All about AI/machine learning
helps vehicles navigate and make decisions safely and efficiently. In personalized medicine, machines analyze patient data to allow doctors to better diagnose and treat patients. The ability to process and learn from diverse data sources is crucial for advancements in machine learning. Real-world applications of machine learning Algorithms need data, and there are a wide number of sensors available for use today. Often the outputs of a variety of sensors are combined – a process called sensor fusion. Sensor fusion combines the data from multiple sensors in different proportions to create a dataset product that is greater than the sum of its parts. For example, smartphones combine data from GPS, accelerometers, gyroscopes, magnetometers, and the camera to determine where it is in space. The result allows for a variety of augmented reality experiences, such as Pokemon Go!, or Google Maps indoor navigation.
In automobiles Modern automobiles have dozens of sensors, including
Figure 2: Neural networks are often shown as a diagram similar to the one pictured above. Data arrives and is stored in an input layer (the 4 neurons shown on left). From there, there are ‘hidden-layers’ that form connections to output layers. In this image there are two sets of 6 neurons in the hidden layers, and two neurons in the output layer. The training task determines how input layers are ultimately connected to output layers through the hidden layers
transmission-speed, wheel-speed, steering-angle, suspension-height, tire-pressure, accelerometers, yaw-rate, microphone arrays, and more. I chose to list these sensors in particular because a company called Tactile Mobility fuses the data from these sensors to produce a very detailed picture of road conditions. Application specific algorithms are able to analyze all of the sensor data coming out of a vehicle and determine which vehicle-type produced the data, what condition the tires are in, what road the vehicle was traveling on, and even the specific lane of travel – all without using GPS. In healthcare Healthcare combines several data points together to provide a picture of patient health. A patient’s vitals include: blood pressure, pulse rate, respiratory rate, and temperature. Other data, such as the results of a Complete Blood Count (7 tests) and a Comprehensive Metabolic Panel (14 tests) can identify which
Reinforcement review This basic process of positive and negative reinforcement – increasing or decreasing the likelihood of a behavior – is how humans, insects, and machines learn. Brains make neuronal connections, ants make trails, and computers increase or decrease the values of variables. It is important to train using data that is appropriate for the task and free of unnecessary information when possible. This approach reduces the computational complexity of the training process. Relation to machine learning Machines interact with the world around them through a series of inputs and outputs. The inputs often come from sensors or data files, while the outputs can be screens, actuators, or data files. By training models on data obtained from sensors, you can develop systems that understand and respond to their environments in near real-time. For instance, in autonomous driving, sensor data
the likelihood of an undesirable action being repeated.
groupings, and outliers.
pheromones will dissipate. With a higher pheromone concentration, there is an increased likelihood that more ants will follow. With fewer pheromones, the lower the likelihood that ants will follow that trail.
The image above illustrates unsupervised learning. In unsupervised learning, a large assortment of data is fed into an algorithm whose job it is to find patterns, coincidences, and anomalies. The algorithm sorts the data in a manner that suits it. As an example: Every day, newspapers around the world produce articles on a wide variety of topics. These articles are often published online, which makes them available as raw input data for search engines. Using unsupervised learning, it is possible for Google and other websites to analyze the data and cluster similar articles together for a reader. Reinforcement learning Reinforcement learning is a process where positive feedback is given after a desired action to increase the likelihood of it happening again, and negative feedback is provided to decrease
Learning piano Imagine a student learning to play a piano, but the piano has all its strings cut. Without sound from key presses, there is no feedback, making it impossible for the student to know if they played the correct or incorrect key. Consequently, no reinforcement learning can take place. Now, picture the same student with a properly tuned piano. Each key press produces a sound. If the sound matches the student’s expectations, it serves as positive reinforcement. If the sound is unpleasant or unexpected, it acts as negative reinforcement. Ant tra I ls An ant that follows a pheromone trail to a food supply will strengthen the trail by depositing more pheromones. If the trail leads to an exhausted food supply, the ants will wander elsewhere and the
Early machine learning In the mid-1990s, computer scientists developed TD-
Gammon, an algorithm that learned to play backgammon through reinforcement learning. Initially, it made random moves and received feedback based on the outcomes of those moves. Actions that increased the game’s expected score received positive reinforcement by increasing a reward variable, while actions that decreased the expected score received negative reinforcement by decreasing the reward variable. The algorithm was programmed to maximize the cumulative reward. Through trial and error, it gradually learned which decisions maximize the reward and win the game!
Figure 3: An example of a model that’s ‘Just right!’ and Overfitting
Just right!
Overfitting
we get technical
62
63
Powered by FlippingBook