of media and advertising in order to publicize its products. But the lack of unified basic testing in the industry makes it impossible for AI products to evaluate quantitatively, which is the essence of the problem.
Other industries, no matter how advertising is advertising, will have its own scale in the industry, and by benchmarking, people can always be divided, but in the AI industry, except for images and voice, the recognized benchmarks do not exist at all.
Before Li Feifei set up the ImageNet image test set, there was no uniform evaluation standard in the field of image recognition. It is difficult to quantitatively evaluate the advantages and disadvantages of various algorithms.
We can see that since the integration of ImageNet data as the benchmark for industry benchmarking, it has greatly promoted the development of image recognition. Image recognition quickly becomes the fastest growing field of artificial intelligence.
It naturally has the applicability of the algorithm, but it is of great significance for the whole industry to establish a unified evaluation standard to make the industry quantitative analysis of the effectiveness of the algorithm. This is why Li Feifei is so respected in the AI field.
We come back to see that in the field of health care, although IBM has worked for many years, it has hardly published any industry recognized test results, or has developed a rigorous system of evaluation for the outside world.
In turn, IBM, through a variety of non academic channels, publicized its achievements, which led to high expectations, but could not verify its effectiveness in the application, which was easy to distrust.
2. There is no good business model
Business models are crucial to technology realisation, but IBM has not found a business model for its AI products that both sides accept.
IBM's business model is very old-fashioned, that is, through technology service contracts to lock customers, and then send personnel to partner there to carry out project implementation.
This model of cooperation is generally applicable to the traditional IT project. Because of the clear target, clear demand and relatively predictable input-output, both partners can control their input in a controllable range. But at present, some AI projects have huge investment scale, but their profits can not be measured.
Take the well-known IBM and MD Anderson Cancer Research Centers as an example. MD Anderson reportedly paid IBM $39 million in fees, but it also noted that:
Another case was the cooperation between the IBM and the Singaporean government. According to the titanium media, the traffic management was affirmed at the time of the experiment. But in the later period, the traffic management department in Singapore needed to pay a huge amount of expenditure in the later period. The huge cost was prohibitive to the relevant departments.
Because of the rise of cloud computing, now IT services have gradually changed to pay per - demand, and more and more companies are reluctant to pay a huge amount of fixed cost for the indeterminate effects. AI services need to find similar patterns to get customers' trust. The mode of kidnapping customers by traditional contracts is definitely not the future.
3. Lack of effective industry model and training data
We know that AI computing now requires a mature industry application model and a large amount of annotated data for the system
And many times such models and data are very scarce.
We can also see this dilemma in Forbes's report to IBM Watson.
In order to provide data support for Watson's health, IBM has made a lot of acquisitions in recent years, many of which are medical data analysis and solutions companies. This includes Truven, a medical data company that bought $2 billion 600 million in 2016, Merge, a medical imaging company that bought $1 billion in 2015, and the medical care management company, Phytel, which was also bought in 2015.
But even with so much investment, IBM doesn't seem to get too much quality data, and its AI training performance is not as good as expected. The Forbes report quoted experts as saying:
The medical data and service companies that had been acquired were the main part of the layoffs. They also proved that they did not bring much value to IBM.
Watson of IBM is the earliest pioneer of AI industrialization, and the difficulty it faces is also the predicament faced by the whole industry.
The new AI enterprises need to overcome these difficulties: to establish an industry accredited evaluation standard, to launch a more flexible implementation scheme, to control the cost of the enterprise, to establish a real industry application model and to get mass data for training, and to get the available solutions. Only by completing these points can AI really be applied from concept to application.