Allegro’s licensable neural networks can speed up machine learning cases
Allegro.AI – an early-stage machine learning and vision company based in Tel Aviv – has announced a major strategic partnership with NetApp – a US$16 billion veteran in the data storage and cloud services business. Allegro was established in 2016, and in 2018 raised over US$11 million from a consortium of backers which included Samsung, Hyundai and Bosch.
At present every business that wants to build computer vision into its system has to develop its own neural-network-created suite of “learnt objects”. In reality, the network is not learning “objects” but is instead deducing a set of biases and weights to apply to data at each neuron connector in the neural network. The final set of weights and biases is the one that generates the fewest incorrect identifications at the output layer. Even a really “intelligent” AI system is not actually intelligent or “aware” – it is simply a really well-tuned set of mathematical functions.
So a business dealing with cats would have to teach its system how to “see” cats, and one dealing with widgets would have to do the same for widgets. The “learning” process requires “backpropagation”. Backpropagation is the process in which a given set of weights and biases is tested against a large body of examples (multiple “is this a cat?” images), and the multiple results are then scored by humans. The results create a new function – the average “cost” or “wrongness” of that set of biases and weights.
The neural network uses these “costs” to adjust its biases and weights, using the scale and direction of “wrongness” to suggest what those changes should be. Wrongness is identified mathematically as “the gradient between where we are and where we want to be”.
The maths to solve those gradients perfectly is complex and slow to compute. While the adjustment process is executed by open-source software (a “framework” in neural net parlance), Allegro’s software controls how the inputs to the framework process are selected and managed.
The framework uses estimated steps to test changes to biases and weights to reduce “wrongness”, and so find the optimised solution stochastically. The neural network, with its new set of biases and weights, is run again on a selected validation data set. The process is repeated until the “cost”, or “wrongness”, is minimised, because the weights and biases have been optimised. This optimisation is the “learning” process.
The human cost
So neural network development uses (expensive) humans to score the network’s decisions – this is a cat, this is not a cat, this is a widget, this is not a widget – to provide the input for backpropagation, and so allow the neural network to tune itself to cats or widgets.
The process that is generally called “learning” is in maths simply an algorithm seeking to minimise the “cost function” of being wrong. Of course, the learning process is complex and expensive. A given object may have multiple angles of view, lighting conditions and colours, and all of these will need to be scored (or learnt).
The scale of this task jumps into focus when one considers that a given application (say an oil rig) will have thousands of objects with tens of thousands of view-angles and lighting states. So tens of millions of images must be processed, scored and learnt.
Allegro has set out to make two changes to this process. First, it has created neural networks that are pre-tuned to specific business segments, which can be accessed and licensed with an Allegro licence. So instead of everyone inventing their own machine-vision wheel to recognise cats, the idea is that they license Allegro’s wheel.
Secondly, by using these highly tuned neural networks Allegro reduces the need for human classification by (it claims) around 90%. The customer sets a target probability of “correct identification”, which generates a smaller number of “edge cases” for human classification. Obviously, the higher the probability threshold set by the customer, the more edge cases are generated.
In fact most users do not develop their own neural networks, but download ones already created by others (often in academia). The problem that flows from that choice is that the downloaded network is not “trained” on the user’s specific problem/image set.
Training takes a large quantity of expensive specialist time. Part of Allegro’s solution is to “associate” the downloaded network to the target data set in minutes. Allegro’s objective, therefore, is not to be a provider of neural networks (in fact Allegro’s networks are provided free to its customers) but rather to be the source of powerful infrastructural tools that wrap around the management of the product lifecycle (both images and sets of biases and weights).
From collecting the data and adding insights, to bringing version control to the network training process, to managing the image data, to quality assurance, and into post production, Allegro has created a suite of building blocks to allow the customer to focus on creating the right network for its task.
While Allegro accepts that it is treading a path already trampled by the FANG community, it believes that many solutions developers will be wary of resting their business plans on tech owned by such large and aggressive operators, which might ultimately take the view that the markets created belong to them, not to their partners (and that the data uploaded belongs to them too).
Allegro also believes that the large datasets owned by the FANG group are thin when it comes to machine-vision data, so why partner with them in the first place? Finally, the use of cloud solutions requires customers to upload and manipulate petabytes of image data – slow, cumbersome and expensive, and fraught with version-control traps. When a customer uses Allegro the customer’s datasets stay firmly and safely under customer control. How big is a petabyte? To gain an accurate impression, a 50 GB line would transmit one petabyte of data in 44 hours.
Allegro’s objective is substantial – it aims to become nothing less than the centre of a new machine-vision ecosystem, an operating system if you like, which leverages development across a multitude of sub-markets to make each new machine-vision step easier, quicker and cheaper.
Applying machine learning to multiple sectors generates those multi-petabyte databases and datasets. These may be located in multiple physical places – at the “edge”, close to where they were collected, in multiple private data centres, or even in cloud accounts of the customer’s customers. Allegro abstracts this location issue so that the customer’s product teams do not need to worry about where the data being worked on is physically hosted.
Allegro’s partnership with NetApp is designed to enable the transmission of these datasets repeatedly (sometimes billions of times) into powerful graphics processors. The data-bus can become a critical bottleneck. Allegro believes that combining its toolset with NetApp’s hardware caching skills will allow network training to be much faster and more processor-efficient.
Allegro includes the oil and gas industry in its set of target markets. Applications already in sight include analysis of geological data, interpreting drone and satellite images to monitor remote installations and pipelines, for tracking assets around large facilities, and even for tracking whether staff are complying with PPE rules out of sight of supervisors.
However, Allegro is still in the early stages of building its sales and distribution network, and has yet to find an oil industry distribution partner.