Shulou(Shulou.com)06/03 Report--

Produced by big data Digest

Compiled by Xiao Jiang, Li Ke, Gou Xiaobai, Zhou Suyun

We are in the midst of an AI startup boom, where salaries for machine learning experts are rising and investors are happy to give generously to AI startups. The popularity of AI has become a symbol of promoting social productive forces, which is bound to change our lives.

But CTO Ric Szopa of Inovo.vc, a former Google engineer who wrote this article, believes that the skills of AI practitioners are depreciating. Starting with a multiple choice question, he told us that AI tools, data sets, capital investment, and the advantages of industry + AI are gradually weakening the advantages of a single AI basic technology.

Let's start with a multiple choice question.

Alice and Bob are two AI entrepreneurs whose companies have raised roughly the same amount of money and competed fiercely in the same market.

Alice spends most of his money hiring the best engineers and brings in a group of PhDs who are experienced in artificial intelligence research.

Bob, on the other hand, chooses to hire moderately qualified but reasonably capable engineers and uses the money saved to get better data.

If it were you, who would you invest in?

Bob, of course.

Why?

In essence, the principle of machine learning is to obtain information from data and convert it into model weight. Better models make the process more efficient (in terms of time or overall quality), but if you assume that the models are relatively well trained, better data will certainly produce better results.

To illustrate this point, let's do another quick and simple test.

Suppose I create two convolution networks with different performance. The last fully connected layer of the "better" model has 128 neurons, while the "slightly worse" model has only 64. I train them on subsets of MNIST datasets of different sizes and draw a line chart of the accuracy of the model on the test set and the number of training samples.

Blue is the "better" model and green is the "slightly worse" model.

Obviously, the size of the training data set has a positive impact (at least until the model begins to overfit and the accuracy is stable). It is worth mentioning that the "slightly worse" model trained on 40000 samples is more accurate than the "better" model trained on 30000 samples.

In my small example, we are dealing with a relatively simple problem and have a more comprehensive dataset. In real life, our conditions are not so perfect. In many cases, adding a dataset often has a very significant effect.

In fact, Alice engineers don't just compete with Bob engineers. Because of the open culture of the AI community and its emphasis on knowledge sharing, their competitors are actually researchers from Google, Facebook, Microsoft and thousands of universities around the world.

Therefore, good engineers are important, but if you are in the AI field, the competitive advantage of data will be even more critical.

The more important question, however, is how you can maintain your advantage.

AI tools are getting easier and easier to use

In 2015, when I was still working at Google, I first started playing DistBelief, which is later known as Tensorflow. The tool was so difficult to use at the time that it was a pipe dream to run outside the system built by Google.

At the end of 2016, I conducted a proof-of-concept study to detect breast cancer in histopathological images. I wanted to use migration learning at the time: using Inception, Google's best image classification architecture at the time, and retraining on my cancer data. I can use a pre-trained initial weight provided by Google to change the top-level structure to match what I'm doing.

After a long period of trial and error on TensorFlow, I finally found a way to manipulate different layers and make it basically work. It takes a lot of perseverance to read TensorFlow materials. But at least I don't have to worry too much about dependencies, because TensorFlow intimately prepares Docker images.

Source: Medium

In early 2018, thanks to Keras, a framework based on TensorFlow, it took only a few lines of Python code to complete the project, and you didn't need to have a deep understanding of what you were doing to use it. But it still has a pain point: hyperparametric tuning.

If you have a deep learning model, you can adjust multiple parameters, such as the number of layers and size. As I was writing these words (early 2019), Google and Amazon offered automatic model tuning service (Cloud AutoML,SageMaker).

I predict that manual tuning will become extinct sooner or later, and engineers will be freed from this tedious task.

The general trend is to make difficult things easier, and you can achieve more without deep understanding. Those great projects of the past sound quite mediocre now, so we should not expect how good our present achievements will be in the future.

It sounds exhilarating, doesn't it, but it can be bad news for companies and individuals who have invested heavily in AI technology. At present, mastering some AI technologies is a competitive advantage for enterprises, because a competent machine learning engineer needs to spend a lot of time reading papers and need a solid mathematical background.

