Kangaroo Court: Examining the Image Data | Association of Certified Electronic Discovery Specialists (ACEDS)
Image classification AI can find existing application in document review. Examples include identifying documents containing specific logos, separating our designs or patent data, and identifying relevant oilfield maps and images. The sophistication of legal AI solutions has grown to include increasingly pragmatic image classification solutions for litigation and compliance functions. The next year or so will see greater thought leadership towards the application of image classification tools and the shift from early adopters to the more pragmatic side of the market. Once technology is aligned with delivery, courts will begin to expect higher standards of sophistication from litigation teams in using these tools. With that in mind, let’s take a look at Image Classification AI, so that we can understand not only what Image Classification AI is, but also how it works.
Your brain is constantly processing the data you have observed from the world around you. You absorb the data, you make predictions about the data, and then you act on it. In vision, a receptor field of a single sensory neuron is the specific region of the retina in which something will affect the triggering of that neuron (i.e., it activates). Each sensory neuron cell has similar receptor fields, and their fields overlap. When light hits the retina (a layer of light-sensitive tissue at the back of the eye), special cells called photoreceptors turn the light into electrical signals. These electrical signals travel from the retina to the brain via the optic nerve. Then the brain transforms the signals into images that you see.
It is important to note that the concept of hierarchy plays an important role in the brain. The neocortex stores information hierarchically, in cortical columns or evenly organized groupings of neurons in the neocortex. 1980, a hierarchical neural network model called neocognitron was proposed by Kunihiko Fukushima, who was inspired by the concept of simple and complex cells. This would form the basis for further research towards the development of image classification AI.
Pretty simple stuff, right? Well not really. So how about reframe the idea and put the anatomy lesson aside so we can think of it like we are the computers. When we see an object, we label it based on what we have already learned in the past. Essentially, we trained on data over a long period of time so that buildings, maps, people and other objects were all clearly classified according to their respective group. The world is data training. Our eyes are the input function, and between the eyes and the brain is a series of dials that process and refine the data into a recognizable final image in our brain. If you’ve never seen a car before, you don’t have any training data to refer to. This data includes the height, width and depth of the physical world. It also includes colors, shadows, and movement. This is where the inspiration for image classification AI models comes from.
Convolutional Neural Networks (CNNs) are a subfield of deep learning. Do you remember AlphaGo? Deep learning (also known as deep neural networks) is a subfield of machine learning. What makes deep learning unique is the way it mimics the workings of the human brain in processing data and creating models to use in decision making. These artificial neural networks use neural nodes that are connected like a layered web between an input layer and an output layer. While traditional programs build analysis with data in a linear fashion, the hierarchical function of deep learning systems allows machines to process data with a non-linear approach.
CNNs have a different architecture than traditional neural networks. Regular neural networks transform an input by passing it through a series of hidden layers. Each layer is made up of a set of neurons, where each layer is fully connected to all neurons in the previous layer. Finally, there is a final fully connected layer – the output layer – which represents the predictions.
CNNs are a little different. To begin with, the layers are organized in three dimensions: width, height and depth. In addition, neurons in one layer do not connect to all neurons in the next layer, but only to a small region of it. Finally, the final output will be reduced to a single vector of probability scores, organized along the depth dimension.
In CNNs, the computer views an input image as an array of pixels. The better the image resolution, the better the training data. CNN’s goal is to enable machines to see and perceive the world as closely as possible to the way humans do. The applications are wide, including image and video recognition, image analysis and classification, media recognition, recommendation systems, and natural language processing (among others).
CNN has two components: feature extraction and classification. The process begins with the convolutional layer, which extracts features from the input image. This preserves the relationship between the pixels by learning the features using small squares of input data. Convolving an image with different filters allows operations such as edge detection, blurring, and sharpening to be performed by applying filters. Next are the strides, or the number of pixels moving on the input matrix. If you had a photo of a tiger, this is the part where the network would recognize its teeth, stripes, two ears, and four legs. During classification, layers of connected neurons serve as a classifier on top of the extracted features. Each layer assigns a probability that the object in the image is a tiger.
Major service providers will see this technology as an opportunity to develop their service capabilities in other areas of the legal arena, such as compliance monitoring, contracts and meeting regulatory standards. More generally, as the police gradually introduce this technology, the expectations of the courts will evolve in parallel. Currently, there are limits to how some agencies and business functions can take advantage of image classification AI, but it is ideal for widespread adoption in the eDiscovery space as companies leverage these algorithms to deliver image learning review solutions to their clients. This is a game-changer for technologically advanced companies as it opens up the potential for new revenue streams in potentially more profitable applications.