AI is beginning to make headway in the agri-food sector. Already, AI is increasing our ability to predict yields and optimum harvest dates, create supply chain efficiencies, improve animal health and reduce water pollution. The applications are only going to grow, with a 400% increase in funding for AI projects forecasted over the next 5 years.
With its ability to improve performance, decision making and sustainability, AI has the potential to revolutionise the agri-food sector. This revolution is urgently needed. Increased competition from overseas and a less friendly EU export market are very real threats. Improving our understanding of these technologies is key to remaining competitive.
This article provides covers 13 high-level definitions of artificial intelligence concepts which you may come across.
A subset of AI, machine learning uses algorithms to learn patterns in data (without being given explicit rules). These patterns are then used to predict outcomes in future or unknown circumstances.
The goal is to transform a data set into an easily comprehensible structure, which can form the basis of future analysis. Grocery retailers are a good example; supermarket chains generate terabytes of data, which are exported to data warehouses. Data scientists then build algorithms to analyse this data, to identify and predict customer buying habits which can be used to improve business decision making.
Within our sector, it has also been used for the automation of tasks, identifying in-season crop type from satellite imagery, predicting disease outbreaks and market segmentation.
An algorithm is a procedure or formula used by a computer to solve a problem. Typically, they will follow a specified sequence of actions to produce an output. Search engine's, such as Google, are probably the best-known example of an algorithm. Based on your search term, the computer will search through published webpages and return the results to you; the order of results is defined by comparing each page with a predetermined ranking criterion. In the agricultural sector, algorithms serve a variety of purposes, including forecasting yields, harvest dates, and optimum windows for spraying
Labelled vs unlabelled data:
In data science, there is a distinction between data which is labelled and data which is not. Data are considered labelled when the correct output is known. For example, we may be trying to identify or count sheep within an image. The labels are used to train our models to detect sheep on new/unlabelled images. Often, labelling data requires human intervention, which is expensive and time consuming, especially as modern techniques such as neural networks require a lot of data. Data without these labels are referred to as raw – or unlabelled – data.
Training data are labelled data used to train a machine learning algorithm. The data sets will be complete with labelled inputs and the expected outputs for a range of possible variations. The algorithm will analyse these variations to build an understanding of how the inputs effect the outputs. Once equipped with this knowledge, the algorithm can produce accurate outputs when presented with new, labelled data.
Supervised Machine learning:
Supervised machine learning is one of three types of machine learning. It deals with labelled data and learns the patterns in the data to produce a correct output.
This type of machine learning has been to identify field boundaries and predict in-season crop types from satellite imagery.
Unsupervised Machine learning:
Unsupervised machine learning involves training an algorithm without the use of labels on the data. The applications for unsupervised machine learning are narrower than with supervised machine learning, typically involving pattern recognition and association. Because it requires no human labelling, Unsupervised machine learning is significantly cheaper and is the majority of data collected.
Semi supervised machine learning:
Semi-supervised machine learning is useful when there is only a small proportion of labelled to data and not enough to use supervised techniques. Unsupervised learning is used to cluster the data into groups and the sparse labels then used to inform the remaining unlabelled data. Supervised learning can then be used to gain more accurate result from the larger dataset. This is useful whenever it is difficult or expensive to label large numbers of examples such as images or documents.
Transfer learning is when a machine learning algorithm is trained for a particular problem, can be applied on a different but related situation. For example, some of the features used by the model when identifying cattle, could be useful when identifying species of sheep.
In this context, black box have nothing to do with aeroplanes. Rather, models are described as black boxes when they involve complex calculations that are difficult to reverse engineer.
As an analogy, although we can approximately know the input ingredients that go into making a cake, it is very difficult to know the exact weighting and proportions of the ingredients once the cake is made and which specific/combination of ingredients are creating a particular characteristic. This can involve complex mathematical equations or formulas involving hundreds of variables. Perhaps the most widely publicised example of this are the newsfeeds of social media platforms such as Facebook and YouTube. Here, the original designers are unaware of the exact process the algorithm used to recommend a piece of content. In this case, it has led to negative results, i.e. the recommendation of extremist content. On the other hand, these large models can produce incredibly accurate solutions to some incredibly complex situations.
Within this branch of AI, originally developed in the 1950's, the algorithms were inspired by thehuman brain. Neural networks work by breaking complex problems down into simplified stages or components of the final result. For example, a network could break down handwritten letters by identifying the individual loops and straight lines separately. How these combine, creates specific letters.
Neural networks and their relations are at the forefront of the current leaps in AI research and evidence-based decision making.
Coupled with increasingly powerful computers, this sub-group of neural networks algorithms work particularly well with the unstructured and non-traditional forms of data, such as images, videos and recordings.
Related to neural networks, an algorithm is considered to be 'deep' when input data goes through multiple hidden layers, each looking at a different aspect of the structure of the data.
In practice, this can facilitate an evolution in the capabilities of AI. For example, rather than simply identifying cattle in a single step, a convolutional neural network may first detect the edges, a deep learning algorithm may begin to recognise a hoof. When asked to identify other species of animal, the algorithm can begin to classify animals into ungulates, as well as their sub-species.
Natural language processing:
Natural language processing is the subset of AI used to interpret human communication. This is a complex form of learning which can utilise advanced neural networks. This form of AI powers translation services and digital assistants such as Siri and Alexa as well as categorising text documents and key word selection.
An expert system is a computer program which uses AI to attempt to simulate the judgement of a human expert. Typically, it will incorporate a large knowledge base full of data, experience (informed by supervised learning and user feedback) and a collection of rule sets which apply to various scenarios. The performance of this system will continue to increase over time, as its judgements are corrected, and more data is added. In our own industry, it's possible that one day expert systems may be able to offer advanced agronomic and veterinary advice, make on-farm management decisions, and streamline purchasing processes and supply chains.