Meet Chris, the Co-Driving AI

Ask an expert over the CognitionX AI Advice Network on some of this week’s interesting articles, such as:

• Microsoft releasing a deep learning model which improves upon Google’s BERT and advances the frontiers of Natural Language Understanding
• “Chris”, the German-born, in-car, AI assistant, which received €7 million in funding
• Automatically modifying sentences to make AI systems more robust

Also in this issue of Speaking Naturally: Biased News Challenge and China’s Smart Speakers.

Under the Hood

Microsoft Transforms Google’s BERT
Microsoft’s new “Multi-Task Deep Neural Network” (MT-DNN) achieves state of the art in 9 of 11 benchmark NLP tasks (outperforming BERT — Google’s pre-trained language model). It integrates the underlying network architecture of BERT, but adds task-specific layers into the model’s architecture to enable multi-task learning (MTL) to fine-tune it.
Read more

Numbers that Matter

Chris, your digital co-driver
How often have you been driving and wished you could ask someone to do something on your phone for you, like dial a number or send a quick text message and ask them to change the car’s navigation or put on a good tune to listen to while sorting out the volume. These seemingly simple tasks become a nightmare when driving and can often distract your attention long enough to cause accidents. This is why German Autolabs (a startup in Berlin) created “Chris”, the world’s first digital co-driver. Chris is an AI assistant which can understand a driver’s voice and hand gestures while they are driving.
Read more

Learn from the Pros

Bixby has already learnt to speak in English
Bixby 2.0 (Samsung’s virtual assistant for Galaxy S10 smartphones and the Galaxy Home smart speaker) is gaining four new languages (UK English, German, Italian, and Spanish), superior speech recognition, session context maintenance (Bixby can recall your immediate conversation without requiring you to repeat details) and personalisation (Bixby can prioritise results and suggest services based on your behaviour and needs).
Read more

Making an Impact

Biased News Challenge
Given some news, develop an AI to decide if it exhibits blind prejudice or unreasonable allegiance to one party, faction, cause, or person. This was the hyperpartisan News Detection Challenge.

GATE won the challenge. Read more to see what they used to win.

Under the Hood

Automatically modifying sentences to make AI systems more robust
Just as image classifiers have been made more robust by slightly changing the training data (e.g. rotating the image, inverting the image, etc.), can text classifiers be made more robust in the same way? The tricky part is, how do you change a sentence automatically, without changing its meaning?

Pivoting is one technique (which assumes that if two English sentences can be translated into the same sentence in a second language, they must have the same meaning), or multi-pivoting (to translate the English sentence into multiple foreign languages and back again to generate a paraphrase) can produce interesting variants like “Great movie!” = “Great film!”. Another method of generating semantically equivalent sentences is to use Semantically Equivalent Adversarial Rules (SEARs). Rules may look like this: “what [NOUN]” -> “which [NOUN]”. “what [VERB]” -> “so what [VERB]”. etc. which would produce variants like “What did Tesla develop?” = “so what did Tesla develop?”.
Read more

CognitionX Research

Chatbots in Telecommunications: Business Case Studies (NEW REPORT)
If you’re a senior leader in the telecommunications sector, curious about the potential of a conversational AI strategy, CognitionX’s latest report on chatbot case studies for telcos is designed to help accelerate your ability to create strategies with confidence.

The questions the report explores:

• What does good look like for chatbots in the telecommunications sector?
• What are some examples of successful case studies?
• What do they cost and return-on-investment windows?
• + many others

£195+VAT for one license, £995+VAT for 10 licenses of this report Register Interest: Chatbots in Telcos

Numbers that Matter

China’s Smart Speakers
Just as the English speaking world have Alexa, Siri and Cortana, the Chinese speaking world have their smart speakers too, including: Alibaba, Xiaomi and Baidu. But which is winning over the Chinese market?

In the first quarter of 2018, Alibaba and Xiaomi accounted for 94% of smart speaker sales in China but Baidu matched Xiaomi’s relative market share in Q3 and solidly surpassed it in Q4. Then Alibaba accounted for over 40% of all smart speaker sales in China in Q4 2018.
Read more