Source | HackerNews
Compile | xiao
In the past two or three years, tepid machine learning suddenly became popular in the industry, so not only the traditional industry, but also the once-popular Internet companies began to wonder: should we engage in machine learning? Should there be an ARTIFICIAL intelligence research institute? What exactly can machine learning, or AI, bring to the company?
HackerNews, a community for tech entrepreneurs at Silicon Valley startup accelerator YC, posed the question: Where does AI, or machine learning, really add value to your company?
Users from tech giants, startups and even traditional industries gave their answers, both sides of the coin.
The AI100 features the following 9:
Altshiftprtscrn answer
I work in manufacturing. We have an acoustic microscope that scans the parts in order to identify internal defects (usually particles trapped in the epoxy). It is difficult to define what equipment is defective by the size, shape, location and quantity of these particles. Our final product test can tell us whether the product is “good” or “bad” in the electrical measurement sense, but this test cannot be applied in the assembly phase where we are concerned with identifying defects.
I recently demonstrated a very simple bagging decision tree model that “predicted” downstream test failures of scanned parts with 95% accuracy. I don’t really have a big background in machine learning, I just apply the principles, I don’t understand them, and it’s entirely possible to do that. (Yes, I feel really guilty.)
The results show that spotting problems early could save $1 million a year in defective product costs (if the model is approved for production). This is just one product, in one factory, in one company, and there are more than 100 such factories around the world.
The experience prompted me to go back to school and learn this stuff in a more organized way. There is great value in using machine learning in complex manufacturing and supply chain environments (in other words, it can avoid waste).
Sidlls answer
The entire product I developed last year could have been reduced to a basic statistical problem (e.g., ratio, probability), but because of the hype, we built “models” and then “predicted” specific outcomes based on the data set.
I work for a company that sells a product that more or less attempts to use “machine learning” to find duplicate items in a large data set.
It’s not machine learning per se that provides value for the product, but rather the hype that has recently swept Silicon Valley: our customers look at a “data science product” and have no idea that it’s just basic predictive analytics. I don’t know if the product would sell without the label.
To be clear: THE company I work for does use machine learning, and I do work on the company’s data science team. I turn off the TV: We don’t need to do this, because our product can do it with basic technology. It doesn’t have to be this complicated.
Ekarulf answer
Amazon personalization.
We use machine learning/deep learning to provide product suggestions and recommendations to our customers. For years, we only used algorithms based on basic statistics, but we found that machine learning models were much easier to do on this occasion.
(The user also included a related Amazon blog post, GitHub address and job listing at the bottom of her answer.)
Strebler answer
We’re a computer vision company, and we do a lot of product detection + recognition + search, mostly for retailers, but we also make money in a lot of other image verticals. My co-founder and I have both written papers in the field of computer vision.
In our part of the world, recent advances in artificial intelligence, machine learning have made possible things that were simply impossible before.
Having said that, the hype around deep learning has become very serious. Several of our competitors have gotten out of the business (even though they are taking advantage of the wonders of deep learning). JustVisual, for example, was sold a few months ago ($20 million +) and Slyce ($50 million +) will be sold later this month.
Yes, deep learning has made some very fundamental advances, but that doesn’t mean it will magically make money!
Jngiam1 answer
Here at Coursera, we use machine learning in a few places:
-
Course advice. We use a low-rank matrix factorization approach to make recommendations, and are also considering integrating other sources of information (such as your career goals).
-
Search. The results are based on correlations among rankings of various signals, from popularity to learner preferences.
-
Learning. There’s a lot of untapped potential. We have looked at some of the peer de-Biasing and worked with people at Stanford on how people learn code [2].
We recently co-organized a workshop on machine learning education for job skill levels: http://ml4ed.cc. There is untapped potential for using machine learning to improve education.
Ksimek answer
At Matterport, our research team is using deep learning to understand the 3D Spaces our clients scan. Deep learning is great for companies like ours, where most of our data is images, and before the advent of deep learning, it was impossible to extract that information in a high-throughput way.
One of our application scenarios is to automatically create panoramas for viewing houses. Real estate is a big market for us and one of the keys to our product is the ability to create a tour mode that automatically plays slideshows or 3D images. The problem is that creating these images manually takes time, as it requires making a 3D model and finding the best view for each room. We know these things can add tremendous value when selling a home, but many of our clients don’t have the time to create them. In our research lab, we use deep learning to automatically create navigation maps by identifying different rooms in a house and whether the picture is more attractive. We trained it with a collection of hand-taken photos from about a million users, some of which marked room types.
It’s not too far off, but we’re also looking at semantic segmentation for 3D geometric evaluation, using deep learning to improve deep data quality, and other applications of deep learning for 3D data. Our clients scanned approximately 370,000 buildings, generating approximately 300 million RGBD images of live action.
Flammy answer
The startup I’m involved with uses machine learning to predict what will drive users away.
We work with B2B, B2C SAAS, mobile applications, games and e-commerce. For each of them, a common solution is customized to let them know which users are most at risk of churn. The time span varies by customer life cycle, but for the longest customer life cycle, we can predict customer churn more than 6 months in advance with high accuracy.
“Than” Who is at risk? More importantly, “Why are they at risk of losing them?” . To answer this question, we looked at patterns and sets of behaviors that are positively and negatively correlated with churn, so that our customers can purposefully encourage, discourage, or modify these specific behaviors.
This allows our customers to try and retain their users in a variety of ways. With our B2B clients, account managers are very confident about who they need to contact and why.
All of this includes regular model retraining, taking into account events and behaviors of new users, new product updates, and so on. We are confident in our solution and are offering our customers a free trial to prove ourselves to them.
I can’t share the details, but we just signed our biggest contract this morning.
Iamed2 answer
We use machine learning to model complex interactions in the grid in order to make decisions to improve grid efficiency, which has been (at least in the short term) more effective than using optimization programs to get better results.
In general, I think if you know your data relationships, then you don’t need machine learning. If you don’t know, machine learning can be especially useful.
Got2surf answer
My company makes software that analyzes customer feedback.
We use “real” machine learning for opinion categorization, as well as some tools for natural language processing and opinion mining. However, most of the valuable results come from simple statistical analyses, probabilities, ratios, as other commentators have mentioned. Machine learning is important for determining whether a customer is angry in a feedback comment. Over time, however, its role in trending topics has become less important