I don't claim to be an expert on machine learning and AI but there is certainly no shortage of hyperbole about it right now. But in this case for good reason I think. It's on every trends/prediction list you read but it is surely the comprehensiveness in which it will be integrated into organisational capability, customer experience (and so competitive advantage) that makes this a true mega-trend. As Fred Wilson describes it 'AI will be the new mobile' (and he should know).
So it's both refreshing and insightful to actually find something that talks about the challenges and limitations that machine learning still faces. Just after Christmas Azeem Azhar (whose Exponential View is an excellent regular read on all things AI) tweeted a photo of a slide that neatly captured some of the current limitations of deep learning. Sadly he wasn't able to attribute the source, but the list focused on aspects such as how data-hungry it can be (requiring millions of examples), and how computationally intensive it can be to train and deploy, the fact that it can be limited in its representation of uncertainty and easily fooled by adversarial examples, the complexity in optimisation (needing high degrees of skill and experimentation) and notably the lack of trust that can come from poor transparency and 'uninterpretable black boxes'.
One of the respondents to Azeem's tweet pointed at this piece by Peter Voss which expands further on current limitations. Whilst current mainstream techniques can be very powerful in narrow domains, they will typically have some or all of a list of constraints that he sets out and which I'll quote in full here:
- Each narrow application needs to be specially trained
- Require large amounts of hand-crafted, structured training data
- Learning must generally be supervised: Training data must be tagged
- Require lengthy offline/ batch training
- Do not learn incrementally or interactively, in real time
- Poor transfer learning ability, re-usability of modules, and integration
- Systems are opaque, making them very hard to debug
- Performance cannot be audited or guaranteed at the ‘long tail’
- They encode correlation, not causation or ontological relationships
- Do not encode entities, or spatial relationships between entities
- Only handle very narrow aspects of natural language
- Not well suited for high-level, symbolic reasoning or planning
Peter usefully goes on to detail a machine learning intelligence checklist of key features.
Reading about the challenges as well as the opportunities with machine learning gives a far more rounded perspective of where we are at, and brings to life the reality of the hurdles we still need to overcome. I've no doubt that many of these challenges will be surmounted in time (perhaps some in short order given the degree of focus on the area) but my overwhelming feeling about this is that it will result in a huge demand for people who can design, build and make sense of this kind of capability. Even if, as an organisation, you can plug into API-accessible machine learning capability or access open-source libraries of machine intelligence (like Tensorflow), you still need to be able to understand where the value is, and design elegant solutions and applications. The past few years has seen a huge growth in demand for data/analytics talent at a time when supply cannot keep up. I think we'll see the same for those skilled in not only the technical side of machine learning, but particularly those who can combine technical expertise with the kind of strategic and empathetic intelligence that can effectively mine the real value that is inherent in this new capability.