Artificial Intelligence – Cognition conundrum

Deep Learning’s limitations laid bare.

• Work on analysing how deep learning neural networks work clearly shows that these systems have no real understanding of what it is they are doing.
• This means that they are going to fall far short of the still rampant AI hype, triggering what I believe will be the 3^rd AI winter.
• Deep learning is essentially the use of large neural nets with many layers (hence deep) to match characteristics of data to desired outcomes.
• The problem has long been that large neural nets are badly understood meaning that improving on the technique is difficult.
• This is why considerable effort is being put into understanding how they work.
• A recent research paper (see here) has attempted to do this by training a neural net to recognise images and then looking at what the neural net could “see” to ascertain how it was making its identifications.
• The first thing that is obvious from the results is that the neural net has no understanding of what it is being presented with but is merely matching characteristics to outcomes.
• For example, the network can easily distinguish between a whale and a shark fin above the water line but adding a picture of a baseball and glove in the corner of the photo completely throws the algorithm off.
• Another example is the discovery that dog breeds are mostly distinguished by the floppiness of their ears.
• This characteristic exists in dogs as a distribution meaning that there will be overlap between breeds as well as potential for confusion with other animals that have floppy ears.
• There are two conclusions to be drawn from this research:
  • First, the good: This research is a step forward in understanding how neural networks and deep learning works.
  • If the process can become well understood, then there is scope to improve the techniques and make deep learning better.
  • This will help to keep AI improving, albeit increasingly incrementally, making it better and more valuable for the tasks at which it excels.
  • Second, the bad: This research lays bare one of the fundamental flaws of deep learning which is that it has no casual understanding.
  • Deep learning can’t reason nor can it extrapolate meaning that the hype of general intelligence being just around the corner will be, once again, unfulfilled (as it was in the 1970s).
  • It points very strongly to RFM’s conclusion (see here) that deep learning had fundamental limitations meaning that new techniques are needed for AI to really take a big step forward.
  • Of these techniques, there is no sign, meaning that the pace of progress in AI must inevitably slow materially as deep learning reaches its limits.
• The real problem here is that AI is now so popular that companies raising money can get higher valuations if they are perceived to be AI companies which has driven hype to fever pitch.
• This creates a real problem for reality which, for the foreseeable future, has no hope of matching those expectations.
• As this realisation dawns the result is likely to be disappointment, disillusionment, falling valuations and falling investments.
• In short, the 3^rd AI Winter.