ML Trends

Newsletter #1 - Feb 17, 2019

Whats real anymore?

It's here.

We're thrilled to bring you the first edition of our ML Trends newsletter. As the field of machine learning progresses with unprecedented pace, we want to help you stay on top of the latest happenings, trends and myths. In each edition, we'll focus on one area out of natural language processing, computer vision, machine learning and artificial intelligence. This edition is all about natural language processing. So, sit back in your armchair and dive in. Don't forget to bring some coffee. 

Straight from the labs
Articles can be faked too.

After people's faces and talking videos of Barack Obama, deep faking comes to text generation. Recently, openAI showed how they were able to generate paragraphs of text by training deep models on troves of data crawled from the Internet.

How does it work? Just prime the model with a short snippet of text and it will write you an article that looks grammatically and semantically coherent. The idea is not particularly new and does require a lot of time and compute power. But, what’s interesting is the ability of this model to generalise well to a variety of domains and downstream NLP tasks like classification, question-answering, etc.

As an optimist, this can be put to good use towards automated writing assistants like the auto-completion and smart reply features within Gmail. On the downside, it makes it extremely easy to generate seemingly real yet fake news, articles, tweets, etc. Probably, one of the reasons why the folks at openAI have publicly released only a smaller model for research and testing purposes.

Check out their demo here. 
Find out more at: OpenAI announcement

Into the realm of business
Machine translation goes B2B.
Google Translate has nearly captured the average consumer's requirements with their quirky-but-free offering. But, the B2B space is still warming up to it. Their main concerns? Accuracy and speed. Especially, given the context. SaaS companies are now trying to solve the accuracy problem with a human-in-the-loop approach.

How does it work?
1. Generate a rough translation using a neural net or Google's translation APIs
2. Bring in an expert translator to vet the results.
3. Pass the feedback to the system and learn.
Rinse and repeat till you gather enough domain-specific data. It's a neat strategy for companies that lack Google-scale data. The other challenge remains - How do you provide real-time translation results?

Couple of companies to watch out for: Lilt and Verbit
Out in the real world
It's all subjective.
HSBC recently used NLP to analyse the earnings calls for MSCI US and EU companies with extensive Chinese exposure. The results found analysts to be overly bearish about the future prospects of these companies. Interestingly, the bearish attitude did not come from objective indicators but from negative sentiments surrounding the strenuous political relationship between the US and China.

The results also showed that the company management were quite positive and highlighted a bullish undertone; a direct contradiction to the analysts' negative outlook. In fact, companies which had stronger negative analysts' sentiments vastly outperformed the analysts' expectations. This example serves to remind us that opinions are influenced by highly subjective indicators, even if these opinions come from seemingly objective sources (aka analysts).

Here's the full article.
We hope you enjoyed the first edition of our newsletter. If you've got a couple of minutes to spare, do take this survey and help us get better.

If you found this interesting, you can help us grow by asking your friends, colleagues and enemies to subscribe to our newsletter.
Enjoy the rest of your weekend.

The ML Trends team
P.S. We planned on sending out the newsletter last Sunday. Unfortunately, Google's algorithms decided that our website was phishing for information. We're safe and cool now. We're also relieved to find out that Skynet is still a few years away.
Copyright © 2019 mlTrends, All rights reserved.