All about AI/machine learning
for training. Normalizing the data provides mathematical reductions in the computational complexity and reduces the amount of required memory since all values will fall in a known range. Finally, split the data into two batches - one large batch for training the algorithms, and a second smaller batch for testing the quality of the training. Training and deployment resources Training often requires a substantial number of addition and multiplication calculations and a lot of available memory. At times you can train a model at home on powerful computing devices such as the NVidia Jetson series of Single-Board-Computers or graphics cards, and other times training involves server farms at data-centers. Training will take anywhere between a few hours and a few months depending on the complexity of the model and the available processing power.
ant analogy, lots of ants are initially needed to find a food source, many more are needed to refine and streamline the trail, but once the trail is established, a single ant can mindlessly follow it without much effort. Training tips For neural networks, the data has to be sufficiently shuffled to properly train. If you were to feed a digit- recognition model six-thousand images of the number 0, followed by six-thousand images of the number 1, then 2, etc. the neural net would not properly form, and no useful predictions could be made. Proper shuffling ensures that the model learns to generalize from the data rather than memorizing the order of the inputs. The scatterplots above are often used to illustrate the goal of training. Algorithms attempt to find models that fit the data reasonably well without over-training. Overtraining means the model will detect the training data with high accuracy, and the testing data with very low accuracy. It is important not to over-train a model. There is always a bit of statistical variation, gaussian noise, or error in the datasets. A properly trained model will ignore the noise of the training dataset. An overtrained model will learn that the noise is important, and part of the dataset. As an example, imagine you’ve got a list of temperatures,
measured at noon at a single location over the course of a year. A properly trained dataset will fit a smooth sinusoidal curve to the scatterplot, while an overtrained dataset will resemble a scatter plot connected with sharp lines. Summary Machine learning is a powerful tool, but it is task specific and lacks the general intelligence seen in Science Fiction movies and fantasy novels. The models used to generate outputs are dependent on the quality of the training process and are not capable of generalization to new areas - a model trained to correct grammar cannot be used to identify license plate numbers in security footage. In contrast, the human brain has no such limitations - its ability to identify, abstract, and generalize the world around it is seemingly boundless. The hopes of having an Android companion are well beyond the limits of technology for the foreseeable future. In recent years, advances in computers have made training and deploying models accessible and affordable to everyone, democratizing AI for everyone with even basic programming skills. To learn more, please see the excellent work done by our very own Shawn Hymel here at Digikey.com!
Machine learning is a powerful tool, but it is task specific and lacks the general intelligence seen in Science Fiction movies and fantasy novels.
Data collection Substantial amounts of data are required to properly train a model. In the handwriting example mentioned at the start of the article, sixty-thousand of the seventy- thousand handwritten digits were used to train a detection algorithm. Data preprocessing Computer scientists enjoy saying “Garbage in equals Garbage out.” Which means the data fed into an algorithm will impact its later performance. For this reason, if you’re training an algorithm to detect handwritten digits, you cannot accidentally include a few hundred photographs of your cat or newborn child. You must evaluate the data for missing values, outliers, duplicates, etc. to ensure that the best possible dataset is fed to the computer
process that requires significant amounts of data and processing power. The steps include: data collection, error detection, and deployment. Neural networks, a type of processor, feed sensor data into an input layer. The input layer then passes the data to a series of hidden layers containing neurons, and eventually to the output layer. The computationally intensive part of training involves determining the values of the individual neurons. 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 blue neurons shown on left). From there,
organs might be functioning poorly and might require further testing. Doctors learn to interpret the data and generate results after years of intense education and practice. But researchers also train computers to fuse the data from the various sources and produce a list of likely causes of an illness. As another example, amputees today have more options than the peg-legged pirate of centuries past. Modern prosthetics now can include accelerometers, gyroscopes, magnetometers, pressure sensors, force sensors, and electromyography (muscle) sensors to provide enhanced capabilities and natural movement. The data from the sensors allows a prosthetic to know where it is in space and interact with the nervous system without having to merge with the nervous system. Looking forward Now, it is time to learn how neural networks work, and hopefully demystify the process enough that you’ll be confident enough to try it out on your own.
there are multiple ‘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
One reason training is so computationally expensive compared to deployment is
the sheer volume of data used in training as well as the need for both forward and backward propagation to determine the value of each neuron. Once training determines the value of each neuron, deployment is far more straight-forward. Remembering the
The learning process
Machine learning is an involved
we get technical
64
65
Powered by FlippingBook