However, as the tools improve, this will no longer be the case. Reading papers is more likely to turn to tool tutorials. If you don't quickly realize what you should focus on, a team of interns with better data may take your job.

Want to maintain a competitive advantage for a long time? extremely difficult!

Let's go back to the example at the beginning of the article. With excellent data sets, Bob successfully competed with Alice, launched its own product, and steadily increased its market share. He can also gradually start hiring better engineers because his company is rumored to be a good place to go.

But at this time, there is another Chuck, although he entered the game late, but he is richer than Bob.

Money is crucial when building datasets. But it is very difficult to speed up the progress of the project by throwing money. In fact, using too many newcomers may slow down progress, but building a dataset is different. The dataset requires a lot of manual manipulation, and you can fix it by hiring more people. Another possibility is that someone owns the data, so all you have to do is pay for the use of the data.

In any case, money can make data sets come faster.

But the question is, why can Chuck raise more money than Bob?

When founders propose a round of financing, they try to balance two potentially contradictory goals. They need to raise enough money to compete in the market, but not too much, as this will lead to excessive dilution. The founding team must maintain sufficient shares in the start-up company so as not to lose the motivation to start a business.

On the other hand, investors want to invest in ideas with huge upward potential, but they must control the risk. As the expected risk increases, they will demand a larger proportion of shares for every dollar paid.

When Bob raised money, "artificial intelligence does help the product" was just a belief. No matter how good he is as a founder and how good her team is, it is possible that the problem he is trying to solve is simply too difficult. The situation in Chuck is very different. He knows that the problems he faces can be solved!

In this case, Bob's response is likely to be to propose another round of financing in order to be in a good position because he is still (temporarily) ahead of the competition. But what if Chuck has solid access to data through strategic partnerships? In the case of a cancer diagnostics start-up, for example, Chuck may use his internal position in an important medical institution to reach an internal agreement with that institution. At this point, Bob may not be able to compete.

Your product needs to be defensive, preferably a moat.

Leverage effect of AI

One way to classify a business is whether it directly adds value or provides leverage for some other source of value. Take an e-commerce company as an example, adding value is like creating a new product line, establishing a new distribution channel is a lever, and cutting costs is also a lever.

Leverage may be more effective than applying force directly. However, leverage works only when coupled to a direct source of value. A tiny number, doubled, tripled, is still very small. If you don't have parts to sell, it's a waste of time to open up new distribution channels.

What should we think of AI in this situation? There are many companies trying to use AI as their direct product (API for image recognition, etc.), which may be attractive to an AI expert.

However, this is often not a good choice. First of all, you are competing with big companies such as Google and Amazon. Second, it is very difficult to develop really useful general-purpose AI products. For example, I've always wanted to use Google's Vision API. Unfortunately, we have never encountered a situation where customer needs and products fully match. There are always all kinds of situations that are either underdeveloped or overdeveloped.

A better option is to think of AI as leverage.

You can adopt an existing, effective business model and enhance it through AI. For example, if the production process depends on human cognitive labor, automating it may lead to a significant increase in gross profit margin. The examples I can think of here are: ECG analysis, industrial inspection, satellite image analysis. It is also exciting that because AI is an auxiliary backend, it can still take advantage of non-AI business to maintain the company's competitive advantage.

Conclusion

AI is a truly transformative technology. But starting a business on this basis is a tricky business. You should not rely entirely on AI skills, because the market trend is that technology will depreciate.

Building an AI model can be fun, but what really matters is having better data than your competitors.

It's hard to maintain a competitive edge, especially with competitors who are better funded than you are, which is likely to happen when you start an AI business. Your goal should be to create a scalable data collection process that is difficult to replicate by competitors.

AI is well suited for subverting industries that rely on low value-added and repetitive labor because it makes it possible to automate the work.

Related reports:

Https://www.kdnuggets.com/2019/01/your-ai-skills-worth-less-than-you-think.html

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.