382 Little Collins Street
Melbourne VIC, 3000
(03) 9008 5922
Deep Learning is a form of artificial intelligence that automates repetitive tasks to save time, save money, and make new innovations possible. It involves training computers to perform a task by drawing from examples.
Deep Learning has become more accessible over the last 12 months, driven by the availability of data and developments in computing power. Machines are now able to process data quickly enough to make the process worth the investment.
Here we briefly explain what Deep Learning is, how it adds value to organisations, and how you can get started.
Deep Learning automates tasks involving the senses, training models to understand raw perceptual data such as images, video and audio. This allows us to automate tasks which were previously costly and time consuming for humans to perform, such as:
(identifying what’s in an image)
(identifying what’s in an image, and where it is)
(identifying what’s in an image, and which pixels it’s in)
(generating an image)
(labelling a given text)
(generating a paragraph or sentence for a given document)
Language Modelling (generating text eg. a new article headline)
Questions Answering (Answer a user’s question from text corpus)
(identifying what’s in an audio sample)
Speech Recognition (converting audio to spoken words)
(identifying a pattern in time series data)
(predicting the next value in time-series data)
(generating or simulating time-series data)
Every organisation starts from a different point, but we suggest everyone trials the process on a small experiment centering around a problem that has the potential to have a sizeable impact on the business. Working through the following points will help guide your thinking
When selecting a problem to solve, define it carefully. Is it specific, measurable, achievable, relevant and timebound? Identify the impacts that a solution could have on your organisation, in terms of time saved, money saved, and reputation
Custom solutions are ideal if:
> Your use case is novel (off-the-shelf solutions only support a few common problems)
> Your data set centers around proprietary assets, and/or you want to own the trained model as IP
> Your use case is novel
> Your data set is highly sensitive, and requires a system to be built around it
> You have access to Machine Learning experts, either on staff or as hired consultants
Non-custom solutions are available from the big players like Google and Amazon. These are ideal if:
> Your use case is common
> You don’t mind about owning the resulting trained model as IP
> Your data is not sensitive
> You have a relatively small budget
> Your company doesn’t have, or can’t afford, Machine Learning experts on staff or as hired consultants
The more examples they have to learn from, the better a Deep Learning model will perform. For a computer vision project, a minimum would be hundreds to thousands of quality images. By quality, we mean relevant to the outputs of the project, rather than high resolution.
If you do not have a data set ready to go, the cost of collecting data will be key to the project. It is also worth investigating ready-made data sources available online, both for purchase and for free.
Investing in your own GPU workstation is a great option, both for proof of concept and long term projects. This option is especially useful when your data set is sensitive. Downsides include managing the server, dealing with spike failures, and changed priorities.
Cloud computing power is an option, however it is generally more expensive. Utilise this option if you are experimenting on a lightweight product. This option is also attractive to those whose data is already hosted with the cloud provider.
While the theory of deep learning has existed for many years, there has been a resurgence of activity as more powerful computer hardware (GPUs) and larger datasets needed to train models have become available.
As a result, deep learning methods have recently achieved state-of-the-art (in some cases human-level or better) results in a large and varied number of fields including image processing, natural language processing, time series modelling and predictive analytics, as well as game playing and decision making.
Now is your chance to become an early adopter of this technology.