This week we will look at an AI that tries to predict upcoming rain and an AI that can detect lung cancer from a blood test. We will also have a quick look at Automated Trucks and how hackers are now able to take control of your phone’s voice assistant.
Everything I read lately is about Covid-19. I will try and keep this newsletter a Covid Free Zone (CFZ) however there is a significant amount of interesting technology being very rapidly developed to deal with this virus. I will therefore produce a bonus newsletter full of Covid-19 news and tech. I hope to send that out tomorrow for your weekend reading pleasure. I will include a few novel activities for you to try during the long hours of lockdown.
Predicting rain with AI
A group of Google Engineers in Amsterdam have developed a Neural Model for predicting the rain that will fall within the next 8 hours. Predicting the weather is fraught with danger. If you are right nobody notices but if you are wrong, even slightly wrong, everyone is a critic. Forecasts are currently based upon physical models of the atmosphere that require massive computational energy and that are sensitive to approximations of the physical laws of nature.
This alternative approach employed a deep neural network (DNN) rather than relying on encoding the laws of nature into a physical model (which, despite our progress are still very complex and difficult to model). DNN’s discover patterns in the data and learn complex transformations from inputs to the desired outputs. Computations can be carried out in parallel on specialized hardware (Graphical Processing Units and Tensor Processing Units which are specifically built for this purpose).
Here is a comparison of a prediction and actual outcome in the US .
The system currently outperforms the current state of the art physics model for predictions up to 7-8 hours ahead and makes a prediction for the entire US at a 1 kilometer resolution in a matter of seconds. The traditional system takes an hour to process a prediction.
The model output is a probability distribution that is used to infer the most likely precipitation rates with associated uncertainties for each geographical region. From these distributions predictions can be made.
The team is continuing to research how to improve global weather forecasting particularly in regions where the impact of rapid changes in weather are most profound.
AI that can detect Lung Cancer from a simple blood test
A group of UK scientists have developed a prototype AI that can detect Lung Cancer. The system analyses the patient’s blood for DNA mutations that drive lung cancer. The system is very much still in development and needs to undergo a clinical trial however if it proves its’ worth at scale, a routine blood test will boost lung cancer screening rates and aid early detection.
The system works by examining free floating DNA that circulates in the blood. The majority of this DNA is from harmless cells however tumors also shed DNA as they form and grow larger. The AI is trained to detect the mutated DNA which is likely to be from a tumor. The system is a screening technology. It doesn’t diagnose lung cancer it identifies individuals that should be further examined.
Currently lung cancer screening tests are carried out by using low dose chest X-rays to check the lungs for tumors. In the US only 5% of people eligible for this test are screened.
In early trials the system showed a 2% false positive rate, while identifying 55% of stage 2 cancers and 70% of stage 3. Currently 60% to 70% of lung cancers are identified at stage 4 (the most advance stage of cancer). Earlier detection makes it easier to save the life of the cancer patient. Lung Cancers have one of the lowest 5 year survival rates (18.1% in the US). The technically minded can read their paper here.
Automated Trucks
Many people believe that Automated Trucks will be on our roads before Self Driving Cars. There is a huge shortage of truck drivers, particularly long haul drivers. The first step in the process of automating trucking is to automate long haul.
Initially this probably mean a person rides inside the truck whilst it is driving but they do not drive the truck. They are there for when the truck stops or breaks down. The “driver” will most likely sleep whilst the truck drives on the highway. Trucks can easily be speed limited (e.g. 80kph) and given they don’t have to have rest stops they can still keep to similar schedules as current long haul drivers.
The likelihood is that Automated Trucks will initially drive between major cities that have good highway connections. The trucks will drive from and to large warehouses located on the outskirts of these cities. Here they will decouple their freight from the cabin and a human will deliver the freight within the city. Driving within cities is complex and still has a lot of development required.
There are a number of companies working in this space. The giants like Waymo (Google’s self driving car company) and Uber are developing automated trucks as are many of the truck manufacturers (Volvo, Mercedes etc.). As a side note, in 2016 Uber purchased Otto, a self driving truck company that was established by Anthony Levandowski (ex Waymo employee). However Uber were sued by Google due to the alleged theft of trade secrets. On March 19 this year Levandowski pled guilty to one of 33 charges of theft, Uber and Waymo have settled their case by Uber granting Waymo an equity stake worth approximately US$245m at the time of settlement.
There are also a number of smaller startups developing the technology required to self drive trucks. Two examples are TuSimple from Tucson Arizona and Starsky Robotics from San Fransico. Starsky claimed to have carried out the first truck trip on a major highway with no human on board in February 2018. In late 2019 however they were unable to find the funding to continue with their development and closed down on 19 March 2020. This is despite being named of the 100 most promising startups to watch in November 2019 by CNBC. Startups can be brutal and they don’t last long if there is no funding. TuSimple is still powering ahead with development.
The A-Z of AI
I recently came across this guide to AI. It is a simple to read and easy to understand guide that will help to make sense of Artificial Intelligence. It is a beginner’s A-Z guide that came from a collaboration between the Oxford Internet Institute at the University of Oxford and Google. The guide is simple enough that young kids can explain it to you in need.
Hacking Smartphones Voice Assistants
Every new technology is soon followed by an inventive way of hacking that new technology. Researchers have discovered that by sending ultrasonic waves through solid materials a hacker can interact with and compromise many of the voice assistants on our phones and computers.
This type of attack is called a “Surfing Attack”. It leverages the unique properties of acoustic transmission in solid materials e.g a table. An attacker can have multiple rounds of interactions with the voice controlled assistant without the need to be in the line-of-sight (there are other line-of-sight laser attacks that have been discovered). Hackers can use the voice assistant to hijack SMS two factor authentication codes and place fraudulent calls without the victim knowing.
Products with a MEMS microphone, standard in many voice controlled assistants, contain a small built in plate call the diaphragm. When the diaphragm is hit with sound or light waves, the wave is translated in an electrical signal that is decoded into the actual commands of the voice assistant. The hacker transmits high frequency sounds, inaudible to the human ear using a $5 piezoelectric transducer attached to the table’s surface. Attacks can be carried out from up to 30 feet away.
To conceal the attack the hacker reduces the volume of the device low enough to make the voice responses unnoticeable whilst still be able to record the voice responses from the assistant via a hidden device closer to the victims computer or phone. Hackers can use Text to Speech systems to make fraudulent calls or send commands such as “read all messages”.
The Surfing Attack was tested with a variety of phones including Google Pixel, iPhones, Samsung Galaxy and Xiaomi all of which were found to be vulnerable. Only the Huawei Mate 9 and the Samsung Galaxy Note 10+ were not vulnerable (the Samsung phone body structure and materials caused the recorded sounds to be too weak to use). Amazon home and Google Echo were not impacted by this attack.
Paying it Forward
If you have a start-up or know of a start-up that has a product ready for market please let me know. I would be happy to have a look and feature the startup in this newsletter. Also if any startups need introductions please get in touch and I will help where I can.
If you have any questions or comments please email me via my website craigcarlyon.com or comment below.
I would also appreciate it if you could forward this newsletter to anyone that you think might be interested.
Till tomorrow’s bonus issue and stay safe from Covid-19.