2023 Author: Bryan Walter | [email protected]. Last modified: 2023-05-21 22:24
Waymo has announced a partnership with DeepMind to train neural network algorithms for self-driving cars. Now, for some of the algorithms, Waymo uses a technique developed in DeepMind, in which the training hyperparameters are selected in parallel on a set of models using a method reminiscent of the evolutionary development of living organisms.
The control system of an unmanned vehicle is based on neural network algorithms, the efficiency of which directly depends on the quantity and quality of training data. It is because of this that leading developers like Waymo are expanding their car park and testing them, driving millions of kilometers on real roads and billions of kilometers in simulation. However, data itself is only part of the prerequisites for making a car safer than humans on the road. Equally important is the structure of certain neural network models, as well as how they are trained.
There are two types of parameters in neural network algorithms. One kind is the direct parameters of the neural network that change during training, for example, the weights of the neurons. Another view is hyperparameters. They are responsible for how the learning takes place. For example, one of the key hyperparameters is learning rate, that is, how quickly the neural network adjusts its parameters during training. At the same time, the learning rate must be kept at a balanced level, because if the learning rate is too low, it will take a lot of time and computational resources, and if it is too high, the parameters can change dramatically and, as a result, never reach the optimal value.
Typically, the process of selecting hyperparameters occurs semi-automatically. During it, many neural network models are trained in parallel and the hyperparameters for each of them are randomly selected, after which the best trained models "win". In 2017, specialists from DeepMind, which, like Waymo and Google, is part of the Alphabet holding, proposed a significantly improved training method, which is now used to train drone algorithms.
The method can be thought of as evolution. Initially, the models start parallel training with a random set of hyperparameters. After a short time, the worst models from the “population” are replaced by a new generation - copies of the best models with slightly modified training hyperparameters. At the same time, the copies completely inherit the state of the parent model, so they do not have to be retrained from scratch and spend resources on it. Since some hyperparameters may not provide the final good result quickly, the researchers implemented the division of the entire "population" into isolated "subpopulations", competing only with each other, similar to how real evolution occurs on islands. In addition, during each training segment, the models are not trained in complete isolation, but can "peep" the hyperparameters of more successful models.
Waymo tested a learning method with an algorithm that highlights areas of pedestrians, cyclists, and motorcyclists from cameras. The algorithm trained in this way ultimately had a 24 percent lower false-positive rate, in which it recognizes a target where it actually does not exist. At the same time, half of the computing resources were spent on training.
Earlier this year, the California Department of Motor Vehicles released annual statistics on the frequency of interventions by test engineers with self-driving cars while testing them on the streets. Just like last year, Waymo was the leader in this indicator. Its cars, on average, cover almost 18,000 kilometers without the need for intervention. However, this data must be taken with a grain of salt, because the official statistics include only cases that, as a result of the simulation on the simulator, were capable of leading to a serious accident if the engineer had not taken over control.