A venture capitalist recently joked that all you have to do to fund a startup is choose a URL that ends in “.ai”. He wasn’t serious, but companies pursuing AI are getting a lot of attention, and if one of these AI startups develops a “killer app” that earns unicorn status, it will be left behind by the investment community. (FOMO) showed recognition that there is a fear. without them.
As someone who has been developing products for 30 years, one of the most frequent questions my clients ask me is how do I keep my products relevant in a world where AI seems to be growing in almost every industry? is how we apply AI to our products. There’s also the idea that Google, Apple, Amazon, Meta, and Microsoft have all the data, but how can others compete on AI?
The good news about data is that every company has a deep knowledge of the field and has important data relevant to their business that is different from the data held by the top five tech companies listed above. That’s it. The real question is whether the data is in a format that can be accessed to build models that work for your business.
Data classification
Many AI applications focus on classification tasks that require data to be labeled in order to be useful. A classic example would be to collect and label images like Fei Fei Li did, create an ImageNet database containing 100,000 synonyms known as “synsets”, and store 1,000 images of her in each synset. How to create This labeled image database has helped advance object recognition tasks in machine learning applications since ImageNet began working on it in 2009, showing how well AI algorithms recognize cats, dogs, and other objects in images. I am involved as much as I can.
Although the terms artificial intelligence (AI) and machine learning (ML) are often used interchangeably, ML is the process of building models to classify data into subgroups, or using regression to predict temperature, etc. The process of representing data along a continuum of continuous data. . AI is the term assigned to the output obtained from a machine learning model.
Dedicated ML
Current AI models are said to function more like purpose-built appliances than the human brain. This is because each application acquires expertise in a very narrow area, which does not generalize well to other areas. To continue the analogy, both dishwashers and washing machines are specialized appliances, and you would not be happy with the results if you put the dishes in the washing machine. This is also true for most AI models.
As an example, the Amazon Alexa app excels at natural language processing, allowing you to play music, answer questions, set timers, and even tell jokes. However, if Alexa has a camera connected, it will require additional training to recognize cats and dogs. While these AI applications provide users with convenience in confined spaces, none of them reach the level of generalized intelligence typical of 3-year-olds.
In the spirit of building a purpose-built appliance, let’s look at an example of a practical device that tracks the activities of patients who want to improve their health. This activity tracker is a coin-sized sensor tile device with Bluetooth Low Energy (BLE) and a multi-axis inertial measurement unit (IMU) sensor in the user’s pocket. Consider how the software for this product could be developed using traditional methods and ML approaches.
In the traditional model, programmers create features that determine the orientation of an activity tracker before developing features that capture the data generated when different users walk, run, or jump. To do. This task will take a considerable amount of time. The programmer inspects each user’s her IMU data, recognizes how this data translates into her sequence of labeled motions, and tunes the algorithm to recognize each data her sequence as an activity. because there is a need.
When the same activity tracker is built using ML, users are given a mobile app and a sensor tile device to put in their pocket. When the sensor tile detects motion, it notifies the mobile her app to ask what activity the user is doing. User responses “label” each activity, and this is repeated for all users, making it easy to collect a bunch of labeled data from a group of sensor tile devices. At the end of each day, sensor data from all ’tiles’ is fed back into the ML model along with labeled activity from the mobile app, and the revised model is downloaded to all users’ sensor tile units. Each sensor tile now recognizes more activities as the user selects activities and labels the data. As the model becomes more mature, the mobile app will suggest detected activities and allow users to review or correct activities as needed. It essentially crowdsources data labeling, resulting in a more robust activity sensor as more users join its use.
This is an example of a supervised ML application with a classifier that learns to associate IMU data patterns with activities such as walking, running, and skipping. If it was trained with only one user, it might work reliably with that user, but it probably does a poor job of identifying activity from different users. This is a classic case of so-called “overfitting”, where the ML model does not generalize well to other users’ data. For this reason, it is important to collect data from a wide enough range of users, most of which (80%) will go into the model’s training set. The remainder (20%) of the collected data is stored as a test set used to qualify and test ML algorithms, validating functionality against previously untested data.
This is a very simple example, but AI and ML have many other applications. In our next column, we will examine the challenges faced by a major player in the mobile ECG market using ML to analyze his ECG waveforms.
