Tech NewsHow to make AI greener and more efficient

DarylJanuary 10, 2022
Wirth Research, a computational fluid dynamics engineering firm, has become increasingly concerned with environmental sustainability.

It initially focused on racing car design, allowing clients to replace costly wind tunnel work with computerized modelling, but in recent years it has designed equipment that reduces the aerodynamic drag of lorries, as well as a device that reduces cold air escaping from open-fronted supermarket fridges, reducing energy use by a quarter.

The Bicester-based company also wanted to reduce the energy consumed by its detailed computerized modelling, which simulates around half a billion tiny cells of air for car aerodynamics. It had already changed the cell resolution within each model, with a finer sub-millimetre mesh used near sharp edges.

Wirth then moved its computing from its own site to a renewable energy-powered data center in Iceland run by Verne Global during the pandemic, when it realized staff could work effectively from home. The new hardware has reduced energy consumption by three-quarters, and the power used is carbon neutral.

According to engineering manager Rob Rowsell, the total cost of the new equipment over several years, including the use of the Icelandic facility and connectivity, is less than the old UK power bill. Furthermore, in order to continue with hybrid working, the company has relocated to smaller offices in an environmentally friendly building.

Wirth wishes to improve the efficiency of its computing processes. It can already halt iterations of virtual models when they stabilize rather than running them a fixed number of times, but it is investigating how it can use artificial intelligence (AI) trained on previous work to predict a stable version of a model that would normally take much longer to reach in a handful of iterations.

The prediction would not have to be completely accurate because the company would then run a few more iterations to ensure the model was stable. “You’d be able to do 15 or 20 iterations rather than 100,” Rowsell says.

According to Peter van der Putten, director of decisions and AI solutions at Massachusetts-based software provider Pegasystems and assistant professor at Leiden University in the Netherlands, there is a lot of potential for using AI to combat climate change.

However, in recent years, AI has increasingly meant the use of deep learning models that require a large amount of computing and electricity to run, such as OpenAI’s GPT3 language model, which was trained on nearly 500 billion words and 175 billion parameters.

“Until recently, it was fashionable to come up with yet another larger model,” van der Putten says. However, environmental considerations are emphasizing the benefits of improving AI efficiency, with rising electricity costs providing additional economic justifications. “From both a financial and a climate standpoint, small is beautiful.”

Another reason is that simpler, more efficient models can yield superior results. Van der Putten co-organized a challenge in 2000 in which participants attempted to predict which customers of an insurance company would be interested in buying caravan insurance based on dozens of variables on thousands of people.

This included real-world noisy data, which can cause complex models to fail. “You might start seeing patterns where there aren’t any.” “You start to overfit data,” van der Putten says. This issue arises when the training data and the data for which predictions are required are not exactly the same – for example when they cover two different groups of people. Simpler models work well when there are clear relationships or only a few data points.

It can also be difficult and costly to revise large models that have been trained on massive amounts of data. Lighter “online learning” models designed to adapt quickly based on new information may be the best option for evolving situations, such as allocating work to a group of employees with many joiners and leavers.

According to Van der Putten, these models are not only less expensive and have a lower environmental impact, but they are also easier to interpret. There is also the option of using traditional machine learning algorithms, such as support vector machines, to classify items, which tend to be lighter because they were developed in times when computing power was much more limited.

According to Van der Putten, AI specialists divided into tribes favouring specific techniques in the late 1980s and early 1990s, but practitioners eventually learned to use different approaches in different situations or combine them. “Returning to a more multidisciplinary approach would be beneficial,” he says now, given that alternatives to big data-driven deep learning generally require much less computing power.

Got to start somewhere

According to Jon Crowcroft, professor of communication systems at Cambridge University and founder of Cambridge-based data discovery firm iKVA, one option is to give AI models a starting point or structure.

Language models, rather than being based on the analysis of billions of words, are used to include structural rules, and science-focused models can benefit from having relevant principles programmed in. This is especially true when analyzing language, videos, or images, where data volumes are typically very large.

For example, if given a sample spike shape, an AI system could learn to identify coronavirus spike proteins more efficiently. “Instead of having zillions of images and someone labeling them, you have a ground truth model,” Crowcroft says.

He goes on to say that this approach is appropriate when each result has significant ramifications, such as with medical images. It may require specialists to provide initial material, but this may not be a significant disadvantage if those setting up the model are experts in the first place, as is likely in the case of academic use. Such early human input can significantly reduce the computing power required to develop an AI model while also making the model easier to explain.

Carrying out AI locally would mean much less data being sent across networks, saving power and money

It could also help to shift where and how AI works. In contrast to current meters, which send usage data every few minutes, a federated machine learning model could involve genuinely smart meters analyzing a customer’s electricity use and sending an occasional update of a resulting model to the supplier.

“The electricity company cares about having a model of everyone’s use over time,” Crowcroft says, rather than what each customer is doing in near real-time.

Local AI would require much less data to be sent across networks, saving power and money, and would provide greater privacy because detailed usage data would not leave the property. “You can flip the thing around,” Crowcroft adds. This type of “edge learning” could be useful for personal healthcare monitors, where privacy is especially important.

Reducing the energy required for AI

There are ways to make a centralized deep learning model more efficient if one is required. TurinTech, a London-based code optimization specialist, claims it can typically reduce the energy required to run an AI model by 40%. According to chief executive Leslie Kanthan, if a slightly less accurate fit is acceptable, much greater savings are possible.

A model trained on past financial trading data cannot predict its future behavior in the same way that overfitting, or fitting a model to a specific group of people who make up training data, cannot. A simpler model can provide good predictions, be much cheaper to develop, and set up and change much faster – a significant advantage in trading.

TurinTech’s optimiser employs a hybrid of deep learning and genetic or evolutionary algorithms to adopt a model based on new data rather than recreating it from scratch. “It will attempt to bend the deep learning model to fit,” Kanthan explains.

According to Harvey Lewis, an associate partner at Ernst and Young UK and the chief data scientist of the firm’s tax practice, evolutionary algorithms and Bayesian statistical methods can help make deep learning more efficient. However, it is common to use brute force to tune parameters in models, running through vast numbers of combinations to see what works, which “is going to be pretty computationally expensive” for billions of parameters.

According to Lewis, the costs of such work can be reduced by using specialized hardware. Graphics processing units (GPUs), which are designed to perform calculations quickly in order to generate images, outperform general-purpose personal computers. Field programmable gate arrays, which can be configured by users, and tensor processing units designed specifically for AI by Google are becoming even more efficient, and quantum computing is poised to advance even further.

However, Lewis believes that it is necessary to first determine whether complex AI is actually required. Deep learning models excel at analyzing massive amounts of consistent data. “They are excellent at the specific task for which they have been trained,” he says. However, in many cases, there are simpler, less expensive alternatives that have a lower environmental impact.

Lewis prefers to start with the simplest AI model that can generate a reasonable answer. “Once you’ve got that, do you need to go any further, or does it provide everything you need?” he asks. Aside from saving money, electricity, and emissions, simpler models like decision trees are easier to understand and explain, which is a useful feature in areas like taxation that must be open to checking and auditing.

He goes on to say that combining human intelligence with artificial intelligence is frequently advantageous. Before automated work begins, this can include manual checks for basic data quality issues, such as whether fields marked as dates are recognisable as such.

It is often more efficient to divide processes between machines and humans, with software performing high-volume sorting such as spotting images with dogs and humans performing more difficult judgments such as classifying the breed. “Integrating a human into the loop is a way to improve performance and make it much more sustainable,” Lewis explains.

Leave a Reply

Your email address will not be published. Required fields are marked *